数据来源:ArXiv Domain

LLM Domain Papers

1. DeepUnfolding_Agent: A Classical_Quantum Adaptive Control API

Abstract:Adaptive optimization workflows in machine learning, federated learning, quantum machine learning, and hybrid classical quantum systems often embed controller logic directly inside task specific training scripts. This coupling limits reuse, reproducibility, safety validation, and fair comparison across execution environments. This paper presents DeepUnfolding Agent, a schema-driven FastAPI platform that exposes deep unfolded adaptive control as an external software service. External executors register a strategy schema, submit round-level diagnostic variables, receive schema-bounded control actions, execute one optimization round, and return feedback for controller update. The framework separates controller intelligence from domain specific execution through a strategy registry, diagnostic to state adapter, controller factory, safety guard, executor abstraction, feedback update path, SQLite persistence layer, Docker deployment, and reproducible benchmark scripts. The implementation supports classical, quantum-aware, hybrid, Qiskit Aer, and optional IBM quantum processing execution pathways. We validate the software through unit, schema, API, safety, orchestration, Quantum execution, checkpointing, and Docker tests, and demonstrate external user fixed versus adaptive workflows across finance, healthcare, genomics, transport, noisy quantum simulation, and hardware backed IBM Quantum execution. The results show that DeepUnfolding-Agent provides a reusable and traceable adaptive control layer for heterogeneous optimization workflows. Our quantum experiments are interpreted as API level quantum control and hardware portability evidence.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20615 (HTTP 429)

Authors: Shanika Iroshi Nanayakkara, Shiva Raj Pokhrel

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20615.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20615

Published: 2026-10-19T01:39:44.857Z


2. PRISM: PRIvileged Sibling-guided Multi-Agent Topology Design

Abstract:Multi-agent systems (MAS) based on large language models (LLMs) have become a powerful way to address complex problems, and their effectiveness depends critically on the collaboration topology that decides which agents participate and how they communicate. Recent work learns to generate this topology from the task query, fitting the generator to corpora of execution-verified graphs. The post-training method for such generators remains a limitation: a corpus lists the graphs that solved a query but never ranks them, leaving continued supervised fine-tuning (SFT) with no per-query preference to exploit. Without such a preference, the generator drifts toward the topologies the corpus contains most often rather than those that suit each query, so a gain on one task can come at a cost on another. We propose PRISM, which recovers per-query signal from the same corpus by learning with privileged information: a teacher reads, through a zero-initialized low-rank branch, a sibling graph that also solved the query and is unavailable at deployment, and the student matches its per-decision targets on trajectories the student samples itself. Across six diverse benchmarks PRISM improves over every topology-design baseline evaluated, and matched-teacher controls support the contribution of privileged query information. Code is available in this link.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Q: 这篇论文试图解决什么问题? 这篇论文针对的是\*\*基于大语言模型的多智能体系统(MAS)中,自动化的协作拓扑生成器在后训练阶段面临的核心监督信号缺失问题\*\*。 具体而言,现有方法通过收集执行验证的成功查询-图对(query-graph pairs)来训练拓扑生成器,但这种语料库存在两个根本性局限: 1. \*\*缺乏查询级别的偏好信号\*\*:语料库仅记录了哪些图成功解决了查询,却未对这些成功图按查询进行排序或标注优劣。这使得持续的监督微调(SFT)只能追求与记录图的最大似然匹配,导致生成器倾向于拟合语料库中最频繁出现的拓扑模式,而非针对每个特定查询选择最合适的拓扑——这往往造成在部分任务上提升、在另一些任务上退步的权衡现象。 2. \*\*训练与推理之间的轨迹不匹配(暴露偏差)\*\*:在SFT中,教师强制(teacher forcing)仅在记录的构造轨迹上提供监督。然而,在部署时,生成器必须基于自身先前决策所产生的部分图(partial graphs)进行自回归决策。一旦学生的采样轨迹偏离训练记录,就会进入语料库未覆盖的状态,从而缺乏直接的监督信号。 为解决上述问题,论文提出了 \*\*PRISM\*\*(Privileged Sibling-guided Multi-agent topology design)。该方法的核心思想是:在训练阶段引入\*\*特权信息\*\*——即一个与当前查询匹配、且同样成功解决了该查询的“兄弟图”(sibling graph)——作为教师模型的额外输入;学生模型则在自己采样得到的轨迹上,通过策略内蒸馏(on-policy self-distillation)来匹配该特权教师的决策分布。这样既修正了训练-推理轨迹不匹配的问题,又利用了查询相关的成功替代拓扑信息,且无需在部署时保留特权通道或进行额外的多智能体执行。 Q2: 有哪些相关研究? 论文将相关研究归纳为以下两个主要方向: ### 1. 多智能体系统通信拓扑设计(MAS Communication Topology Design) 该方向的研究脉络可从预设结构逐步过渡到学习式生成: - \*\*早期预设结构\*\*:包括链式推理(Chain-of-Thought, Wei et al., 2022; Hong et al., 2024)、树状搜索(Tree of Thoughts, Yao et al., 2023)以及辩论协议(Du et al., 2024)。 - \*\*通信结构的重要性\*\*:研究表明,多智能体系统的性能高度依赖于通信结构(Qian et al., 2025; Shen et al., 2025),同时协作的计算成本也推动了效率优化方面的研究(Li et al., 2026a)。 - \*\*自动化拓扑设计\*\*: - \*\*AgentPrune\*\*(Zhang et al., 2025a):通过剪枝通信边来优化拓扑。 - \*\*AgentDropout\*\*(Wang et al., 2025):在通信轮次中动态移除智能体与边。 - \*\*G-Designer\*\*(Zhang et al., 2025b):基于变分图自编码器(variational graph autoencoder)生成任务条件化的通信图。 - \*\*ARG-DESIGNER\*\*(Li et al., 2026b):采用自回归方式生成智能体角色与通信边,允许智能体数量及其角色组成随查询动态变化,扩展了设计空间。 - \*\*补充研究\*\*:包括基于失败轨迹进行策略内细化以实现智能体安全对齐(Yin et al., 2026),以及通过对抗性操纵增加工具增强推理的计算开销(Li & Wang, 2026)。 本文的研究定位是在已基于执行验证图完成监督学习初始化的自回归拓扑生成器之上,进行后训练(post-training)的优化;研究保留了 ARG-DESIGNER 的部署架构,并通过引入特权教师(privileged teacher)在策略内提供额外监督。 ### 2. 策略内蒸馏(On-Policy Distillation) 该方向关注如何从学习策略自身访问的状态中进行蒸馏,以缓解训练与推理之间的分布不匹配: - \*\*DAgger\*\*(Ross et al., 2011):从策略自身访问的状态中学习,而非仅从记录者(demonstrator)记录的轨迹中学习。 - \*\*GKD\*\*(Agarwal et al., 2024)与 \*\*MiniLLM\*\*(Gu et al., 2024):将上述思想引入大语言模型蒸馏,即从学生模型采样并由教师模型打分;MiniLLM 还进一步启发了基于模式寻求(mode-seeking)的反向 KL 散度目标。 - \*\*自蒸馏(Self-Distillation)\*\*:近期研究将教师与学生设为同一模型,而非传统的大-小模型对: - Zhao et al.(2026):基于模型自身的策略内样本蒸馏推理能力。 - Dat et al.(2026):将类似方案扩展至扩散语言模型(diffusion language models)。 本文提出的 \*\*PRISM\*\* 属于策略内自蒸馏框架,其特点在于: - 教师与学生通过\*\*特权输入(privileged input)\*\*分离,而非通过模型规模差异或早期检查点分离。 - 决策空间为离散的小规模分类空间(角色选择 + END 标记;边存在/不存在),因此 KL 散度可被\*\*精确计算\*\*,无需估计。 Q3: 论文如何解决这个问题? 该论文提出 \*\*PRISM\*\*(Privileged Sibling-guided Multi-agent topology design),通过引入\*\*训练时特权信息\*\*与\*\*策略内自蒸馏\*\*,解决自回归拓扑生成器在后训练阶段面临的监督信号缺失与训练-推理轨迹不匹配问题。具体解决方案可分为以下四个阶段及核心技术组件: ### 1. 总体思路 PRISM 的核心在于:对于每个训练查询,利用一个“兄弟图”(sibling graph)——即同一查询在语料库中另一个同样成功、但结构不同的已验证拓扑——作为\*\*仅在训练时可用\*\*的特权输入,构建一个特权教师模型。该教师模型在学生策略自身采样得到的轨迹上提供决策级软目标,使学生模型即使偏离了原始记录轨迹,仍能获得有效监督。同时,通过监督锚定损失保留对已验证图的直接拟合。部署时,特权通道被完全移除,学生模型以与基线相同的架构进行推理。 ### 2. 四阶段流程 #### Stage 0:冷启动(Cold Start) 遵循既有管线(ARG-DESIGNER),通过执行验证构建监督语料库 D = (G(c), Q) mid S(Q, G(c)) = 1 ,并训练初始拓扑生成器至收敛,获得部署检查点 θ_0 。 #### Stage 1:构建特权教师(Privileged Teacher) 从 θ_0 克隆一个教师模型,并为其注入一个额外的特权输入分支: - \*\*兄弟图采样\*\*:对于每个已验证目标图 G ,从其所在查询 Q 的其他成功图中均匀采样一个\*\*文本表示不同\*\*的兄弟图 a 。若查询仅有一个成功拓扑,则特权输入置为零向量。 - \*\*FiLM 调制\*\*:将兄弟图编码 f_a 通过一个小型编码器(层归一化、Dropout、线性映射至秩 r=16 、 tanh 激活)后,生成 FiLM 参数,调制查询条件向量:

c(teacher) = TaskProc(f_Q) odot (1 + γ(f_a)) + β(f_a)
其中 γ, β: R^r to R^d 为可学习的仿射变换。 - **零初始化**: γ 与 β 的输出层初始化为零,确保教师模型在训练初期与学生模型行为完全一致,消除初始功能错配。 - **训练目标**:教师模型以加权负对数似然(Eq. 2)在完整语料库上进行监督微调,更新全部参数;此阶段学生模型保持冻结。 #### Stage 2:策略内自蒸馏(On-Policy Self-Distillation) 此阶段学生模型开始更新,教师模型冻结: - **策略内采样**:对每个训练查询,从当前学生策略 P
θ(· mid Q) 中采样 M=4 条完整拓扑轨迹。这些轨迹在**每个优化步骤都重新采样**,确保教师始终在**学生当前实际访问的状态**上提供纠正信号。 - **教师重打分**:对每条采样轨迹,教师模型在特权输入 a (同一查询的固定参考兄弟图)的条件下进行前向传播,输出每步的节点与边决策分布;学生模型在无特权输入的条件下对同一条轨迹进行打分。 - **精确 KL 蒸馏**:由于师生共享相同的节点与边符号集,每步的分类 KL 散度可被精确计算:

L(KL) = α · E(node steps)[KL(Pθ ,|, P(teacher))] + (1-α) · E(edge slots)[KL(Pθ ,|, P(teacher))]
其中采用**反向 KL**(reverse KL),即 KL(P
θ | P(teacher)) ,以鼓励学生模型集中于教师偏好的少数决策模式,避免在多个成功拓扑间简单平均。 - **监督锚定**:为防止学生偏离有效的拓扑模式,引入对已验证参考图 G(real) 的监督损失:
L(anchor) = -log Pθ(G_(real) mid Q)

  • **联合目标**:
    L(PRISM) = (1-λ)L(KL) + λ L(anchor)
    其中 λ = 0.3 。梯度仅通过重打分传播,采样过程本身不可微;教师不计算梯度。 #### Stage 3:推理部署(Inference) 部署时仅使用学生模型。特权输入通道被完全移除,因此: - 推理架构与 ARG-DESIGNER 完全一致; - 无需额外的多智能体执行或计算开销; - 每次查询仍保持约 3.04 个智能体、 4.04 次 LLM 调用的低成本。 ### 3. 关键设计动机 - **解决训练-推理不匹配**:不同于教师强制仅在记录轨迹上监督,PRISM 通过策略内采样确保监督覆盖学生自身决策所到达的离轨状态(off-trajectory states)。 - **恢复查询级偏好**:兄弟图提供了关于“同一查询的另一成功构造方式”的信息,使蒸馏目标包含查询相关的拓扑偏好,而非语料库全局的边际分布。 - **参数高效与可控**:秩 r=16 的 FiLM 分支仅增加约 20k 参数(约为全宽通道的 1/30 ),且通过零初始化与冻结教师策略,严格隔离训练特权与部署行为。 Q4: 论文做了哪些实验? 论文在六个基准测试上开展了一系列系统实验,涵盖主结果、诊断分析、控制实验与消融研究,具体如下: ### 1. 主结果与基准对比 在 **GSM8K、MultiArith、SVAMP、AQuA、HumanEval、MMLU** 六个基准上,以 **Qwen2.5-7B-Instruct** 为骨干,将 PRISM 与以下基线/对照进行比较: - **Vanilla**(单智能体) - **AgentPrune、AgentDropout、G-Designer、ARG-DESIGNER**(既有拓扑设计方法) - **SFT(对照)**:从同一部署检查点 θ_0 出发,持续进行监督最大似然训练,与 PRISM 共享训练轮数与学习率 结果显示,PRISM 在所有六个基准上均取得最高 Pass@1/准确率(Table 1)。值得注意的是,SFT 在五个基准上持平或提升,但在 **AQuA 上较基线下降 12.60%**,而 PRISM 较基线提升;同时,PRISM 的平均每次查询调用成本仅从 4.00 增至 4.04 次 LLM 调用(Table 2)。 ### 2. 输出分布集中度分析 为验证持续 SFT 是否导致输出分布退化,论文测量了生成拓扑的 **top-1 形状占比**(即最频繁角色组合被发射的查询比例): - ARG-DESIGNER(冷启动):**0.171** - SFT:**0.346**(显著上升,分布更集中) - PRISM:**0.157**(低于冷启动,保持多样性) 此外,Table 3 报告了 SFT 模态角色分配在各数学基准上的条件准确率:该模态形状在 GSM8K 上带来 +2.0% 增益,但在 AQuA 上造成 -3.4% 损失,说明集中本身不等于更好。 ### 3. 难度匹配控制实验 为避免不同问题集难度差异干扰比较,论文按 **SFT 是否发射其默认角色组合** 将问题二分,并让两个臂在完全相同的问题子集上评分(Table 4): - 在 AQuA 上,PRISM 在 SFT 默认组提升 ** +18.7% **( p=0.0001 ),在非默认组提升 ** +11.0% **( p=0.003 ) - 在 GSM8K 与 SVAMP 上,两组增益较小且不显著,说明 AQuA 的退化并非单纯由问题难度引起 ### 4. 特权通道信息含量验证 为证明 FiLM 通道确实读取了特权内容而非仅利用额外容量,论文设计了两种测量(§6.1): - **特权增益**: Delta = log P
    (teacher)(G mid Q, a) - log P(teacher)(G mid Q) ,在留出的已验证图上真实通道达到 ** +2.45 nats**,而**乱序控制**(输入另一查询的图)始终接近零 - **FiLM 范数**:真实通道的 |film| 上升至 3.10,乱序通道仅上升至 1.15(Figure 4) 这表明优化器确实训练了通道,但只有真实兄弟图内容才产生可预测的似然增益。 ### 5. 内容 vs. 程序分离实验 为区分“策略内蒸馏过程本身”与“特权内容”各自的贡献,论文在 Stage 2 中固定其余所有设置(初始化、采样、超参数、计算预算),仅更换教师信号(Figure 2): - **无教师(SFT)** - **乱序教师**(另一查询的图) - **常数教师**(固定特权输入) - **真实兄弟图教师(PRISM)** 结果显示: - 乱序教师平均劣于常数教师,说明错误匹配的查询信息会误导学生 - PRISM 在全部六个基准上均优于常数教师(平均 +1.66% ,符号检验 p=0.031 ),证明**查询匹配的特权内容**在蒸馏程序之外提供了额外收益 ### 6. 消融实验 #### 6.1 特权通道形式 固定 Stage 2 其余设置,比较向教师提供不同类型的特权输入(Figure 3): - **context**(问题与答案文本) - **rationale**(推理过程/测试用例/正确选项) - **graph stats**(查询的成功图统计摘要) - **sibling graph**(默认,另一成功拓扑) **sibling graph** 在六个基准上均为最优,且在 AQuA 上领先次优对手达 14.96%。context 与 rationale 表现最差,甚至低于 SFT。 #### 6.2 KL 散度方向 在特权通道固定为 sibling graph 的条件下,比较三种散度目标(Figure 5): - **Reverse KL**(默认): KL(Pθ | P(teacher)) - **Forward KL**: KL(P(teacher) | P_θ) - **JSD**( β=0.5 ) Reverse KL 在数学基准上均最优,平均准确率最高(85.52),符合其“模式寻求”(mode-seeking)特性——避免在多个已验证拓扑间做简单平均。 ### 7. 附加诊断与稳健性实验 - **种子稳健性**(Appendix D):每个臂在种子 42/43/44 下各运行三次。PRISM 与 SFT 的基准平均差距(约 3.18)远大于运行间波动(标准差多低于 0.6),且 SFT 在 AQuA 上持续低迷(约 63%),确认结果非随机噪声(Table 8) - **FiLM 秩**(Appendix E):将低秩通道从 r=16 增至 r=64, 256 及全宽,各设定与默认差异不超过 2.7%,说明该超参数不敏感(Figure 6) - **语料库统计**(Appendix C, F):报告 MIXCHAIN、MMLU、MBPP 三个语料库的蓝图家族构成、求解率、去重后规模,以及具备兄弟图的查询占比(76.9%、87.1%、90.3%) - **输出条件性**(Appendix G):计算互信息 I(Q;G) ,PRISM 的查询条件熵盈余为 0.160 nats,显著高于 SFT(0.051)与冷启动(0.017),表明其输出更依赖查询(Table 10) - **拓扑干预**(Appendix H):在 91 个 AQuA 问题上固定链式连接与推理超参数,仅替换角色组合。SFT 的模态组合仅达 53.85%–61.54%,而 PRISM 的随机排列组合与随机大小匹配组合均显著更高(约 75%–79%),为拓扑因果效应提供直接证据(Table 11, 12) - **定性案例**(Appendix J):在 GSM8K 上展示 6 个具体案例,对比 ARG-DESIGNER 与 PRISM 构建的 crew 及最终答案,显示 PRISM 更倾向于分配 **Math Solver** 等针对性角色,而基线常重复 **Inspector** 或 **Programming Expert** 导致错误。 Q5: 有什么可以进一步探索的点? 基于论文结论与实验诊断,以下方向值得进一步探索: —- ### 1. 查询级拓扑分配的因果验证 论文的结果表明 PRISM 优于基线,但并未严格确立**查询特定的角色分配(query-specific crew assignment)**本身是增益的来源。附录 H 的干预实验显示,将 PRISM 的 crew 随机排列后,在部分子集上仍能达到相近表现。未来可设计更具代表性的留出集实验,直接检验“将特定拓扑匹配到特定查询”相较于“从查询无关的成功拓扑池中随机选取”的因果效应。 ### 2. 匹配计算量下的方法比较 PRISM 的 Stage 1 需额外训练教师模型(40 epochs),Stage 2 涉及策略内采样与重打分,整体后训练开销高于 SFT。尽管论文在内容 vs. 程序对照中固定了 Stage 2 预算,但**全程匹配计算量**的严格比较(例如,将 SFT 的总梯度步数或浮点运算扩展至与 PRISM 相当时)尚未进行,这会影响对方法效率的公允评估。 ### 3. 跨骨干模型的迁移与通用性 全部实验基于 **Qwen2.5-7B-Instruct** 完成。需验证 PRISM 的增益是否可迁移至: - 不同规模的骨干模型(如 3B 或 70B); - 不同架构或系列的模型(如 Llama、GPT 等); - 更长的上下文窗口或多模态场景。 ### 4. 特权信息形式的进一步扩展 论文测试了 context、rationale、graph stats 与 sibling graph 等特权通道,其中 sibling graph 最优。其他潜在形式尚未探索: - **执行轨迹**:智能体间的完整消息传递序列; - **中间推理状态**:如思维链的中间步骤或程序执行的中间变量; - **对抗或失败案例**:提供未成功图作为负例特权信号,以明确禁止某些拓扑模式。 ### 5. 多轮通信(K > 1)场景下的有效性 实验固定通信轮数 **K = 1**。当智能体进行多轮交互时,拓扑的动态变化(如每轮重连、AgentDropout 式的动态剪枝)可能引入新的训练-推理不匹配。需验证 PRISM 的特权蒸馏框架在动态或时变拓扑中的扩展性。 ### 6. 兄弟图采样策略的优化 当前 sibling 从查询的成功图中**均匀随机采样**。可探索更具结构意识的采样: - 基于与当前学生轨迹的相似度或互补性采样; - 基于图编辑距离或功能多样性选择,以最大化教师信号的覆盖范围; - 对无 sibling 的查询(MIXCHAIN 中占 23.1%)设计更好的回退机制,而非简单的零向量。 ### 7. 输出集中度与准确率的因果机制 论文观察到 SFT 提高输出集中度(top-1 share 从 0.171 升至 0.346)但伴随部分基准退化,而 PRISM 降低集中度至 0.157 却全面提升。需进一步厘清: - 集中度本身是否是准确率的充分/必要条件; - 是否存在“最优集中度”随任务或查询类型变化的规律; - 除 reverse KL 外,其他模式寻求目标(如基于能量的模型、GFlowNets)能否更精细地控制多样性-质量权衡。 ### 8. 教师-学生架构的替代设计 论文采用 **FiLM 调制 + 零初始化低秩分支** 注入特权信息。其他参数高效融合机制值得对比: - Cross-attention 层让查询表示 attend 到兄弟图编码; - Adapter 或 LoRA 在特定层插入; - 将特权信息作为软提示(soft prompts)前置到输入序列。 ### 9. 与统一跨域设计框架的结合 相关工作 **OFA-MAS**(Li et al., 2026c)提出基于混合专家图生成模型的统一拓扑设计。未来可探索将 PRISM 的特权蒸馏作为后训练阶段,嫁接到此类统一模型上,检验跨领域(如从数学推理迁移到代码生成)的泛化能力,而非为每个域单独构建语料库。 ### 10. 失败模式与鲁棒性分析 论文主要关注平均准确率提升,对以下失败模式缺乏深入剖析: - 哪些查询在 SFT 后显著退化、PRISM 又如何修复; - 特权教师本身在何种情况下产生误导性高置信度分布; - 学生策略在蒸馏过程中是否会出现模式坍塌(mode collapse)至某一特定兄弟图。 Q6: 总结一下论文的主要内容 该论文针对基于大语言模型的多智能体系统(MAS)中**查询条件化的协作拓扑生成器**的后训练问题,提出了一种利用特权信息进行策略内蒸馏的新方法。以下是主要内容的结构化总结: —- ### 1. 研究背景与核心挑战 - **问题设定**:多智能体系统的有效性依赖于协作拓扑(即哪些智能体参与、信息如何流动)。近期研究通过自回归生成器,从任务查询中直接生成拓扑图,并在执行验证的语料库上进行后训练。 - **核心局限**: - **缺乏查询级偏好信号**:验证语料库仅记录“哪些图成功解决了查询”,但未对同一查询的多个成功图进行优劣排序。持续的监督微调(SFT)因此只能追求与记录图的最大似然匹配,导致生成器倾向于拟合语料库中最频繁出现的拓扑模式,而非针对每个查询选择最合适的拓扑。实验发现,SFT 在部分基准(如 GSM8K)上提升,却在另一些基准(如 AQuA)上出现显著退化(下降 12.60% )。 - **训练-推理轨迹不匹配**:SFT 采用教师强制(teacher forcing),仅在记录的构造轨迹上提供监督;而部署时生成器必须基于自身先前决策产生的部分图(partial graphs)自回归推理。一旦采样偏离记录轨迹,模型将进入缺乏直接监督的离轨状态(暴露偏差)。 —- ### 2. PRISM 方法概述 论文提出 **PRISM**(Privileged Sibling-guided Multi-agent topology design),其核心思想是:**在训练时引入一个仅在训练阶段可用、部署时移除的“兄弟图”(sibling graph)作为特权信息**,构建特权教师模型,对学生的策略内轨迹进行决策级蒸馏。 方法包含四个阶段: - **Stage 0(冷启动)**:基于执行验证构建查询-图语料库,训练初始生成器 θ_0 。 - **Stage 1(特权教师训练)**:克隆 θ_0 为教师,通过一个零初始化的低秩 FiLM 分支读取查询对应的另一成功拓扑(兄弟图),在完整语料库上进行监督微调。 - **Stage 2(策略内自蒸馏)**:学生模型从当前策略中采样多条轨迹(on-policy),由特权教师在这些轨迹的每个决策点提供软目标;学生通过反向 KL 散度匹配教师,同时以监督锚定损失保留对已验证图的直接拟合。 - **Stage 3(部署)**:完全移除特权通道,学生模型以与基线相同的架构进行推理。 —- ### 3. 关键技术组件 - **FiLM 特权调制**:

c(teacher) = TaskProc(f_Q) odot (1 + γ(f_a)) + β(f_a)
其中 f_a 为兄弟图编码,经小型编码器(秩 r=16 )生成 FiLM 参数 γ, β 。零初始化确保教师初始行为与学生一致。 - **兄弟图采样**:对每查询,从其成功图中均匀采样一个与目标图文本表示不同的兄弟图;若查询仅有一个成功拓扑,则特权输入置零。 - **精确反向 KL 蒸馏**:
L
(KL) = α · E(node)[KL(Pθ ,|, P(teacher))] + (1-α) · E(edge)[KL(Pθ ,|, P(teacher))]
采用反向 KL( KL(Pθ | P(teacher)) )以鼓励学生集中于教师偏好的模式,避免在多个成功图间做简单平均。 - **监督锚定**:
L(PRISM) = (1-λ)L(KL) + λ L(anchor), quad L(anchor) = -log Pθ(G(real) mid Q)
其中 λ=0.3 ,防止学生漂移出有效拓扑分布。 —- ### 4. 实验与主结果 在 **Qwen2.5-7B-Instruct** 骨干上,于 **GSM8K、MultiArith、SVAMP、AQuA、HumanEval、MMLU** 六个基准进行评估: - **主结果**(Table 1):PRISM 在所有六个基准上均优于所有基线(ARG-DESIGNER、G-Designer、AgentPrune、AgentDropout 及单智能体 Vanilla)。相比 SFT 对照,PRISM 在 AQuA 上修复了 13.78% 的巨大差距,在其他基准上也取得一致提升。 - **成本**(Table 2):PRISM 平均每次查询调用 3.04 个智能体、 4.04 次 LLM 调用,仅比基线增加 1% 的调用量。 - **输出分布**:SFT 使最频繁拓扑的占比从 0.171 上升至 0.346 (分布更集中),而 PRISM 将其降至 0.157 ;同时 PRISM 的查询条件互信息更高,表明其输出更依赖查询本身而非退化为常数。 —- ### 5. 诊断、控制实验与消融 - **难度匹配控制**(Table 4):按 SFT 是否发射其默认角色组合将问题二分后,PRISM 在 AQuA 的两组子集上均显著提升(默认组 +18.7% , p=0.0001 ;非默认组 +11.0% , p=0.003 ),表明增益不限于修正单一退化模式。 - **特权信息含量验证**(§6.1, Figure 4):通过“特权增益” Delta = log P(teacher)(G mid Q, a) - log P(teacher)(G mid Q) 与 FiLM 范数监测,证明真实兄弟图通道产生 +2.45 nats 的似然增益,而乱序控制(另一查询的图)始终接近零,确认教师确实读取了特权内容。 - **内容 vs. 程序分离**(Figure 2):固定 Stage 2 的全部设置,仅更换教师信号。结果显示“常数输入教师”优于“乱序教师”,而 PRISM(真实兄弟图)又在全部六个基准上优于常数教师(平均 +1.66% , p=0.031 ),证明查询匹配的特权内容本身在蒸馏程序之外提供了额外收益。 - **特权通道形式**(Figure 3):对比 context、rationale、graph stats 与 sibling graph,**sibling graph** 在全部基准上最优,且在 AQuA 上领先次优对手达 14.96% 。 - **KL 方向**(Figure 5):反向 KL 在数学基准上全面优于前向 KL 与 JSD,支持其模式寻求(mode-seeking)特性更适合多成功拓扑场景。 - **拓扑干预**(Appendix H):在固定连接方式与推理超参数的条件下,直接替换角色组合。SFT 的模态组合表现显著劣于 PRISM 的随机排列组合与大小匹配随机组合,为拓扑的因果效应提供了直接证据。 —- ### 6. 结论与待探索方向 - **结论**:PRISM 通过训练时特权兄弟图与策略内自蒸馏,有效缓解了 SFT 的查询级偏好缺失与训练-推理不匹配问题,在六个 diverse 基准上实现了稳定且显著的提升,且部署成本与架构均无额外负担。 - **局限与未来方向**: - 尚未在匹配总计算量的条件下与 SFT 进行严格对比; - 未完全确立“查询特定拓扑分配”对下游准确率的独立因果效应; - 需扩展至不同骨干模型规模与系列、多轮通信( K>1 )场景; - 可探索兄弟图的结构化采样策略、其他特权信息形式(如执行轨迹、失败案例)以及教师-学生架构的替代融合机制。 Q7: 想要进一步了解论文 以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文,请点击 这里 为你跳转Kimi AI网页版,并启动一个与该论文相关的新会话。

Authors: Phuong Tuan Dat, Qi Li, Xinchao Wang

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20628.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20628

Published: 2026-10-19T01:39:44.857Z


3. A Large-Scale Dataset for Multi-Class Financial News Classification in Brazilian Portuguese

Abstract:The development of Financial Natural Language Processing (NLP) systems for Brazilian Portuguese is hindered by the scarcity of domain-specific datasets. To address this, we present a large-scale corpus of approximately 160,000 Brazilian Portuguese financial news articles collected from web sources. To establish initial benchmarks for a weakly supervised multi-class classification approach based on semantic similarity, we annotated subsets of this corpus both manually and using a Large Language Model (LLM). We evaluated three similarity-based methods — cosine similarity, K-Nearest Neighbors (KNN), and FAISS-based nearest-neighbor search — using multilingual semantic embeddings. Our experiments reveal that the task is highly sensitive to label quality and class distribution. On the manually labeled subset, KNN achieved the best performance (47.3% accuracy, 31.9% F1-score), closely followed by FAISS. Conversely, on the LLM-annotated subset, while accuracy reached approximately 50%, F1-scores dropped to near 20%, highlighting the negative impact of label imbalance and noisy weak annotations. These findings establish foundational baselines for a challenging task, underscoring the critical need for high-quality manual annotations. Ultimately, this dataset and our preliminary results provide a reference point for future research in Brazilian Portuguese Financial NLP.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20630 (HTTP 429)

Authors: Leomax Bandeira, Gabriel Aguiar, Dalton Cézane Gomes Valadares, Luiz Pereira, Helder Fernando de Araújo Oliveira, Gustavo Xavier, Andson M. Balieiro

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20630.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20630

Published: 2026-10-19T01:39:44.857Z


4. Real-time Chatbot Romance Scam Detection in Stage Classification

Abstract:Romance scam has become a serious problem via social media, and the improvement of artificial intelligence enable to fake identify easily. In the study, a chatbot has developed to detect romance scam in real-time through Line Messaging API. Using the seven stage theories in the social science field, the stages of romance scam can be detected through large language model. In addition, once a romance scam is detected, the chatbot provides explanations of the stage and prevention advices through large language model. The contribution of the work provides a practical application, and can be used in real scenario for reducing the risk of romance scam.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20631 (HTTP 429)

Authors: Yi-cheng Liu, Guan-ru Lai, Yi-cheng Lai, Lyn Chao-ling Chen

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20631.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20631

Published: 2026-10-19T01:39:44.857Z


5. MeddeX: A Retrieval-Augmented LLaMA-3 System with Gemini Vision for Evidence-Backed Medical Report Analysis

Abstract:Modern healthcare generates increasingly complex medical reports that challenge clinical professionals in terms of time and interpretive burden. Manual review processes are slow, error-prone, and difficult to scale. This paper presents MeddeX, a Medical Report Analyzer that combines a QLoRA fine-tuned LLaMA-3 (8B) model with Retrieval-Augmented Gen eration (RAG) and a multimodal Gemini Vision layer to deliver accurate, concise, and evidence-backed summaries of medical documents. The system handles text-native PDFs, scanned image based PDFs, and direct radiological images through an intelligent hybrid extraction pipeline that includes a novel garbage-text detection mechanism for corrupted and UTF-16-encoded doc uments. Beyond single-report Q&A, MeddeX introduces report scoped retrieval for precise contextual grounding, an automated longitudinal report comparison engine, and a self-healing backfill pipeline that lazily processes legacy documents on demand. A FastAPI backend, Supabase vector store, and this http URL fron tend complete the end-to-end architecture. Evaluation against the base LLaMA3-Med42-8B model demonstrates consistent improvements across ROUGE-1 (+9.2%), ROUGE-2 (+10.0%), ROUGE-L (+11.8%), BLEU (+71.4%), and BERTScore (+0.8%), validating the benefit of domain-specific fine-tuning. MeddeX represents a practical and robust step toward AI-assisted clinical decision support.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20632 (HTTP 429)

Authors: Rajdeep Pandhare, Rohit Kundaikar, Yakub Kojar, Shreyansh Tiwari

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20632.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20632

Published: 2026-10-19T01:39:44.857Z


6. Context Injection and Long-Horizon Reliability: Separating Duplicate Recognition, Local Processing, and Counter Tracking

Abstract:Long-horizon failures of language models are usually measured as end-task success or as a horizon length. We instead score, step by step, three components of a 400-step ledger task in which every step is easy in isolation and the model’s own answers are re-injected into its context: recognising a record already seen 5 to 300 steps earlier, processing the current record locally, and writing a counter incremented at every step. On Qwen3-8B (non-thinking, 20 seeds, replicated on a second machine and backend), with every step scored component by component, two results hold. First, duplicate-recognition errors rise late in the episode — from 0.0019 to 0.011 per step at risk across step 200 — while the counter makes no error in 8,000 steps and local processing is descriptively stable; the rise combines long-distance recall errors and rising false alarms. Second, a block of irrelevant content inserted before each step makes the counter fail in 20 of 20 seeds on each machine, against none without it, while its effect on local processing is much smaller and not distinguishable from none. Three pre-registered follow-ups probe the counter. Its value depends on the value present in the response history: a corrupted value persists without correction in every seed, and an exact external reminder, repeated before each message, re-aligns it within one to two steps. Under the irrelevant block, no counter error was observed when an exact compact reminder of the state, supplied externally by an oracle, preceded each message (0 of 20 seeds with an error, against 15 without it), and this reminder outperformed a length-matched neutral text; whether a single reminder suffices was not conclusive, and the absence of damage to the other components was not established. An aggregate or exclusive error classification can hide the difference we report.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20633 (HTTP 429)

Authors: Thomas Serval

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20633.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20633

Published: 2026-10-19T01:39:44.857Z


7. Jev-as-a-Judge for RAG Claim Verification: Confidence-Gated Cascading Between a Non-Generative Verifier and a Frontier LLM

Abstract:Retrieval-augmented generation asks one question at volume: does this evidence support this claim? The answer is one bit, yet most pipelines obtain it from a text generator and pay generation prices for the prose wrapped around it. We compare a non-generative decision model, which returns a typed value from a declared enumeration together with a native probability and no text, against a frontier LLM configured as an engineer would deploy it for binary classification. Both answer the same 495 human-labelled claims from LLM-AggreFact, so every comparison is within-item. We report three things. First, the two systems are statistically indistinguishable on the metric this task is conventionally scored by, a paired difference of -0.56 points, Jev minus Astra, at a 187-fold cost difference. Second, that tie carries almost no information about deployment: the systems disagree on 15.2% of claims, and the cheaper one admits 23.2% of the claims annotators marked unsupported against the frontier model’s 14.1%. A component whose job is to catch unsupported claims should be judged on that rate, and the conventional headline hides it. Third, because the errors differ they compose: routing the claims the cheap model is confident about and escalating the rest reaches 74.10 at USD 2.92 per thousand claims against 73.83 at USD 8.93, with the threshold chosen out of fold. But a matched random-escalation control at in-sample thresholds, spending the same budget by coin flip, does as well, so what buys the saving is the escalation share and not the probability that picks what to escalate. Measured against the cheap model it is built from, the cascade costs 61-fold more for 0.83 points, 95% CI [-3.36, 4.86], and it is the more permissive guard of the two. Equal accuracy conceals unequal errors. All predictions, prompts, pricing and code are released.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20634 (HTTP 429)

Authors: Alden Do Rosario, Kirollos Atef

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20634.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20634

Published: 2026-10-19T01:39:44.857Z


8. Exploring the Design Space of Permutation Diffusion LLMs

Abstract:Permutation generation with language models requires both text-conditioned ranking and a global one-to-one assignment between candidates and output positions. We study this problem through three diffusion-LLM parameterizations: hard-assignment prediction over shuffled candidate-index sequences, soft-assignment prediction over doubly stochastic matrices, and direct ordering-score prediction followed by sorting. The three families place permutation constraints at different stages of prediction and decoding and therefore support different training objectives and sampling procedures. We evaluate these designs on recommendation reranking across three Amazon domains with candidate sets of size 10 and 20. At one function evaluation, the hard- and soft-assignment variants achieve Recall@1 comparable to a constrained masked-diffusion baseline, whereas direct ordering-score prediction guarantees valid permutations without Hungarian assignment or Sinkhorn normalization. Additional function evaluations do not consistently improve ranking quality, and their effect depends strongly on the state representation and training objective. These results show that corruption, prediction parameterization, objective, and decoder should be designed jointly for permutation diffusion with language-model denoisers.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20635 (HTTP 429)

Authors: Zhe Xu, Mingfu Liang, Kavosh Asadi, Jean-Pascal Billaud, Tianyi Chen, Yanhui Guo, Christina Huang, Baekjin Kim, Jimmy Kim, Yufei Li, Daqi Sun, Jiachen Sun, Dian Yu, Zaiwei Zhang, Kaushik Rangadurai, Frank Shyu, Sandeep Pandey, Xi Liu, Luke Simon

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20635.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20635

Published: 2026-10-19T01:39:44.857Z


9. AbridgeKV: Compressing Long Reasoning into the KV Cache of Learned Pause Tokens

Abstract:Long chain-of-thought decoding is memory-bound: the KV cache of the generated tokens grows linearly with the test-time compute being spent. Evicting cache entries does not fix this, because a model never trained on a cache with entries removed generates longer, often up to the length limit. We instead train the model to compress its own output. AbridgeKV compresses each completed block of generated text into the KV entries of a few learned pause tokens. These are extra input tokens with trained embeddings that produce no output. Once the pause tokens have encoded the block, AbridgeKV frees the block’s original cache; subsequent tokens attend to the pause tokens’ KV entries in its place. Three novel choices make this work: the pause tokens’ KV entries take positions inside the block they replace, the cache budget is allocated across layers according to the measured value of each layer’s entries, and compression operates on relatively large blocks. On 446 competition problems from AIME, HMMT, AMC, BRUMO, CMIMC and SMT with 32K-token generations, keeping a quarter of the generated cache matches an identically trained full-KV model in accuracy and output length; the strongest published evictor at a matched budget writes far longer, and the closest trained compressor, retrained on the same data, loses 10 points. Served in vLLM, the cache is 3.5x smaller per sequence and a saturated H100 finishes a problem 3.2x faster than with full KV.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20641 (HTTP 429)

Authors: Jack Cai, Lijie Yang, Vasilis Kontonis, Dimitris Papailiopoulos, Shivam Garg

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20641.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20641

Published: 2026-10-19T01:39:44.857Z


10. CoTS: Competitive Token Selection for Latent Reasoning in Large Language Models

Abstract:Recent latent reasoning methods move beyond discrete Chain-of-Thought (CoT) by replacing sampled tokens with continuous soft embeddings or hidden states, enabling richer intermediate representations and preserving alternative reasoning paths. However, the need for latent reasoning varies across decoding steps: confident predictions favor discrete commitment, whereas ambiguous steps benefit from preserving multiple plausible alternatives. Existing approaches typically handle this variation through either global soft activation or hard switching based on uncertainty thresholds, treating latent intervention as an all-or-nothing decision rather than continuously adapting its strength. To address this limitation, we propose Competitive Token Selection (CoTS), a training-free framework that reframes latent intervention from binary mode selection into continuous allocation at the token level. CoTS uses the probability competition between the two most likely candidates to dynamically balance discrete commitment and soft exploration at each reasoning step. The second candidate represents the strongest alternative to the model’s preferred continuation, so the separation between the two directly reflects the stability of the current discrete choice. When one candidate clearly dominates, CoTS keeps the representation close to the sampled discrete token to preserve a stable reasoning trajectory; as the competition strengthens, it increases the contribution of the soft representation to retain plausible alternatives. This competition-based allocation requires neither manually tuned switching thresholds nor mixing coefficients, and introduces no additional forward passes. Experiments across six reasoning benchmarks demonstrate that CoTS consistently outperforms standard CoT and state-of-the-art training-free latent reasoning methods. Our code is available at this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20642 (HTTP 429)

Authors: Jie Yu, Yuxi Sun, Zhaoxu Li, Chunxiao Liu, Xudong Jiang

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.20642.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20642

Published: 2026-10-19T01:39:44.857Z


Agent Domain Papers

1. The World Won’t Wait: Evaluating and Teaching Agents Under a Live Clock

Abstract:Large language model (LLM) agents can solve complex interactive tasks, yet reasoning and tool calls take time. In continuously changing environments, an otherwise correct action may arrive too late. We introduce RealtimeBench, with 180 tasks across six game types, to evaluate agents while games continue running during inference. Paired evaluations show that GPT-6.1-sol outperforms the faster Jev when games are paused, but falls behind when they keep running. We propose Two-Timescale Offline Recursive Self-Improvement (TORSI): a slow teacher studies failures and tests corrections in paused games between attempts, then writes reusable instructions for a fast student. The student acts without teacher calls during gameplay, and validation scores determine which updates are adopted. Two rounds raise Jev’s average test score from 13.51 to 24.19 on a 0—100 progress scale, improving all six game types. The instructions also improve average scores for three other models without retuning. These findings point toward real-time intelligence: enabling powerful LLM agents to remain effective in a continuously changing world by organizing deliberation, learning, and action across timescales.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20665 (HTTP 429)

Authors: Ningxin Pan, Hanyu Li, Yeqing Shen, Ning Ding, Yehui Tang, Wentao Zhang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20665.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20665

Published: 2026-10-19T01:39:59.081Z


2. Evaluating Model Capabilities After Benchmark Saturation

Abstract:Rapid advances in AI capabilities are causing benchmarks to saturate. Once several models approach perfect scores, a benchmark loses its power to differentiate between them, and with it its utility for informing AI safety cases and frontier policies. Yet benchmark items produce far more information than the scores we typically extract from them: a full transcript of how each model solved the task. We investigate whether these transcripts can be used to assess model capabilities even after benchmark scores have saturated. We simulate saturation retrospectively using historical benchmark and model releases, and test whether protocols that access only correctly solved transcripts from saturated benchmarks can predict capabilities measured on held-out, unsaturated items. We evaluate representative contemporary approaches for extracting information from model transcripts: pairwise judging, single-transcript absolute scoring, and predictions from deterministic transcript features or embeddings. We find that saturated transcripts retain capability information beyond benchmark accuracy. Amongst our evaluated protocols, pairwise judging performs best, reducing prediction error by 5.3 times compared to a baseline using only saturated benchmark scores. This capability-relevant information is distributed across transcript sections, but concentrated in specific channels of information such as reasoning and tool-calls. However, the predictive power of protocols does not always generalise across distribution shifts — for example by underestimating the capabilities of the first reasoning models introduced in early 2025. Saturated benchmarks thus retain capability-predicting signal in their task transcripts, but extending the lifetime of existing benchmarks requires methods designed to extract these signals robustly.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20667 (HTTP 429)

Authors: Morio Hamada, Jehyeok Yeon, Matteo Migliarini, Justin Olive, Magda Dubois

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20667.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20667

Published: 2026-10-19T01:39:59.081Z


3. Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion

Abstract:Reward-guided discrete diffusion enables test-time alignment with downstream objectives, but applying Sequential Monte Carlo (SMC) guidance at every denoising step is computationally expensive. Active guidance steps require additional reward evaluations and, for gradient-based proposals, value-gradient computations. Sparse guidance reduces this cost by guiding only a subset of timesteps, but commonly used schedules are fixed heuristics whose effectiveness can vary substantially across tasks. We propose VISTA, a training-free method that automatically selects guided timesteps under a fixed budget. VISTA formulates sparse guidance as approximation of the ideal full-guidance probability path and derives an estimable upper bound on (\ell_1) path deviation. Using value statistics collected from a small number of full-guidance warmup generations, VISTA optimizes the resulting empirical surrogate exactly via dynamic programming. The warmup generations can be retained as model outputs, and after schedule selection VISTA has the same per-generation guidance complexity as other sparse schedules with the same budget. Across text, biological sequence, and image generation tasks, VISTA discovers task-dependent schedules and outperforms hand-designed sparse schedules in most tested settings, while approaching or exceeding full-guidance performance on several task metrics. In toxic-text generation, for example, VISTA attains a 0.99 unique-toxic ratio compared with 1.00 for full guidance while reducing post-selection per-generation latency from 204.23 s to 38.16 s, a 5.35$\times$ speedup.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20669 (HTTP 429)

Authors: Jaehyeon Kim, Ziseok Lee, Siyoung Choi, Kyungsu Kim

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20669.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20669

Published: 2026-10-19T01:39:59.081Z


4. AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks

Abstract:Agents pair reasoning models with harnesses that organize model calls, execute tools, load skill instructions, and retain history in memory. Understanding agent performance requires linking execution conditions to outcomes, behavior, and resource use. We introduce AlphaDiana, a measurement system with modular interfaces for datasets, agents, environments, and scorers. It preserves harness-native control loops while controlling file reuse and memory retention separately. AlphaDiana supports macro analyses of complete agents and micro interventions on agent components such as tools, skills, and memory under matched conditions. We study three open-source models and three open-source harnesses on five answer-based benchmarks, with two models also tested on two interactive benchmarks. We find that harnesses do not consistently improve the accuracy of reasoning: gains on questions unsolved by the model alone can be outweighed by losses on questions it answers correctly without tools. This inconsistency extends to component interventions: neither skill guidance nor retained history reliably improves accuracy. Behavior and output-token use must be assessed with correctness: low token entropy can accompany wrong answers, and less output per execution can mean more output tokens per correct answer when accuracy falls. AlphaDiana links answer correctness and execution status to traces and resource records for comparing agents and diagnosing failures under shared task and resource policies.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20675 (HTTP 429)

Authors: Zhanke Zhou, Zongze Li, Weikai Huang, Xuan Li, Chentao Cao, Xiao Feng, Xiangyu Lu, Jinbo Hu, Yi Xie, Nico Pelleriti, Shiyang Liu, Menghan Lu, Max Zimmer, Brando Miranda, Jiangchao Yao, Bo Liu, Sanmi Koyejo, Sebastian Pokutta, Bo Han

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20675.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20675

Published: 2026-10-19T01:39:59.081Z


5. When Correctness Feedback Erases Evidence: Phase-Separated Completeness Auditing in Verus

Abstract:Correctness feedback suppresses the very evidence that specification-completeness auditing is designed to uncover: programs that satisfy a written contract but violate intended behavior. We demonstrate this objective reversal on 12 Verus/Rust tasks across four model families under matched call budgets. Independent no-feedback discovery more than doubles certified-witness yield relative to IntentShield’s corrective-feedback loop (59.0% versus 27.1%). A format-matched iterative control places the larger share of this gap in feedback content, not independent restarts. The advantage holds in every model family and survives parser repair. A separate phase-separated execution matches the aggregate yield without reusing responses; most certified witnesses also fail held-out cases. Preserving these witnesses delivers stronger specification repair: Phase-Revised contracts achieve 81.3% implementation success, versus 62.5% for IS-Revised, with Direct-Revised close behind. Both no-feedback witness sources substantially outperform weakened contracts. The effect does not reproduce in the available Dafny and Lean evaluations, establishing an empirical boundary around the studied Verus/Rust workflow. Completeness auditing and correctness repair have opposing immediate objectives. Preserve and certify the admitted failure before repairing it.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20676 (HTTP 429)

Authors: Adarsh Agrawal

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20676.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20676

Published: 2026-10-19T01:39:59.081Z


6. Counterfactual Neighborhood Risk Fields: Evidence-Grounded Selective Safety for Vision-Language Models

Abstract:Representation-based safety defenses for vision-language models (VLMs) often learn to separate harmful from benign inputs, but separation alone does not establish what a risk score means or when it should be trusted. We introduce Counterfactual Neighborhood Risk Fields (CNRF), an evidence-grounded framework for selective VLM safety. CNRF constructs context-matched benign-harmful counterfactual pairs, uses their representation differences as capability-related risk directions, and estimates neighborhood support separately from risk. This separation prevents unfamiliarity from being conflated with harmfulness and enables selective intervention rather than blanket rejection. A victim-adaptive interface accommodates different safety observations across VLM architectures and routes selected requests for same-victim response review, which either releases acceptable candidates or produces constrained alternatives with reduced harmful capability. Across three VLMs and five multimodal jailbreak families, CNRF achieves 86.06-89.30% macro defense success rate while retaining substantial benign-task utility. Controlled experiments further show that query-matched references and preserved context-direction correspondence materially improve risk ranking in the low-false-positive regime. CNRF thus reframes representation-based VLM safety from learning harmfulness boundaries to constructing applicable evidence for selective risk decisions.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20679 (HTTP 429)

Authors: Tianshi Wang, Jingsong Wang, Fengling Li, Jingjing Li, Xin Li, Lei Zhu

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20679.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20679

Published: 2026-10-19T01:39:59.081Z


7. IneqAgent: Agentic Inequality Proving in Lean with Symbolic Computation

Abstract:Formal inequality proving requires mathematical reasoning and exact computation, often connected through intermediate lemmas and goal reformulations. We introduce IneqAgent, an agentic framework combining LLM reasoning with symbolic computation for inequality proving in Lean4. Lemma-level Monte Carlo Tree Search coordinates lemma construction, goal reformulation, and symbolic backend selection. Complementary sum-of-squares, Bernstein, and successive difference-substitution backends produce exact certificates for local Lean proof generation. The agent integrates these proofs with LLM-generated steps and uses Lean feedback to construct complete proofs checked against the original theorem statements. On four benchmarks, IneqAgent with DeepSeek-V4-Flash solves 175 of 186 problems (94.09%), achieving the highest overall success rate among the evaluated methods. Across four backbones, it achieves higher success rates at low LLM inference cost thresholds than Direct Agent using the same backbone, with lower mean inference costs over each method’s successful runs. We also obtain Lean-verified strengthenings of Dittert’s conjecture in dimensions four and five, and certify the Holens—Djokovic $(4,4)$ inequality, implying Merris’s conjecture in dimension four.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20681 (HTTP 429)

Authors: Gaolei He, Jinhui Li, Ruobing Zuo, Xia Zeng, Zhengfeng Yang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20681.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20681

Published: 2026-10-19T01:39:59.081Z


8. L2G-SR: Learning to Guide LLM-based Symbolic Regression

Abstract:Symbolic regression searches for an equation that explains observed data. Numerical error is commonly used to guide the search, but it alone can favor an accurate surrogate over a candidate with the right structure. We propose L2G-SR, a framework for LLM-based symbolic regression that guides search with a learned structural verifier. We construct structural supervision signal using tree edit distance (TED), accounting for commutative operators and alternative expression forms. A verifier first aligns numerical observations with the formulas, then learns to score a candidate’s structural similarity to the target formula. At search time, the target formula is hidden and the verifier complements numerical error in parent selection. An analyst then uses verifier feedback and diagnostic tools to propose revisions, which a generator converts into refined expressions. On LLM-SRBench, L2G-SR achieves state-of-the-art numerical fit, equation level accuracy and TED across the evaluated LLM backbones and domains. L2G-SR demonstrates the value of learned structural guidance for LLM-based symbolic regression.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Q: 这篇论文试图解决什么问题? 该研究针对\*\*基于大语言模型(LLM)的符号回归(Symbolic Regression, SR)中搜索指导信号不足\*\*的问题,提出了一种结合数值拟合与结构相似性反馈的新框架。具体而言,其核心动机与待解决问题可归纳如下: ### 1. 数值误差作为唯一指导信号的局限性 现有符号回归方法普遍依赖数值拟合误差(如均方误差 L_(MSE) )来指导搜索和评估候选公式。然而,有限观测数据往往支持多种不同结构的表达式,导致低经验误差\*\*无法可靠地反映候选公式与目标公式在结构上的真实接近程度\*\*。正如文中所示: - x - x^3/6 在零点附近可近似 sin x ,但两者结构完全不同; - exp(1.05x) 与 exp(x) 结构形式一致,却可能累积显著的数值误差。 因此,\*\*仅凭数值指标难以区分“参数偏差但结构正确”与“数值巧合但结构错误”的候选公式\*\*,容易在搜索过程中偏好准确的替代结构(surrogate)而非具有正确结构的候选。 ### 2. 结构反馈在搜索过程中的缺失 在典型的 LLM-based 符号回归流程中,目标公式在搜索阶段是\*\*隐藏\*\*的。现有方法缺乏一种可在搜索时实时评估候选公式与未知目标之间\*\*结构相似度\*\*的机制。传统的树编辑距离(TED)虽然能够度量表达式结构差异,但未能充分考虑交换律、等价代数形式及常数抽象等符号回归场景中的特殊性。 ### 3. 需建立可学习的结构验证机制 为弥补上述缺口,该研究试图解决如何\*\*从观测数据和候选公式中预测其与隐藏目标的结构距离\*\*。这要求: - 设计一种\*\*改进的结构相似性度量\*\*,对常数取值、运算数顺序及常见代数重写具有鲁棒性; - 构建一个\*\*多模态验证器(verifier)\*\*,能够对齐数值观测与符号表达式的表征,并学习预测候选公式的结构距离; - 将该结构反馈\*\*集成到 LLM 驱动的搜索循环\*\*中,与数值误差共同指导父代选择、修订建议与候选保留。 简言之,该论文试图解决的核心问题是:\*\*在目标公式不可见的条件下,如何为基于 LLM 的符号回归提供可靠、可学习的结构指导信号,以突破单纯依赖数值误差的搜索瓶颈,从而发现结构更接近真实目标、且具有良好数值拟合的符号表达式。\*\* Q2: 有哪些相关研究? 根据论文第2节(Related Work),相关研究可分为以下两大方向: ## 1. 符号回归与神经搜索方法(Symbolic and Neural Search) ### 经典搜索与演化方法 - \*\*遗传编程与演化算法\*\*:包括经典的 Genetic Programming(Koza, 1992)、Eureqa(Schmidt & Lipson, 2009)、Operon(Burlacu et al., 2020)、PySR 与 GP-GOMEA(Cranmer, 2023; Virgolin et al., 2021)等,通过进化、物理结构或概率模型搜索表达式。 - \*\*物理启发与贝叶斯方法\*\*:如 AI Feynman(Udrescu & Tegmark, 2020; Udrescu et al., 2020)利用物理结构,以及 Bayesian SR(Jin et al., 2019)进行概率建模。 - \*\*基准测试\*\*:SRBench(La Cava et al., 2021)系统比较了各类代表性方法。 ### 学习驱动的生成与搜索 - \*\*策略学习\*\*:DSR 与 uDSR(Petersen et al., 2021; Landajuela et al., 2022)学习表达式生成策略。 - \*\*神经引导的搜索\*\*:neural-guided GP、DGSR-MCTS 与 TPSR(Mundhenk et al., 2021; Kamienny et al., 2023; Shojaee et al., 2023)将学习的提议分布与树搜索或遗传编程结合。 - \*\*预训练与端到端模型\*\*:在大规模合成语料上进行预训练以摊销发现成本(Biggio et al., 2021; Kamienny et al., 2022; Holt et al., 2023)。 ### 数值-符号联合表示 - \*\*多模态表示学习\*\*:SNIP(Meidani et al., 2024)、MDLformer(Yu et al., 2025)与 MOD-SR(Xiang et al., 2026)致力于连接数值观测与符号表达式的表示空间。L2G-SR 利用此类表示来评分候选公式的结构相似性,服务于外部搜索过程。 ## 2. LLM 引导的发现与结构比较(LLM-Guided Discovery and Structural Comparison) ### 基于 LLM 的公式发现 - \*\*基础 LLM-SR 框架\*\*:LLM-SR(Shojaee et al., 2025a)、LaSR(Grayeli et al., 2024)与 DrSR(Wang et al., 2025)利用大语言模型的先验知识,结合数值反馈来提出并改进表达式。 - \*\*增强型智能体系统\*\*:后续工作进一步引入诊断工具、可重用结构、多角色协作或引擎控制,如 DE(Pang et al., 2026)、RESTART(Li & Pan, 2026)、SR-Scientist(Xia et al., 2026)、A-SR(Zhao et al., 2026)等。L2G-SR 的区分度在于\*\*将监督学习的结构信号显式集成到选择、诊断与保留环节\*\*。 ### 结构比较与等价性 - \*\*树编辑距离\*\*:经典 TED(Zhang & Shasha, 1989)用于比较标记树之间的差异。 - \*\*结构评估基准\*\*:SRSD(Matsubara et al., 2024)与 PHYBench(Qiu et al., 2025)关注恢复出的表达式结构的评估。 - \*\*等价图表示\*\*:EGG-SR(Jiang et al., 2026)使用等式图(equality graph)编码代数等价关系。L2G-SR 则提出了一种\*\*改进的 TED\*\*,在抽象常数、匹配交换律操作数及比较多种代数视图的基础上,为结构监督提供可学习的监督信号。 Q3: 论文如何解决这个问题? 论文通过 \*\*L2G-SR\*\* 框架解决该问题,核心思路是引入\*\*可学习的结构验证器\*\*来补充传统的数值误差信号,从而指导 LLM 进行更有效的符号回归搜索。具体解决方案包含三个相互关联的组成部分: --- ### 1. 改进的树编辑距离(Modified TED)作为结构监督信号 为衡量候选公式与目标公式的结构相似性,论文提出了一种改进的归一化树编辑距离 d_(str)(f, g) ,使其对符号回归中常见的等价形式具有鲁棒性: - \*\*常数与数值抽象\*\*:将纯数值子树映射为 \`number\`,符号常数映射为 \`param\`,从而忽略具体常数值的差异;递归扁平化 \`Add\` 和 \`Mul\` 节点以消除结合性带来的分组差异。 - \*\*交换律操作数匹配\*\*:在加法或乘法根节点处,除有序编辑外,还通过最小代价分配(minimum-cost assignment)匹配子节点,降低对操作数顺序的敏感性。混合代价定义为:

D(hyb)(A, B) = min D(ord)(A, B),; c(a,b) + min(π) ∑(i=1)^(m) C(i,π(i))
其中 a, b ∈ Add, Mul , C
(ij) 为递归的混合代价或虚匹配代价 kappa(s) 。 - **多视图代数比较**:为容忍常见代数重写,比较三种视图下的距离并取最小值:
d(str)(f, g) = min(k ∈ 0,1,2) D_(hyb)(T(P_k^c(f)),, T(P_k^r(g)))|T(P_k^r(g))|
三种视图包括:结构保留的候选解析视图、普通求值视图(吸收常数)、以及展开视图(展开乘积与幂次)。 此外,子树插入/删除成本采用折扣函数:
kappa(s) = mins,; 0.6(s-5)+5
以避免大规模子树删除导致距离过度膨胀。 —- ### 2. 多模态 TED 验证器(Verifier) 验证器的作用是在**目标公式隐藏**的搜索阶段,仅根据观测数据 D 和候选公式 f ,预测其结构距离 d 。 #### 架构与输入 验证器首先将数值观测与符号公式对齐到共享嵌入空间: - **数值编码**:采用科学计数法,将每个数值编码为三个标记(符号、尾数、指数)。 - **公式编码**:在 MDLformer 前缀序列基础上,为每个标记添加**树地址(tree address)**以显式编码树结构:

a = WA [ ∑(ell=1)^(d) e(level)(ell) odot e(slot)(cell)max(d,1) + e(depth)(d) ]
其中 (c1, dots, c_d) 为根到该节点的路径。 - **回归摘要**:对每个变量计算 12 种简单输入-输出关系(如 log y sim log x_j 、 y sim x_j^2 等)的 Pearson 相关系数,并通过门控注意力机制聚合为 z_R 。 最终,验证器融合观测嵌入 z_D 、公式嵌入 z_f 和回归摘要 z_R ,通过融合头 hθ 输出预测距离:
d = h_θ([z_D;, z_f;, z_R])

两阶段训练 - **Stage 1(符号-数值对齐)**:使用对称批次对比学习,对齐观测 D_i 与其生成目标 f_i^star 的嵌入:

L(align) = -(1) / (2B)∑(i=1)^(B) [ log(exp(ui^top v_i/t)) / (∑(j=1)^(B)exp(ui^top v_j/t)) + log(exp(v_i^top u_i/t)) / (∑(j=1)^(B)exp(v_i^top u_j/t)) ]

  • **Stage 2(结构预测)**:在冻结的 Stage 1 编码器基础上,训练验证器预测归一化 TED。损失函数对远距离样本进行截断,防止预测值向均值聚集:
    L(d, d) = |d - d|, & d ≤ 1 max(0,, 1-d), & d > 1
    —- ### 3. TED 引导的 LLM 符号搜索(L2G-SR) 在搜索阶段,L2G-SR 将数值误差与验证器预测的结构距离结合,形成**双信号搜索循环**: #### 双信号父代选择 维护四个独立的候选池(islands)。每轮随机选择一岛,并以均等概率切换两种选择模式: - **MSE 轮**:在当前岛抽取若干公式,选择验证均方误差最低的作为父代; - **TED 轮**:选择预测 TED d 最低的作为父代:

f(parent) = argmin(f ∈ S) MSE(val)(f), & MSE round argmin(f ∈ S) d(f), & TED round

分析师与生成器代理 - **分析师(Analyst)**:接收父代的验证 MSE、预测 TED、残差模式及变换变量回归摘要。在 TED 轮中,分析师对比低 TED 公式与父代以识别值得保留的结构;在 MSE 轮中则关注残差 y - f(x) 的模式。分析师可调用结构比较工具或残差探测工具,并基于历史编辑结果向生成器提出**四条具体修订建议**。 - **生成器(Generator)**:另一 LLM 根据父代公式、分析师建议及引用的证据,生成四个完整的新公式。 #### 拟合、简化与记忆管理 - 对每个生成的公式,使用 BFGS 在拟合子集上优化未知常数; - 使用 SymPy 进行代数简化,并移除对预测贡献极小的项(避免引入虚假分支干扰后续搜索); - 将有效公式加入当前岛屿,并按验证 MSE 和预测 TED 分别进行岛内排名。当岛屿超过 64 个公式时,保留两个排名中各自前 32 名的公式。 搜索结束后,在验证 MSE 最优的容忍范围内,选择**最简单的公式**作为最终输出。 Q4: 论文做了哪些实验? 论文围绕 **LLM-SRBench** 基准开展了一系列实验,涵盖主性能对比、验证器效能分析、消融研究、案例研究及补充的分布内/外测试。具体实验内容如下: —- ### 1. 主实验:基准性能对比(Main Experimental Results) **设置** - **基准**:LLM-SRBench,共 240 条方程,分属 5 个领域:Transform(111)、Physics(44)、Materials(25)、Chemistry(36)、Biology(24)。 - **Backbone**:Llama3.1-8B-Instruct 与 Qwen3-4B-Instruct-2507。 - **基线**:LLMDirect、LLM-SR、LaSR、SGA、DE、A-SR、RESTART、SR-Scientist。 - **评价指标**: - 归一化均方误差 NMSE(越低越好); - 相对误差容忍度 0.01 下的准确率 Acc (0.01) (越高越好); - 真实归一化树编辑距离 TED(越低越好)。 **结果** - 在两种 backbone 及全部五个领域上,L2G-SR 的 NMSE 与 Acc (0.01) 均达到**第一或第二**(Table 1)。 - 最终所选公式的真实 TED 在每个领域、两种 backbone 下均为**最低**(Figure 3),表明其恢复的结构与目标公式最接近。 —- ### 2. 验证器分析(Verifier Analysis) **研究问题**:验证器预测 TED 的准确性如何?其结构反馈对下游搜索是否有实际增益? - **合成验证集表现(Table 2)**:在 Stage 2 保留的固定验证集上,检查点取得的验证目标值为 0.12613,对 d le 1 样本的 MAE 为 0.13086,对 d > 1 样本的召回率(预测值 ge 1 )为 52.63%。 - **搜索候选公式上的预测质量(Table 3)**:记录在真实搜索过程中出现的候选公式,验证器预测的 TED 与真实 TED 在不同领域上呈现出合理的均方误差与 Pearson 相关性。 - **下游搜索消融(Table 4)**:在 Physics 领域进行严格对照实验,将 TED 反馈替换为 validation MSE(其余设置完全一致:双角色生成、诊断工具、拟合与简化、预算 400 槽位等)。结果显示: - 去除 TED 后,NMSE 从 (2.702 ± 1.109) × 10^(-4) 恶化至 (2.264 ± 2.590) × 10^(-3) ; - Acc (0.01) 从 89.39% 降至 74.24% ; - 最终公式真实 TED 从 0.239 上升至 0.361 。 这表明,即便验证器在绝对数值上仅呈现中等相关性,其提供的**相对结构偏好**对搜索具有重要价值。 —- ### 3. 消融研究(Ablation Study) **RQ3:在匹配预算下,TED 反馈是否提升搜索?** - 如上所述,Table 4 的 Physics 消融实验确认:仅将结构通道中的预测 TED 替换为 validation MSE,就会显著损害数值拟合与结构恢复。 **RQ4:哪些验证器输入与训练阶段对预测至关重要?** - 在固定合成验证集上对比五种配置(Table 5): 1. **w/o tree address**:去除公式编码器的树地址嵌入; 2. **Formula only**:去除全部数值编码器,仅保留公式编码与融合头; 3. **w/o corr. branch**:去除 12 项回归相关性的分支; 4. **One-stage**:从零随机初始化直接优化 Stage 2 目标,跳过 Stage 1 的对比预训练; 5. **Full verifier**:完整模型。 结果中所有消融变体的验证目标与 MAE (le 1) 均劣于完整模型;去除数值输入、相关分支或 Stage 1 预训练均会降低对远距离样本的召回率。这说明**树地址、数值观测、回归摘要以及两阶段训练均为必要组件**。 —- ### 4. 案例研究:Physics 11(Case Study) 为直观展示 TED 反馈的作用,论文详细追踪了 Physics 11 问题第 57 轮的搜索过程(Figure 4): - 在 TED 轮中,系统选择了一个验证 NMSE 较高(1.019 vs. 0.0476)但预测 TED 更低(0.504 vs. 0.758)的父代。 - 基于该父代,LLM 分析师建议将阻尼正弦项与时空项分离并添加。 - 生成的子代最终达到测试 NMSE 2.09 × 10^(-15) 、Acc (0.01)=1 、真实 TED =0 ,而仅使用 MSE 的最终公式真实 TED 高达 0.965。 该案例表明,**结构反馈能够引导搜索跳出局部数值最优,发现真正正确的表达式结构**。 —- ### 5. 补充实验:分布内与分布外测试(Appendix E) **E.2 分布内(ID)振荡数据(Table 18)** - 在两个真实世界振荡问题(Oscillation 1、2)上评估。L2G-SR 在 Oscillation 1 上取得极低的 NMSE( 6.22 × 10^(-30) )与 100% 准确率;在 Oscillation 2 上也达到 100% 准确率。 **E.3–E.4 分布外(OOD)泛化** - **LSR-Synth OOD(Table 19)**:在输入分布偏移的测试集上,L2G-SR 在 Physics、Materials、Chemistry、Biology 四个领域的 NMSE 与 Acc (0.01) 均优于所对比的 A-SR、RESTART、SR-Scientist。 - **Oscillation OOD(Table 20)**:使用官方 OOD 分割的 10,000 条数据(不重新拟合常数),L2G-SR 在两个振荡问题上均取得最低的平均 NMSE 与最高的点准确率(Oscillation 2 为 91.24% )。 **E.5 基线复现与预算对照** - 报告了 RESTART 的复现配置(Table 21)以及各基线方法的搜索预算对比(Table 22),说明 L2G-SR 与 DE、RESTART 均使用 400 个 LLM 提议槽位,而 LLM-SR、LaSR、SGA 使用 1000 个。 **E.6 验证器消融细节** - 补充说明了 Table 5 中各变体的具体训练设置(参数冻结方式、初始化来源、训练步数等),确保消融的可复现性。 Q5: 有什么可以进一步探索的点? 基于论文结论与实验局限,以下方向值得进一步探索: —- ### 1. 更强大的大语言模型与后训练策略 论文受限于计算资源,仅在轻量级、低成本的 LLM(Llama3.1-8B-Instruct、Qwen3-4B-Instruct-2507)上进行了评估。未来可探索: - 将框架迁移至最新的大规模 LLM(如 Llama 4、GPT-4o、Claude 4 等),测试其是否能在更复杂的公式发现任务上带来进一步提升; - 探索**后训练(post-training)LLM 本身作为结构验证器**,而非仅训练一个外部 Transformer 验证器,使模型内在地具备结构距离判断能力; - 直接对**分析师(analyst)角色进行后训练**,以生成更高质量的公式修订建议,而非仅依赖提示工程。 —- ### 2. 结构相似性度量的扩展与完善 当前改进的 TED 已通过常数抽象、交换律匹配与三种代数视图容忍了部分等价形式,但仍存在拓展空间: - 纳入更广泛的**代数与解析等价规则**,如三角恒等式( sin^2 x + cos^2 x = 1 )、指数-对数恒等式、积分/导数关系等; - 探索基于**等式图(equality graph, e-graph)**的结构距离,以系统性地编码和计算更复杂的等价变换; - 研究 TED 归一化方式对具有高度不对称子树结构(如一个极简单目标与一个极复杂候选)的鲁棒性。 —- ### 3. 验证器的训练数据与泛化性增强 验证器目前依赖大规模合成数据训练,在真实 benchmark 上表现良好,但仍有提升余地: - 引入**真实科学观测数据或含噪声的实验数据**进行微调,检验验证器对测量误差、异常值及非理想采样分布的鲁棒性; - 探索**主动学习或课程学习策略**,让验证器在搜索过程中逐步适应目标领域的结构分布,而非仅使用预训练权重; - 扩展至**更高维输入空间**或**隐式微分方程/偏微分方程**的结构恢复任务,检验多模态表征的缩放性。 —- ### 4. 搜索算法的效率与可扩展性 L2G-SR 每轮涉及 LLM 推理、常数拟合(BFGS)、代数简化与验证器前向传播,计算开销显著: - 开发**轻量级验证器**(如知识蒸馏后的小模型或 MLP 替代方案),在保持相对排序能力的同时降低推理延迟; - 将 TED 反馈与**遗传编程、蒙特卡洛树搜索(MCTS)或贝叶斯优化**等非 LLM 搜索框架深度结合,测试结构验证信号在更广泛搜索范式中的通用性; - 设计**自适应的双信号权重调度机制**,根据搜索阶段(早期探索 vs. 后期微调)动态调整数值误差与结构反馈的优先级,而非固定的 1:1 轮次交替。 —- ### 5. 理论理解:为何相对结构偏好足以指导搜索? 实验表明,即便验证器的绝对 TED 预测仅呈现中等相关性(Table 3),其提供的相对排序仍能显著改进搜索(Table 4)。未来可深入研究: - 验证器预测的**排序准确度(ranking accuracy)**与搜索收敛速度之间的理论关系; - 在**组合优化与强化学习**框架下,分析双信号(数值 + 结构)反馈对避免局部最优和欺骗性拟合(deceptive fit)的作用机制; - 建立**样本复杂度界限**,量化合成训练数据的规模与多样性对验证器泛化性能的影响。 —- ### 6. 跨领域应用与可解释性拓展 - 将 L2G-SR 应用于**经济学、气候科学、系统生物学**等领域具有复杂约束(如量纲一致性、单调性、凹凸性)的符号发现任务; - 利用验证器的**注意力权重与回归摘要**提供可解释的诊断信息:不仅告诉用户“结构有多远”,还能指出“哪一部分子树或哪种函数关系与目标不符”,从而辅助科学家进行假设生成与模型选择。 Q6: 总结一下论文的主要内容 该论文提出 **L2G-SR**(Learning to Guide LLM-based Symbolic Regression),一个通过**可学习的结构验证器**来增强大语言模型(LLM)符号回归搜索的框架。以下从背景动机、核心技术、实验验证与结论展望四个方面进行总结。 —- ### 1. 背景与核心问题 符号回归(Symbolic Regression, SR)旨在从观测数据 D = (xi, y_i)(i=1)^n 中同时恢复目标函数 f^star 的**数学结构**与**数值常数**。现有方法(包括基于 LLM 的方法)普遍以**数值拟合误差**(如均方误差 L(MSE) )作为搜索与评估的主要信号。然而,数值误差存在固有缺陷: - **结构不同的公式可能数值接近**:例如 x - x^3/6 在原点附近可近似 sin x ,但结构完全不同; - **结构相似的公式可能数值偏离**:例如 exp(1.05x) 与 exp(x) 具有相同形式,但可累积显著误差。 因此,单纯依赖数值误差容易在搜索中偏好“数值巧合的替代结构”而非“结构正确的候选公式”,迫切需要一种**互补的结构指导信号**。 —- ### 2. 方法:L2G-SR 框架 论文通过三个相互连接的组件解决上述问题: #### (1)改进的树编辑距离(Modified TED) 为度量公式间的结构相似性,提出一种对常数取值、交换律及代数重写鲁棒的归一化 TED d(str)(f, g) : - **常数抽象**:将纯数值子树映射为 `number`,符号常数映射为 `param`,忽略具体常数值;扁平化 `Add`/`Mul` 节点以消除结合性差异。 - **交换律匹配**:对加法/乘法根节点,除有序编辑 D_(ord) 外,额外计算子节点最小代价分配:

D(hyb)(A, B) = min D(ord)(A, B),; c(a,b) + min(π) ∑(i=1)^(m) C(i,π(i))
其中 a,b ∈ Add, Mul 。 - **多视图比较**:通过三种预处理视图(结构保留、普通求值、展开)比较公式,取最小距离:
d
(str)(f, g) = min(k ∈ 0,1,2) D(hyb)(T(P_k^c(f)),, T(P_k^r(g)))|T(P_k^r(g))|

  • **子树折扣成本**:大子树的插入/删除成本采用 kappa(s) = mins,, 0.6(s-5)+5 ,避免距离过度膨胀。 #### (2)多模态 TED 验证器(Verifier) 在目标公式隐藏的搜索阶段,验证器仅从观测数据 D 和候选公式 f 预测其结构距离 d 。 - **输入编码**: - **数值观测**:以科学计数法分词(符号、尾数、指数)编码; - **候选公式**:在前缀序列基础上为每个标记注入**树地址**:

a = WA [ ∑(ell=1)^(d) e(level)(ell) odot e(slot)(cell)max(d,1) + e(depth)(d) ]
显式编码根到节点的路径; - **回归摘要**:计算 12 种输入-输出关系(如 log y sim log x_j , y sim x_j^2 )的 Pearson 相关系数,经门控注意力聚合为 z_R 。 - **两阶段训练**: - **Stage 1(对齐)**:通过对称批次对比学习,将观测嵌入 u_i 与生成目标嵌入 v_i 对齐; - **Stage 2(预测)**:在冻结编码器基础上,训练融合头预测归一化 TED,损失对远距离样本进行截断:
L(d, d) = |d - d|, & d ≤ 1 max(0,, 1-d), & d > 1

(3)双信号 LLM 搜索 搜索循环同时利用**数值误差**与**预测 TED**: - **父代选择**:维护四个候选池(islands),每轮以均等概率切换: - MSE 轮:选择验证 MSE 最低的公式; - TED 轮:选择预测 TED d 最低的公式:

f(parent) = argmin(f ∈ S) MSE(val)(f), & MSE round argmin(f ∈ S) d(f), & TED round

  • **分析师(Analyst)与生成器(Generator)**:分析师基于数值与结构反馈、残差模式及诊断工具提出四条修订建议;生成器据此产出新公式。 - **后处理**:BFGS 常数优化、SymPy 代数简化、移除贡献极小项,最终由验证器评分并更新记忆。 —- ### 3. 实验验证 论文在 **LLM-SRBench**(240 条方程,5 个领域)上开展实验,使用 Llama3.1-8B-Instruct 与 Qwen3-4B-Instruct-2507 作为 backbone。 #### 主实验结果 - **数值拟合**:L2G-SR 的 NMSE(越低越好)与 Acc (0.01) (越高越好)在所有领域和两种 backbone 上均达到**第一或第二**。 - **结构恢复**:最终公式的**真实 TED 在所有领域均为最低**(Figure 3),表明恢复的结构最接近目标。 #### 验证器与消融分析 - **验证器效能**(Table 3-4):在 Physics 领域的严格消融中,将 TED 反馈替换为 validation MSE(其余设置完全相同)后,NMSE 恶化约 8.4 倍,Acc (0.01) 从 89.39% 降至 74.24% ,真实 TED 从 0.239 升至 0.361 。这表明结构反馈是搜索性能的关键贡献者。 - **组件必要性**(Table 5):去除树地址、数值输入、回归摘要分支或 Stage 1 预训练,均会劣化验证器性能,证明各组件不可或缺。 #### 案例研究 在 Physics 11 的第 57 轮,TED 轮次选择了一个验证 NMSE 更高(1.019 vs. 0.0476)但预测 TED 更低(0.504 vs. 0.758)的父代。其子代最终达到测试 NMSE 2.09 × 10^(-15) 与真实 TED =0 ,而纯 MSE 导向的最终公式真实 TED 高达 0.965。这直观展示了结构反馈帮助搜索跳出局部数值最优、发现正确结构的能力。 #### 补充实验 - **分布内/外测试**:在 Oscillation 真实数据及 LSR-Synth OOD 场景下,L2G-SR 同样取得最优或接近最优的泛化性能(Appendix E)。 —- ### 4. 结论与未来方向 论文的核心结论是:**为基于 LLM 的符号回归引入可学习的结构引导,能够显著同时提升数值拟合精度与结构恢复准确度**。主要贡献包括: - 一种改进的 TED 度量,为符号回归提供超越数值拟合的结构监督; - 一个多模态验证器,使结构反馈在目标隐藏时仍可被获取; - 一个整合数值与结构双信号的 LLM 搜索代理。 未来工作可探索:采用更先进的 LLM 并进行后训练、将验证器训练扩展至真实科学数据、优化搜索效率与自适应信号调度,以及深化对“为何相对结构排序足以指导搜索”的理论理解。 Q7: 想要进一步了解论文 以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文,请点击 这里 为你跳转Kimi AI网页版,并启动一个与该论文相关的新会话。

Authors: Letian Huang, Xiangyu Wang, Haowei Lin, Di He

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20685.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20685

Published: 2026-10-19T01:39:59.081Z


9. When Reality Outpaces Cognition: A New Scaling Principle for Reliable Intelligence

Abstract:This paper challenges the understanding capabilities of existing AI paradigms and emerging world models by asking a fundamental question: How quickly can reality change before an intelligent system can no longer reliably perceive, reason, and act? We formulate this as a theory- and model-agnostic challenge in which systems must infer reliable actions for increasingly distant future states across a growing observation—reality gap. We operationalize this challenge through the effective Speed of Coherent Cognition ((S_C^{\mathrm{eff}})), defined as the highest tested Speed of Reality ((S_R)) through which a system continuously satisfies a predefined behavioural reliability criterion from baseline. As a proof of concept, reinforcement-learning agents are trained at the baseline (S_R) and evaluated, without retraining, as (S_R) is increased by up to (50\times) while computational resources remain fixed. We compare Transformer and gated-MLP baselines with (Co^4), an architecture inspired by cellular mechanisms implicated in conscious processing and awake thought. Across the tested scenarios, (Co^4) achieves substantially higher (S_C^{\mathrm{eff}}) than the baselines, motivating (S_C^{\mathrm{eff}}) as a new benchmark and the scaling of coherent cognition as a principle for reliable AI.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20764 (HTTP 429)

Authors: Ahsan Adeel, Talha Bin Riaz

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20764.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20764

Published: 2026-10-19T01:39:59.081Z


10. ARIA: Runtime Inspection and Amendment for Multi-Agent Systems

Abstract:In LLM-based multi-agent systems (MAS), a local mistake can propagate as other agents build on it, turning a correctable error into a failed task. Diagnosing completed runs or repeatedly re-executing them to search for corrections does not provide a learned policy for intervening while execution is still in progress. We present Agent Runtime Inspection and Amendment (ARIA), a framework for learning when and how to repair ongoing multi-agent executions. Our key intuition is to turn execution-verified repairs into reusable supervision, moving repair search into data construction and training so that a learned model can intervene at runtime without repeatedly testing alternatives. First, we collect over 24{,}000 successful and failed trajectories across four multi-agent frameworks, six benchmarks, and eight host models. From these trajectories, we construct the ARIA dataset, to our knowledge the first large-scale dataset pairing failed multi-agent trajectories with execution-verified repairs. It contains 3{,}953 verified repair pairs, each specifying what went wrong, where to intervene, and how execution should continue. Then, we develop the ARIA framework for MAS-agnostic runtime inspection and amendment, designed to integrate with any MAS that exposes execution trajectories and supports continuation replacement. Using verified repairs as training examples, we teach a compact model to recognize when intervention is needed and how to correct an ongoing run. The aim is to learn repair strategies that apply to related failures in new configurations, rather than merely reproduce individual corrections. Finally, the trained ARIA model demonstrates strong repair performance in both post-hoc and runtime settings, with generalization to unseen frameworks and host models. Our code is available at this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20780 (HTTP 429)

Authors: Bolin Shen, Yushun Dong

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20780.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20780

Published: 2026-10-19T01:39:59.081Z


Evaluation Domain Papers

1. The World Won’t Wait: Evaluating and Teaching Agents Under a Live Clock

Abstract:Large language model (LLM) agents can solve complex interactive tasks, yet reasoning and tool calls take time. In continuously changing environments, an otherwise correct action may arrive too late. We introduce RealtimeBench, with 180 tasks across six game types, to evaluate agents while games continue running during inference. Paired evaluations show that GPT-6.1-sol outperforms the faster Jev when games are paused, but falls behind when they keep running. We propose Two-Timescale Offline Recursive Self-Improvement (TORSI): a slow teacher studies failures and tests corrections in paused games between attempts, then writes reusable instructions for a fast student. The student acts without teacher calls during gameplay, and validation scores determine which updates are adopted. Two rounds raise Jev’s average test score from 13.51 to 24.19 on a 0—100 progress scale, improving all six game types. The instructions also improve average scores for three other models without retuning. These findings point toward real-time intelligence: enabling powerful LLM agents to remain effective in a continuously changing world by organizing deliberation, learning, and action across timescales.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20665 (HTTP 429)

Authors: Ningxin Pan, Hanyu Li, Yeqing Shen, Ning Ding, Yehui Tang, Wentao Zhang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20665.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20665

Published: 2026-10-19T01:40:11.761Z


2. Evaluating Model Capabilities After Benchmark Saturation

Abstract:Rapid advances in AI capabilities are causing benchmarks to saturate. Once several models approach perfect scores, a benchmark loses its power to differentiate between them, and with it its utility for informing AI safety cases and frontier policies. Yet benchmark items produce far more information than the scores we typically extract from them: a full transcript of how each model solved the task. We investigate whether these transcripts can be used to assess model capabilities even after benchmark scores have saturated. We simulate saturation retrospectively using historical benchmark and model releases, and test whether protocols that access only correctly solved transcripts from saturated benchmarks can predict capabilities measured on held-out, unsaturated items. We evaluate representative contemporary approaches for extracting information from model transcripts: pairwise judging, single-transcript absolute scoring, and predictions from deterministic transcript features or embeddings. We find that saturated transcripts retain capability information beyond benchmark accuracy. Amongst our evaluated protocols, pairwise judging performs best, reducing prediction error by 5.3 times compared to a baseline using only saturated benchmark scores. This capability-relevant information is distributed across transcript sections, but concentrated in specific channels of information such as reasoning and tool-calls. However, the predictive power of protocols does not always generalise across distribution shifts — for example by underestimating the capabilities of the first reasoning models introduced in early 2025. Saturated benchmarks thus retain capability-predicting signal in their task transcripts, but extending the lifetime of existing benchmarks requires methods designed to extract these signals robustly.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20667 (HTTP 429)

Authors: Morio Hamada, Jehyeok Yeon, Matteo Migliarini, Justin Olive, Magda Dubois

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20667.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20667

Published: 2026-10-19T01:40:11.761Z


3. Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion

Abstract:Reward-guided discrete diffusion enables test-time alignment with downstream objectives, but applying Sequential Monte Carlo (SMC) guidance at every denoising step is computationally expensive. Active guidance steps require additional reward evaluations and, for gradient-based proposals, value-gradient computations. Sparse guidance reduces this cost by guiding only a subset of timesteps, but commonly used schedules are fixed heuristics whose effectiveness can vary substantially across tasks. We propose VISTA, a training-free method that automatically selects guided timesteps under a fixed budget. VISTA formulates sparse guidance as approximation of the ideal full-guidance probability path and derives an estimable upper bound on (\ell_1) path deviation. Using value statistics collected from a small number of full-guidance warmup generations, VISTA optimizes the resulting empirical surrogate exactly via dynamic programming. The warmup generations can be retained as model outputs, and after schedule selection VISTA has the same per-generation guidance complexity as other sparse schedules with the same budget. Across text, biological sequence, and image generation tasks, VISTA discovers task-dependent schedules and outperforms hand-designed sparse schedules in most tested settings, while approaching or exceeding full-guidance performance on several task metrics. In toxic-text generation, for example, VISTA attains a 0.99 unique-toxic ratio compared with 1.00 for full guidance while reducing post-selection per-generation latency from 204.23 s to 38.16 s, a 5.35$\times$ speedup.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20669 (HTTP 429)

Authors: Jaehyeon Kim, Ziseok Lee, Siyoung Choi, Kyungsu Kim

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20669.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20669

Published: 2026-10-19T01:40:11.761Z


4. AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks

Abstract:Agents pair reasoning models with harnesses that organize model calls, execute tools, load skill instructions, and retain history in memory. Understanding agent performance requires linking execution conditions to outcomes, behavior, and resource use. We introduce AlphaDiana, a measurement system with modular interfaces for datasets, agents, environments, and scorers. It preserves harness-native control loops while controlling file reuse and memory retention separately. AlphaDiana supports macro analyses of complete agents and micro interventions on agent components such as tools, skills, and memory under matched conditions. We study three open-source models and three open-source harnesses on five answer-based benchmarks, with two models also tested on two interactive benchmarks. We find that harnesses do not consistently improve the accuracy of reasoning: gains on questions unsolved by the model alone can be outweighed by losses on questions it answers correctly without tools. This inconsistency extends to component interventions: neither skill guidance nor retained history reliably improves accuracy. Behavior and output-token use must be assessed with correctness: low token entropy can accompany wrong answers, and less output per execution can mean more output tokens per correct answer when accuracy falls. AlphaDiana links answer correctness and execution status to traces and resource records for comparing agents and diagnosing failures under shared task and resource policies.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20675 (HTTP 429)

Authors: Zhanke Zhou, Zongze Li, Weikai Huang, Xuan Li, Chentao Cao, Xiao Feng, Xiangyu Lu, Jinbo Hu, Yi Xie, Nico Pelleriti, Shiyang Liu, Menghan Lu, Max Zimmer, Brando Miranda, Jiangchao Yao, Bo Liu, Sanmi Koyejo, Sebastian Pokutta, Bo Han

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20675.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20675

Published: 2026-10-19T01:40:11.761Z


5. When Correctness Feedback Erases Evidence: Phase-Separated Completeness Auditing in Verus

Abstract:Correctness feedback suppresses the very evidence that specification-completeness auditing is designed to uncover: programs that satisfy a written contract but violate intended behavior. We demonstrate this objective reversal on 12 Verus/Rust tasks across four model families under matched call budgets. Independent no-feedback discovery more than doubles certified-witness yield relative to IntentShield’s corrective-feedback loop (59.0% versus 27.1%). A format-matched iterative control places the larger share of this gap in feedback content, not independent restarts. The advantage holds in every model family and survives parser repair. A separate phase-separated execution matches the aggregate yield without reusing responses; most certified witnesses also fail held-out cases. Preserving these witnesses delivers stronger specification repair: Phase-Revised contracts achieve 81.3% implementation success, versus 62.5% for IS-Revised, with Direct-Revised close behind. Both no-feedback witness sources substantially outperform weakened contracts. The effect does not reproduce in the available Dafny and Lean evaluations, establishing an empirical boundary around the studied Verus/Rust workflow. Completeness auditing and correctness repair have opposing immediate objectives. Preserve and certify the admitted failure before repairing it.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20676 (HTTP 429)

Authors: Adarsh Agrawal

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20676.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20676

Published: 2026-10-19T01:40:11.761Z


6. Counterfactual Neighborhood Risk Fields: Evidence-Grounded Selective Safety for Vision-Language Models

Abstract:Representation-based safety defenses for vision-language models (VLMs) often learn to separate harmful from benign inputs, but separation alone does not establish what a risk score means or when it should be trusted. We introduce Counterfactual Neighborhood Risk Fields (CNRF), an evidence-grounded framework for selective VLM safety. CNRF constructs context-matched benign-harmful counterfactual pairs, uses their representation differences as capability-related risk directions, and estimates neighborhood support separately from risk. This separation prevents unfamiliarity from being conflated with harmfulness and enables selective intervention rather than blanket rejection. A victim-adaptive interface accommodates different safety observations across VLM architectures and routes selected requests for same-victim response review, which either releases acceptable candidates or produces constrained alternatives with reduced harmful capability. Across three VLMs and five multimodal jailbreak families, CNRF achieves 86.06-89.30% macro defense success rate while retaining substantial benign-task utility. Controlled experiments further show that query-matched references and preserved context-direction correspondence materially improve risk ranking in the low-false-positive regime. CNRF thus reframes representation-based VLM safety from learning harmfulness boundaries to constructing applicable evidence for selective risk decisions.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20679 (HTTP 429)

Authors: Tianshi Wang, Jingsong Wang, Fengling Li, Jingjing Li, Xin Li, Lei Zhu

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20679.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20679

Published: 2026-10-19T01:40:11.761Z


7. IneqAgent: Agentic Inequality Proving in Lean with Symbolic Computation

Abstract:Formal inequality proving requires mathematical reasoning and exact computation, often connected through intermediate lemmas and goal reformulations. We introduce IneqAgent, an agentic framework combining LLM reasoning with symbolic computation for inequality proving in Lean4. Lemma-level Monte Carlo Tree Search coordinates lemma construction, goal reformulation, and symbolic backend selection. Complementary sum-of-squares, Bernstein, and successive difference-substitution backends produce exact certificates for local Lean proof generation. The agent integrates these proofs with LLM-generated steps and uses Lean feedback to construct complete proofs checked against the original theorem statements. On four benchmarks, IneqAgent with DeepSeek-V4-Flash solves 175 of 186 problems (94.09%), achieving the highest overall success rate among the evaluated methods. Across four backbones, it achieves higher success rates at low LLM inference cost thresholds than Direct Agent using the same backbone, with lower mean inference costs over each method’s successful runs. We also obtain Lean-verified strengthenings of Dittert’s conjecture in dimensions four and five, and certify the Holens—Djokovic $(4,4)$ inequality, implying Merris’s conjecture in dimension four.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20681 (HTTP 429)

Authors: Gaolei He, Jinhui Li, Ruobing Zuo, Xia Zeng, Zhengfeng Yang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20681.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20681

Published: 2026-10-19T01:40:11.761Z


8. L2G-SR: Learning to Guide LLM-based Symbolic Regression

Abstract:Symbolic regression searches for an equation that explains observed data. Numerical error is commonly used to guide the search, but it alone can favor an accurate surrogate over a candidate with the right structure. We propose L2G-SR, a framework for LLM-based symbolic regression that guides search with a learned structural verifier. We construct structural supervision signal using tree edit distance (TED), accounting for commutative operators and alternative expression forms. A verifier first aligns numerical observations with the formulas, then learns to score a candidate’s structural similarity to the target formula. At search time, the target formula is hidden and the verifier complements numerical error in parent selection. An analyst then uses verifier feedback and diagnostic tools to propose revisions, which a generator converts into refined expressions. On LLM-SRBench, L2G-SR achieves state-of-the-art numerical fit, equation level accuracy and TED across the evaluated LLM backbones and domains. L2G-SR demonstrates the value of learned structural guidance for LLM-based symbolic regression.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Q: 这篇论文试图解决什么问题? 该研究针对\*\*基于大语言模型(LLM)的符号回归(Symbolic Regression, SR)中搜索指导信号不足\*\*的问题,提出了一种结合数值拟合与结构相似性反馈的新框架。具体而言,其核心动机与待解决问题可归纳如下: ### 1. 数值误差作为唯一指导信号的局限性 现有符号回归方法普遍依赖数值拟合误差(如均方误差 L_(MSE) )来指导搜索和评估候选公式。然而,有限观测数据往往支持多种不同结构的表达式,导致低经验误差\*\*无法可靠地反映候选公式与目标公式在结构上的真实接近程度\*\*。正如文中所示: - x - x^3/6 在零点附近可近似 sin x ,但两者结构完全不同; - exp(1.05x) 与 exp(x) 结构形式一致,却可能累积显著的数值误差。 因此,\*\*仅凭数值指标难以区分“参数偏差但结构正确”与“数值巧合但结构错误”的候选公式\*\*,容易在搜索过程中偏好准确的替代结构(surrogate)而非具有正确结构的候选。 ### 2. 结构反馈在搜索过程中的缺失 在典型的 LLM-based 符号回归流程中,目标公式在搜索阶段是\*\*隐藏\*\*的。现有方法缺乏一种可在搜索时实时评估候选公式与未知目标之间\*\*结构相似度\*\*的机制。传统的树编辑距离(TED)虽然能够度量表达式结构差异,但未能充分考虑交换律、等价代数形式及常数抽象等符号回归场景中的特殊性。 ### 3. 需建立可学习的结构验证机制 为弥补上述缺口,该研究试图解决如何\*\*从观测数据和候选公式中预测其与隐藏目标的结构距离\*\*。这要求: - 设计一种\*\*改进的结构相似性度量\*\*,对常数取值、运算数顺序及常见代数重写具有鲁棒性; - 构建一个\*\*多模态验证器(verifier)\*\*,能够对齐数值观测与符号表达式的表征,并学习预测候选公式的结构距离; - 将该结构反馈\*\*集成到 LLM 驱动的搜索循环\*\*中,与数值误差共同指导父代选择、修订建议与候选保留。 简言之,该论文试图解决的核心问题是:\*\*在目标公式不可见的条件下,如何为基于 LLM 的符号回归提供可靠、可学习的结构指导信号,以突破单纯依赖数值误差的搜索瓶颈,从而发现结构更接近真实目标、且具有良好数值拟合的符号表达式。\*\* Q2: 有哪些相关研究? 根据论文第2节(Related Work),相关研究可分为以下两大方向: ## 1. 符号回归与神经搜索方法(Symbolic and Neural Search) ### 经典搜索与演化方法 - \*\*遗传编程与演化算法\*\*:包括经典的 Genetic Programming(Koza, 1992)、Eureqa(Schmidt & Lipson, 2009)、Operon(Burlacu et al., 2020)、PySR 与 GP-GOMEA(Cranmer, 2023; Virgolin et al., 2021)等,通过进化、物理结构或概率模型搜索表达式。 - \*\*物理启发与贝叶斯方法\*\*:如 AI Feynman(Udrescu & Tegmark, 2020; Udrescu et al., 2020)利用物理结构,以及 Bayesian SR(Jin et al., 2019)进行概率建模。 - \*\*基准测试\*\*:SRBench(La Cava et al., 2021)系统比较了各类代表性方法。 ### 学习驱动的生成与搜索 - \*\*策略学习\*\*:DSR 与 uDSR(Petersen et al., 2021; Landajuela et al., 2022)学习表达式生成策略。 - \*\*神经引导的搜索\*\*:neural-guided GP、DGSR-MCTS 与 TPSR(Mundhenk et al., 2021; Kamienny et al., 2023; Shojaee et al., 2023)将学习的提议分布与树搜索或遗传编程结合。 - \*\*预训练与端到端模型\*\*:在大规模合成语料上进行预训练以摊销发现成本(Biggio et al., 2021; Kamienny et al., 2022; Holt et al., 2023)。 ### 数值-符号联合表示 - \*\*多模态表示学习\*\*:SNIP(Meidani et al., 2024)、MDLformer(Yu et al., 2025)与 MOD-SR(Xiang et al., 2026)致力于连接数值观测与符号表达式的表示空间。L2G-SR 利用此类表示来评分候选公式的结构相似性,服务于外部搜索过程。 ## 2. LLM 引导的发现与结构比较(LLM-Guided Discovery and Structural Comparison) ### 基于 LLM 的公式发现 - \*\*基础 LLM-SR 框架\*\*:LLM-SR(Shojaee et al., 2025a)、LaSR(Grayeli et al., 2024)与 DrSR(Wang et al., 2025)利用大语言模型的先验知识,结合数值反馈来提出并改进表达式。 - \*\*增强型智能体系统\*\*:后续工作进一步引入诊断工具、可重用结构、多角色协作或引擎控制,如 DE(Pang et al., 2026)、RESTART(Li & Pan, 2026)、SR-Scientist(Xia et al., 2026)、A-SR(Zhao et al., 2026)等。L2G-SR 的区分度在于\*\*将监督学习的结构信号显式集成到选择、诊断与保留环节\*\*。 ### 结构比较与等价性 - \*\*树编辑距离\*\*:经典 TED(Zhang & Shasha, 1989)用于比较标记树之间的差异。 - \*\*结构评估基准\*\*:SRSD(Matsubara et al., 2024)与 PHYBench(Qiu et al., 2025)关注恢复出的表达式结构的评估。 - \*\*等价图表示\*\*:EGG-SR(Jiang et al., 2026)使用等式图(equality graph)编码代数等价关系。L2G-SR 则提出了一种\*\*改进的 TED\*\*,在抽象常数、匹配交换律操作数及比较多种代数视图的基础上,为结构监督提供可学习的监督信号。 Q3: 论文如何解决这个问题? 论文通过 \*\*L2G-SR\*\* 框架解决该问题,核心思路是引入\*\*可学习的结构验证器\*\*来补充传统的数值误差信号,从而指导 LLM 进行更有效的符号回归搜索。具体解决方案包含三个相互关联的组成部分: --- ### 1. 改进的树编辑距离(Modified TED)作为结构监督信号 为衡量候选公式与目标公式的结构相似性,论文提出了一种改进的归一化树编辑距离 d_(str)(f, g) ,使其对符号回归中常见的等价形式具有鲁棒性: - \*\*常数与数值抽象\*\*:将纯数值子树映射为 \`number\`,符号常数映射为 \`param\`,从而忽略具体常数值的差异;递归扁平化 \`Add\` 和 \`Mul\` 节点以消除结合性带来的分组差异。 - \*\*交换律操作数匹配\*\*:在加法或乘法根节点处,除有序编辑外,还通过最小代价分配(minimum-cost assignment)匹配子节点,降低对操作数顺序的敏感性。混合代价定义为:

D(hyb)(A, B) = min D(ord)(A, B),; c(a,b) + min(π) ∑(i=1)^(m) C(i,π(i))
其中 a, b ∈ Add, Mul , C
(ij) 为递归的混合代价或虚匹配代价 kappa(s) 。 - **多视图代数比较**:为容忍常见代数重写,比较三种视图下的距离并取最小值:
d(str)(f, g) = min(k ∈ 0,1,2) D_(hyb)(T(P_k^c(f)),, T(P_k^r(g)))|T(P_k^r(g))|
三种视图包括:结构保留的候选解析视图、普通求值视图(吸收常数)、以及展开视图(展开乘积与幂次)。 此外,子树插入/删除成本采用折扣函数:
kappa(s) = mins,; 0.6(s-5)+5
以避免大规模子树删除导致距离过度膨胀。 —- ### 2. 多模态 TED 验证器(Verifier) 验证器的作用是在**目标公式隐藏**的搜索阶段,仅根据观测数据 D 和候选公式 f ,预测其结构距离 d 。 #### 架构与输入 验证器首先将数值观测与符号公式对齐到共享嵌入空间: - **数值编码**:采用科学计数法,将每个数值编码为三个标记(符号、尾数、指数)。 - **公式编码**:在 MDLformer 前缀序列基础上,为每个标记添加**树地址(tree address)**以显式编码树结构:

a = WA [ ∑(ell=1)^(d) e(level)(ell) odot e(slot)(cell)max(d,1) + e(depth)(d) ]
其中 (c1, dots, c_d) 为根到该节点的路径。 - **回归摘要**:对每个变量计算 12 种简单输入-输出关系(如 log y sim log x_j 、 y sim x_j^2 等)的 Pearson 相关系数,并通过门控注意力机制聚合为 z_R 。 最终,验证器融合观测嵌入 z_D 、公式嵌入 z_f 和回归摘要 z_R ,通过融合头 hθ 输出预测距离:
d = h_θ([z_D;, z_f;, z_R])

两阶段训练 - **Stage 1(符号-数值对齐)**:使用对称批次对比学习,对齐观测 D_i 与其生成目标 f_i^star 的嵌入:

L(align) = -(1) / (2B)∑(i=1)^(B) [ log(exp(ui^top v_i/t)) / (∑(j=1)^(B)exp(ui^top v_j/t)) + log(exp(v_i^top u_i/t)) / (∑(j=1)^(B)exp(v_i^top u_j/t)) ]

  • **Stage 2(结构预测)**:在冻结的 Stage 1 编码器基础上,训练验证器预测归一化 TED。损失函数对远距离样本进行截断,防止预测值向均值聚集:
    L(d, d) = |d - d|, & d ≤ 1 max(0,, 1-d), & d > 1
    —- ### 3. TED 引导的 LLM 符号搜索(L2G-SR) 在搜索阶段,L2G-SR 将数值误差与验证器预测的结构距离结合,形成**双信号搜索循环**: #### 双信号父代选择 维护四个独立的候选池(islands)。每轮随机选择一岛,并以均等概率切换两种选择模式: - **MSE 轮**:在当前岛抽取若干公式,选择验证均方误差最低的作为父代; - **TED 轮**:选择预测 TED d 最低的作为父代:

f(parent) = argmin(f ∈ S) MSE(val)(f), & MSE round argmin(f ∈ S) d(f), & TED round

分析师与生成器代理 - **分析师(Analyst)**:接收父代的验证 MSE、预测 TED、残差模式及变换变量回归摘要。在 TED 轮中,分析师对比低 TED 公式与父代以识别值得保留的结构;在 MSE 轮中则关注残差 y - f(x) 的模式。分析师可调用结构比较工具或残差探测工具,并基于历史编辑结果向生成器提出**四条具体修订建议**。 - **生成器(Generator)**:另一 LLM 根据父代公式、分析师建议及引用的证据,生成四个完整的新公式。 #### 拟合、简化与记忆管理 - 对每个生成的公式,使用 BFGS 在拟合子集上优化未知常数; - 使用 SymPy 进行代数简化,并移除对预测贡献极小的项(避免引入虚假分支干扰后续搜索); - 将有效公式加入当前岛屿,并按验证 MSE 和预测 TED 分别进行岛内排名。当岛屿超过 64 个公式时,保留两个排名中各自前 32 名的公式。 搜索结束后,在验证 MSE 最优的容忍范围内,选择**最简单的公式**作为最终输出。 Q4: 论文做了哪些实验? 论文围绕 **LLM-SRBench** 基准开展了一系列实验,涵盖主性能对比、验证器效能分析、消融研究、案例研究及补充的分布内/外测试。具体实验内容如下: —- ### 1. 主实验:基准性能对比(Main Experimental Results) **设置** - **基准**:LLM-SRBench,共 240 条方程,分属 5 个领域:Transform(111)、Physics(44)、Materials(25)、Chemistry(36)、Biology(24)。 - **Backbone**:Llama3.1-8B-Instruct 与 Qwen3-4B-Instruct-2507。 - **基线**:LLMDirect、LLM-SR、LaSR、SGA、DE、A-SR、RESTART、SR-Scientist。 - **评价指标**: - 归一化均方误差 NMSE(越低越好); - 相对误差容忍度 0.01 下的准确率 Acc (0.01) (越高越好); - 真实归一化树编辑距离 TED(越低越好)。 **结果** - 在两种 backbone 及全部五个领域上,L2G-SR 的 NMSE 与 Acc (0.01) 均达到**第一或第二**(Table 1)。 - 最终所选公式的真实 TED 在每个领域、两种 backbone 下均为**最低**(Figure 3),表明其恢复的结构与目标公式最接近。 —- ### 2. 验证器分析(Verifier Analysis) **研究问题**:验证器预测 TED 的准确性如何?其结构反馈对下游搜索是否有实际增益? - **合成验证集表现(Table 2)**:在 Stage 2 保留的固定验证集上,检查点取得的验证目标值为 0.12613,对 d le 1 样本的 MAE 为 0.13086,对 d > 1 样本的召回率(预测值 ge 1 )为 52.63%。 - **搜索候选公式上的预测质量(Table 3)**:记录在真实搜索过程中出现的候选公式,验证器预测的 TED 与真实 TED 在不同领域上呈现出合理的均方误差与 Pearson 相关性。 - **下游搜索消融(Table 4)**:在 Physics 领域进行严格对照实验,将 TED 反馈替换为 validation MSE(其余设置完全一致:双角色生成、诊断工具、拟合与简化、预算 400 槽位等)。结果显示: - 去除 TED 后,NMSE 从 (2.702 ± 1.109) × 10^(-4) 恶化至 (2.264 ± 2.590) × 10^(-3) ; - Acc (0.01) 从 89.39% 降至 74.24% ; - 最终公式真实 TED 从 0.239 上升至 0.361 。 这表明,即便验证器在绝对数值上仅呈现中等相关性,其提供的**相对结构偏好**对搜索具有重要价值。 —- ### 3. 消融研究(Ablation Study) **RQ3:在匹配预算下,TED 反馈是否提升搜索?** - 如上所述,Table 4 的 Physics 消融实验确认:仅将结构通道中的预测 TED 替换为 validation MSE,就会显著损害数值拟合与结构恢复。 **RQ4:哪些验证器输入与训练阶段对预测至关重要?** - 在固定合成验证集上对比五种配置(Table 5): 1. **w/o tree address**:去除公式编码器的树地址嵌入; 2. **Formula only**:去除全部数值编码器,仅保留公式编码与融合头; 3. **w/o corr. branch**:去除 12 项回归相关性的分支; 4. **One-stage**:从零随机初始化直接优化 Stage 2 目标,跳过 Stage 1 的对比预训练; 5. **Full verifier**:完整模型。 结果中所有消融变体的验证目标与 MAE (le 1) 均劣于完整模型;去除数值输入、相关分支或 Stage 1 预训练均会降低对远距离样本的召回率。这说明**树地址、数值观测、回归摘要以及两阶段训练均为必要组件**。 —- ### 4. 案例研究:Physics 11(Case Study) 为直观展示 TED 反馈的作用,论文详细追踪了 Physics 11 问题第 57 轮的搜索过程(Figure 4): - 在 TED 轮中,系统选择了一个验证 NMSE 较高(1.019 vs. 0.0476)但预测 TED 更低(0.504 vs. 0.758)的父代。 - 基于该父代,LLM 分析师建议将阻尼正弦项与时空项分离并添加。 - 生成的子代最终达到测试 NMSE 2.09 × 10^(-15) 、Acc (0.01)=1 、真实 TED =0 ,而仅使用 MSE 的最终公式真实 TED 高达 0.965。 该案例表明,**结构反馈能够引导搜索跳出局部数值最优,发现真正正确的表达式结构**。 —- ### 5. 补充实验:分布内与分布外测试(Appendix E) **E.2 分布内(ID)振荡数据(Table 18)** - 在两个真实世界振荡问题(Oscillation 1、2)上评估。L2G-SR 在 Oscillation 1 上取得极低的 NMSE( 6.22 × 10^(-30) )与 100% 准确率;在 Oscillation 2 上也达到 100% 准确率。 **E.3–E.4 分布外(OOD)泛化** - **LSR-Synth OOD(Table 19)**:在输入分布偏移的测试集上,L2G-SR 在 Physics、Materials、Chemistry、Biology 四个领域的 NMSE 与 Acc (0.01) 均优于所对比的 A-SR、RESTART、SR-Scientist。 - **Oscillation OOD(Table 20)**:使用官方 OOD 分割的 10,000 条数据(不重新拟合常数),L2G-SR 在两个振荡问题上均取得最低的平均 NMSE 与最高的点准确率(Oscillation 2 为 91.24% )。 **E.5 基线复现与预算对照** - 报告了 RESTART 的复现配置(Table 21)以及各基线方法的搜索预算对比(Table 22),说明 L2G-SR 与 DE、RESTART 均使用 400 个 LLM 提议槽位,而 LLM-SR、LaSR、SGA 使用 1000 个。 **E.6 验证器消融细节** - 补充说明了 Table 5 中各变体的具体训练设置(参数冻结方式、初始化来源、训练步数等),确保消融的可复现性。 Q5: 有什么可以进一步探索的点? 基于论文结论与实验局限,以下方向值得进一步探索: —- ### 1. 更强大的大语言模型与后训练策略 论文受限于计算资源,仅在轻量级、低成本的 LLM(Llama3.1-8B-Instruct、Qwen3-4B-Instruct-2507)上进行了评估。未来可探索: - 将框架迁移至最新的大规模 LLM(如 Llama 4、GPT-4o、Claude 4 等),测试其是否能在更复杂的公式发现任务上带来进一步提升; - 探索**后训练(post-training)LLM 本身作为结构验证器**,而非仅训练一个外部 Transformer 验证器,使模型内在地具备结构距离判断能力; - 直接对**分析师(analyst)角色进行后训练**,以生成更高质量的公式修订建议,而非仅依赖提示工程。 —- ### 2. 结构相似性度量的扩展与完善 当前改进的 TED 已通过常数抽象、交换律匹配与三种代数视图容忍了部分等价形式,但仍存在拓展空间: - 纳入更广泛的**代数与解析等价规则**,如三角恒等式( sin^2 x + cos^2 x = 1 )、指数-对数恒等式、积分/导数关系等; - 探索基于**等式图(equality graph, e-graph)**的结构距离,以系统性地编码和计算更复杂的等价变换; - 研究 TED 归一化方式对具有高度不对称子树结构(如一个极简单目标与一个极复杂候选)的鲁棒性。 —- ### 3. 验证器的训练数据与泛化性增强 验证器目前依赖大规模合成数据训练,在真实 benchmark 上表现良好,但仍有提升余地: - 引入**真实科学观测数据或含噪声的实验数据**进行微调,检验验证器对测量误差、异常值及非理想采样分布的鲁棒性; - 探索**主动学习或课程学习策略**,让验证器在搜索过程中逐步适应目标领域的结构分布,而非仅使用预训练权重; - 扩展至**更高维输入空间**或**隐式微分方程/偏微分方程**的结构恢复任务,检验多模态表征的缩放性。 —- ### 4. 搜索算法的效率与可扩展性 L2G-SR 每轮涉及 LLM 推理、常数拟合(BFGS)、代数简化与验证器前向传播,计算开销显著: - 开发**轻量级验证器**(如知识蒸馏后的小模型或 MLP 替代方案),在保持相对排序能力的同时降低推理延迟; - 将 TED 反馈与**遗传编程、蒙特卡洛树搜索(MCTS)或贝叶斯优化**等非 LLM 搜索框架深度结合,测试结构验证信号在更广泛搜索范式中的通用性; - 设计**自适应的双信号权重调度机制**,根据搜索阶段(早期探索 vs. 后期微调)动态调整数值误差与结构反馈的优先级,而非固定的 1:1 轮次交替。 —- ### 5. 理论理解:为何相对结构偏好足以指导搜索? 实验表明,即便验证器的绝对 TED 预测仅呈现中等相关性(Table 3),其提供的相对排序仍能显著改进搜索(Table 4)。未来可深入研究: - 验证器预测的**排序准确度(ranking accuracy)**与搜索收敛速度之间的理论关系; - 在**组合优化与强化学习**框架下,分析双信号(数值 + 结构)反馈对避免局部最优和欺骗性拟合(deceptive fit)的作用机制; - 建立**样本复杂度界限**,量化合成训练数据的规模与多样性对验证器泛化性能的影响。 —- ### 6. 跨领域应用与可解释性拓展 - 将 L2G-SR 应用于**经济学、气候科学、系统生物学**等领域具有复杂约束(如量纲一致性、单调性、凹凸性)的符号发现任务; - 利用验证器的**注意力权重与回归摘要**提供可解释的诊断信息:不仅告诉用户“结构有多远”,还能指出“哪一部分子树或哪种函数关系与目标不符”,从而辅助科学家进行假设生成与模型选择。 Q6: 总结一下论文的主要内容 该论文提出 **L2G-SR**(Learning to Guide LLM-based Symbolic Regression),一个通过**可学习的结构验证器**来增强大语言模型(LLM)符号回归搜索的框架。以下从背景动机、核心技术、实验验证与结论展望四个方面进行总结。 —- ### 1. 背景与核心问题 符号回归(Symbolic Regression, SR)旨在从观测数据 D = (xi, y_i)(i=1)^n 中同时恢复目标函数 f^star 的**数学结构**与**数值常数**。现有方法(包括基于 LLM 的方法)普遍以**数值拟合误差**(如均方误差 L(MSE) )作为搜索与评估的主要信号。然而,数值误差存在固有缺陷: - **结构不同的公式可能数值接近**:例如 x - x^3/6 在原点附近可近似 sin x ,但结构完全不同; - **结构相似的公式可能数值偏离**:例如 exp(1.05x) 与 exp(x) 具有相同形式,但可累积显著误差。 因此,单纯依赖数值误差容易在搜索中偏好“数值巧合的替代结构”而非“结构正确的候选公式”,迫切需要一种**互补的结构指导信号**。 —- ### 2. 方法:L2G-SR 框架 论文通过三个相互连接的组件解决上述问题: #### (1)改进的树编辑距离(Modified TED) 为度量公式间的结构相似性,提出一种对常数取值、交换律及代数重写鲁棒的归一化 TED d(str)(f, g) : - **常数抽象**:将纯数值子树映射为 `number`,符号常数映射为 `param`,忽略具体常数值;扁平化 `Add`/`Mul` 节点以消除结合性差异。 - **交换律匹配**:对加法/乘法根节点,除有序编辑 D_(ord) 外,额外计算子节点最小代价分配:

D(hyb)(A, B) = min D(ord)(A, B),; c(a,b) + min(π) ∑(i=1)^(m) C(i,π(i))
其中 a,b ∈ Add, Mul 。 - **多视图比较**:通过三种预处理视图(结构保留、普通求值、展开)比较公式,取最小距离:
d
(str)(f, g) = min(k ∈ 0,1,2) D(hyb)(T(P_k^c(f)),, T(P_k^r(g)))|T(P_k^r(g))|

  • **子树折扣成本**:大子树的插入/删除成本采用 kappa(s) = mins,, 0.6(s-5)+5 ,避免距离过度膨胀。 #### (2)多模态 TED 验证器(Verifier) 在目标公式隐藏的搜索阶段,验证器仅从观测数据 D 和候选公式 f 预测其结构距离 d 。 - **输入编码**: - **数值观测**:以科学计数法分词(符号、尾数、指数)编码; - **候选公式**:在前缀序列基础上为每个标记注入**树地址**:

a = WA [ ∑(ell=1)^(d) e(level)(ell) odot e(slot)(cell)max(d,1) + e(depth)(d) ]
显式编码根到节点的路径; - **回归摘要**:计算 12 种输入-输出关系(如 log y sim log x_j , y sim x_j^2 )的 Pearson 相关系数,经门控注意力聚合为 z_R 。 - **两阶段训练**: - **Stage 1(对齐)**:通过对称批次对比学习,将观测嵌入 u_i 与生成目标嵌入 v_i 对齐; - **Stage 2(预测)**:在冻结编码器基础上,训练融合头预测归一化 TED,损失对远距离样本进行截断:
L(d, d) = |d - d|, & d ≤ 1 max(0,, 1-d), & d > 1

(3)双信号 LLM 搜索 搜索循环同时利用**数值误差**与**预测 TED**: - **父代选择**:维护四个候选池(islands),每轮以均等概率切换: - MSE 轮:选择验证 MSE 最低的公式; - TED 轮:选择预测 TED d 最低的公式:

f(parent) = argmin(f ∈ S) MSE(val)(f), & MSE round argmin(f ∈ S) d(f), & TED round

  • **分析师(Analyst)与生成器(Generator)**:分析师基于数值与结构反馈、残差模式及诊断工具提出四条修订建议;生成器据此产出新公式。 - **后处理**:BFGS 常数优化、SymPy 代数简化、移除贡献极小项,最终由验证器评分并更新记忆。 —- ### 3. 实验验证 论文在 **LLM-SRBench**(240 条方程,5 个领域)上开展实验,使用 Llama3.1-8B-Instruct 与 Qwen3-4B-Instruct-2507 作为 backbone。 #### 主实验结果 - **数值拟合**:L2G-SR 的 NMSE(越低越好)与 Acc (0.01) (越高越好)在所有领域和两种 backbone 上均达到**第一或第二**。 - **结构恢复**:最终公式的**真实 TED 在所有领域均为最低**(Figure 3),表明恢复的结构最接近目标。 #### 验证器与消融分析 - **验证器效能**(Table 3-4):在 Physics 领域的严格消融中,将 TED 反馈替换为 validation MSE(其余设置完全相同)后,NMSE 恶化约 8.4 倍,Acc (0.01) 从 89.39% 降至 74.24% ,真实 TED 从 0.239 升至 0.361 。这表明结构反馈是搜索性能的关键贡献者。 - **组件必要性**(Table 5):去除树地址、数值输入、回归摘要分支或 Stage 1 预训练,均会劣化验证器性能,证明各组件不可或缺。 #### 案例研究 在 Physics 11 的第 57 轮,TED 轮次选择了一个验证 NMSE 更高(1.019 vs. 0.0476)但预测 TED 更低(0.504 vs. 0.758)的父代。其子代最终达到测试 NMSE 2.09 × 10^(-15) 与真实 TED =0 ,而纯 MSE 导向的最终公式真实 TED 高达 0.965。这直观展示了结构反馈帮助搜索跳出局部数值最优、发现正确结构的能力。 #### 补充实验 - **分布内/外测试**:在 Oscillation 真实数据及 LSR-Synth OOD 场景下,L2G-SR 同样取得最优或接近最优的泛化性能(Appendix E)。 —- ### 4. 结论与未来方向 论文的核心结论是:**为基于 LLM 的符号回归引入可学习的结构引导,能够显著同时提升数值拟合精度与结构恢复准确度**。主要贡献包括: - 一种改进的 TED 度量,为符号回归提供超越数值拟合的结构监督; - 一个多模态验证器,使结构反馈在目标隐藏时仍可被获取; - 一个整合数值与结构双信号的 LLM 搜索代理。 未来工作可探索:采用更先进的 LLM 并进行后训练、将验证器训练扩展至真实科学数据、优化搜索效率与自适应信号调度,以及深化对“为何相对结构排序足以指导搜索”的理论理解。 Q7: 想要进一步了解论文 以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文,请点击 这里 为你跳转Kimi AI网页版,并启动一个与该论文相关的新会话。

Authors: Letian Huang, Xiangyu Wang, Haowei Lin, Di He

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20685.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20685

Published: 2026-10-19T01:40:11.761Z


9. When Reality Outpaces Cognition: A New Scaling Principle for Reliable Intelligence

Abstract:This paper challenges the understanding capabilities of existing AI paradigms and emerging world models by asking a fundamental question: How quickly can reality change before an intelligent system can no longer reliably perceive, reason, and act? We formulate this as a theory- and model-agnostic challenge in which systems must infer reliable actions for increasingly distant future states across a growing observation—reality gap. We operationalize this challenge through the effective Speed of Coherent Cognition ((S_C^{\mathrm{eff}})), defined as the highest tested Speed of Reality ((S_R)) through which a system continuously satisfies a predefined behavioural reliability criterion from baseline. As a proof of concept, reinforcement-learning agents are trained at the baseline (S_R) and evaluated, without retraining, as (S_R) is increased by up to (50\times) while computational resources remain fixed. We compare Transformer and gated-MLP baselines with (Co^4), an architecture inspired by cellular mechanisms implicated in conscious processing and awake thought. Across the tested scenarios, (Co^4) achieves substantially higher (S_C^{\mathrm{eff}}) than the baselines, motivating (S_C^{\mathrm{eff}}) as a new benchmark and the scaling of coherent cognition as a principle for reliable AI.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20764 (HTTP 429)

Authors: Ahsan Adeel, Talha Bin Riaz

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20764.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20764

Published: 2026-10-19T01:40:11.761Z


10. ARIA: Runtime Inspection and Amendment for Multi-Agent Systems

Abstract:In LLM-based multi-agent systems (MAS), a local mistake can propagate as other agents build on it, turning a correctable error into a failed task. Diagnosing completed runs or repeatedly re-executing them to search for corrections does not provide a learned policy for intervening while execution is still in progress. We present Agent Runtime Inspection and Amendment (ARIA), a framework for learning when and how to repair ongoing multi-agent executions. Our key intuition is to turn execution-verified repairs into reusable supervision, moving repair search into data construction and training so that a learned model can intervene at runtime without repeatedly testing alternatives. First, we collect over 24{,}000 successful and failed trajectories across four multi-agent frameworks, six benchmarks, and eight host models. From these trajectories, we construct the ARIA dataset, to our knowledge the first large-scale dataset pairing failed multi-agent trajectories with execution-verified repairs. It contains 3{,}953 verified repair pairs, each specifying what went wrong, where to intervene, and how execution should continue. Then, we develop the ARIA framework for MAS-agnostic runtime inspection and amendment, designed to integrate with any MAS that exposes execution trajectories and supports continuation replacement. Using verified repairs as training examples, we teach a compact model to recognize when intervention is needed and how to correct an ongoing run. The aim is to learn repair strategies that apply to related failures in new configurations, rather than merely reproduce individual corrections. Finally, the trained ARIA model demonstrates strong repair performance in both post-hoc and runtime settings, with generalization to unseen frameworks and host models. Our code is available at this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20780 (HTTP 429)

Authors: Bolin Shen, Yushun Dong

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.20780.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20780

Published: 2026-10-19T01:40:11.761Z


VLM Domain Papers

1. BD-HazardVLM: Probing Vision-Language Models for Latent Defensive-Driving Hazards in Bangladesh Road Scenes

Abstract:Vision-language models are increasingly used for driving-scene understanding, yet their defensive reasoning under weakly structured Global South traffic remains unclear. We introduce BD-HazardVLM, a diagnostic benchmark of 500 Bangladesh road-scene images, comprising 375 from RSUD20K and 125 from TFP-BD, annotated for hazard presence, primary hazard category, risk timing, severity, defensive actions, visible evidence, and hallucination- and Western-road-bias traps. We evaluate Qwen2.5-VL-7B, InternVL3-8B, and MiniCPM-V-4.6 using a shared zero-shot task definition and structured output schema. Binary hazard-presence accuracy ranges from 88.0 percent to 89.2 percent, essentially matching the 89.0 percent majority-class baseline. Fine-grained reasoning is substantially weaker: category accuracy remains below 40 percent, risk-timing accuracy ranges from 1.6 percent to 12.4 percent, and exact defensive-action matching remains below 2 percent. Although InternVL3-8B achieves the highest severity accuracy at 61.2 percent, it has zero recall on the 58 high-severity cases. Qwen exhibits severe category mode collapse, while all models show weak latent-hazard anticipation and near-zero self-reported trap recognition. For the evaluated models, these findings show that coarse hazard recognition does not imply reliable fine-grained defensive-driving reasoning and motivate geographically diverse safety evaluation.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20627 (HTTP 429)

Authors: Abdullah Al Maruf, Md. Sajedul Islam, Tanmoy Mridha, Irfan Hossain Bhuiyan

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20627.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20627

Published: 2026-10-19T01:40:24.655Z


2. Towards automatic digital garment assembly

Abstract:Creating digital garments from existing two-dimensional patterns remains a time-consuming process. Although patterns are widely available in digital form, their assembly in garment CAD software involves manual definition of sewing operations and arrangement of the pattern panels for drape simulation. We present two learning-based approaches that automate both tasks using the 2D vector pattern as their sole input. Our automatic sewing model operates at the segment level, leveraging a Transformer encoder with multiple prediction heads. Our novel architecture supports many-to-many, ordered segment correspondences, as well as hems, overlapping seams needed for multilayer garments, and construction properties such as seam allowances and stitch types, required for the finished virtual garment but neglected by previous methods. A complementary panel-level model predicts body-region assignments, layers, and placement coordinates to initialize the drape simulation. We evaluate both models on a curated dataset of 4,035 professionally assembled garment instances. The sewing model exactly recovers all seam and hem topologies of 63.3\% of test garments, while the arrangement model meets the specified placement criteria for all panels in 66.2\% of test garments. We also integrate both methods into garment CAD software, demonstrating automatic assembly of successfully predicted garments.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20651 (HTTP 429)

Authors: Alejandro Rodriguez

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20651.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20651

Published: 2026-10-19T01:40:24.655Z


3. Orbia: Do Generated Videos Form a Consistent 3D World?

Abstract:Video generators serve as interactive worlds for exploration and simulation, yet realistic and visually appealing frames do not necessarily depict a consistent 3D world. Existing benchmarks assess selected geometric properties or use vision-language models to judge memory and consistency, leaving systematic 3D evaluation underexplored. We introduce Orbia, the first benchmark dedicated to a systematic and comprehensive evaluation of 3D consistency in interactive world models. Through controlled exploration and revisit, we organize the evaluation into three dimensions: input preservation, generated-content persistence, and 3D self-consistency. For input preservation, registered RGB-D references and annotated anchors support appearance, geometry, and spatial stability measurements. For generated-content persistence, we compare revisits to regions absent from the input image with their first generated observations to test whether models preserve what they create. For 3D self-consistency, cross-view depth and feature correspondence, regional scale agreement, and held-out Gaussian rendering test whether generated views describe the same 3D scene. Minute-long exploration and revisits further test long-term consistency. Our construction and annotation pipeline brings together public datasets, self-captured scenes, and simulation, yielding 720 instances across indoor and outdoor environments and six exploration-and-revisit protocols. Evaluation spans 27 text-, action-, and camera-conditioned models through native-interface adapters. Our Orbia benchmark provides a systematic basis for diagnosing 3D inconsistency and developing more persistent video world models.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20693 (HTTP 429)

Authors: Dongyue Lu, Guian Fang, Hur Lim, Yonggang Wu, Yanrong Wang, Hongshan Chen, Shengju Qian, Yue Liao, Wei Chow, Lingdong Kong, Tianxin Huang, Yu Yang, Hongji Yang, Junchao Huang, Ao Liang, Yihao Wang, Zihan Wang, Rong Li, Hanlin Chen, Xin Wang, Wei Tsang Ooi, Mike Zheng Shou, Shuicheng Yan

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20693.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20693

Published: 2026-10-19T01:40:24.655Z


4. WorldCast: Exploring Intelligent Visual Generation in World State Prediction

Abstract:Humans interpret multimodal observations of the world and forecast how it will evolve. Recent multimodal generative systems have demonstrated similar capabilities: they can integrate multimodal context and synthesize visual outputs that depict changes in the world. To characterize this emerging trajectory, we organize model capabilities into a four-level hierarchy and discover that higher-level tasks remain relatively underexplored despite model progress. To bridge this gap, we explore model performance on multi-step world state prediction, a high-level task that requires models to forecast successive future world states from multimodal observations and diverse references. Such a task requires models to possess entangled understanding and generation capabilities, as well as extensive world knowledge to translate human action sequences, physical transformations, and abstract visual reasoning into coherent visual outputs. We introduce WorldCast, a benchmark comprising 320 manually designed cases across eight categories, and evaluate 12 representative models. The conducted experiments yield three key insights: (a) advances in frontier models are rapidly pushing the boundaries of intelligent world modeling, with the GPT-Image-2.5 System scoring 72.8 and demonstrating early-stage visual state prediction capabilities on real-world tasks; (b) WorldCast reveals substantial gaps between models: even Nano Banana Pro reaches only 42.8 and the best open-source model scores 20.1, with deficits in handling complex context, entangled references, and reasoning-heavy tasks; (c) strong generative models effectively draw on the capabilities of their underlying language models. These findings highlight the need for multimodal generative models that are capable of dynamic planning and reasoning while also being deeply unified to faithfully simulate real-world states.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20731 (HTTP 429)

Authors: Wenhao Yan, Haodong Li, Xiaotong Li, Yaxin Luo, Zhengrong Yue, Jialv Zou, Xu Huang, Haojin Yang, Nan Jia, Yan Feng, Peng Pei, Xunliang Cai

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20731.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20731

Published: 2026-10-19T01:40:24.655Z


5. Purify Before You Align: Improving Representation Alignment for Diffusion Models

Abstract:Representation alignment improves and accelerates diffusion training by encouraging intermediate features to match a pretrained visual encoder. However, standard alignment matches pretrained representations without distinguishing what information may be more useful for generation. We introduce local appearance purification (LAP), which improves generation quality and makes encoders such as MAE more effective alignment targets by removing locally predictable content and aligning to the residual. LAP modifies only the alignment target, and requires no changes to diffusion training or inference. Through controlled interventions within a fixed encoder, we show that effective alignment depends on what information the target emphasizes, not simply how much image information it contains. Our theory gives a variational interpretation of standard alignment and, for fixed local predictors, a sum of token-wise information bounds conditioned on local appearance. To understand why purification benefits some encoders more than others, we derive an exact constant-readout test that measures how well the target can be matched without using the diffusion model’s hidden state. LAP improves generation quality across five pretrained representations (MAE, AIMv2, CLIP, DINOv2, DINOv3), with the largest gain for MAE: on ImageNet-256 with SiT-B/2, it reduces FID from $53.1$ to $42.3$. Its gains persist with stronger alignment recipes and tokenizers. Computed before diffusion training, the constant-readout score predicts the benefit of purification (Spearman $\rho=0.94$ across 13 encoders).

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20761 (HTTP 429)

Authors: Yingheng Wang, Yaoqiang Li, Yaqin Wu, Junwen Bai, Jiatao Gu, Christopher De Sa, Volodymyr Kuleshov

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20761.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20761

Published: 2026-10-19T01:40:24.655Z


6. ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models

Abstract:Image-to-text (I2T) understanding and text-to-image (T2I) generation are dual tasks in unified multimodal models (UMMs). Reconstructive post-training couples these tasks, but existing reconstruction-based methods remain perception-centric and provide limited supervision for reasoning about structure, intent, and transformation. We propose ReasonUni, a reasoning-augmented reconstructive post-training framework that extends this coupling beyond perception. Given an original image and an editing instruction, the UMM first reasons about what and how to edit and describes the target state, then generates the corresponding target image conditioned on this understanding output. To jointly optimize both modules, we introduce TreeGRPO, a tree-structured reinforcement learning method that samples multiple reasoning paths and candidate images per path, then aggregates visual rewards to update both modules under a unified objective. Experiments across multiple benchmarks demonstrate consistent gains in both understanding and generation, showing that reasoning-augmented reconstruction yields genuine mutual gains for UMMs.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20781 (HTTP 429)

Authors: Feize Wu, Zhiyuan Yan, Yuhuan Hong, Yuwei Niu, Zongjian Li, Zheyuan Liu, Bin Lin, Jinbo Yan, Jie Qin, Xudong Mao, Limeng Qiao, Guanglu Wan, Li Yuan

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20781.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20781

Published: 2026-10-19T01:40:24.655Z


7. Blame Errant Words, Not the Canvas: Defect-Aware Spatial Credit Assignment for Visual Text Generation

Abstract:Recent reinforcement learning approaches have improved visual text generation, yet generated images still suffer from incorrect wording and malformed glyphs. We identify two coupled limitations: recognition-based rewards often overlook visible glyph defects, while whole-image scalar feedback obscures the uneven quality of individual text regions. These limitations motivate our central principle: $\textit{Blame errant words, not the canvas}$, to effectively preserve local quality distinctions. First, we introduce $\mathbf{TextDART}$, a defect-aware region-targeted reward model that localizes individual text regions and independently evaluates their semantic correctness and glyph quality, providing fine-grained, region-precise feedback. Building on TextDART, we propose $\mathbf{SCANNER}$, which turns these regional judgments into spatially differentiated policy feedback. It normalizes rewards for each text target across sampled images and casts the resulting regional advantages onto the corresponding latent tokens. Extensive experiments on existing benchmarks and our newly established $\textbf{PolyText-Bench}$ demonstrate that our approach achieves state-of-the-art performance in text accuracy and glyph fidelity. Project page: this https URL

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20794 (HTTP 429)

Authors: Sujie Hu, Huaisong Zhang, Pujing Yang, Qian Wang, Xiaomin Li, Xinhang Leng, Chen Li, Jing Lyu, Xiu Li, Ling Pan

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20794.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20794

Published: 2026-10-19T01:40:24.655Z


8. RIGen: Beyond Action Following with Reciprocal Interaction in Generative Gameplay

Abstract:Generative gameplay requires more than immediate responses to player this http URL should have persistent consequences, while the generated world should act according to its own state and rules, shaping the player’s situation and subsequent choices. Video generation captures local dynamics, but limited history, partially observed game state, and the need for contextual behavior decisions make such interactions difficult to sustain through audiovisual prediction alone. We introduce \rig (Reciprocally Interactive Generation), a framework for reciprocal interaction in generative gameplay. \rig combines persistent state and configurable rules, contextual interaction reasoning, and action-conditioned audiovisual generation. Together, these components make interaction consequences and actor behavior jointly determine generated content while maintaining consistency between game state and visible outcomes. We construct a Xonotic gameplay dataset to train an audiovisual generator for learning dynamics and interactions. Further post-training enables \rigarena, an interactive generative first-person shooter (FPS) prototype. Experiments demonstrate more reliable combat outcomes and reciprocal interaction between the player and the generated world. Project page: this https URL.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Q: 这篇论文试图解决什么问题? 该论文旨在解决生成式游戏玩法(generative gameplay)中\*\*双向交互(reciprocal interaction)\*\*的可持续实现问题。现有基于视频生成的世界模型虽能捕捉局部动态并响应玩家控制,但仅依靠视听预测难以支撑具有以下特征的复杂游戏交互: - \*\*动作需产生持久后果\*\*:玩家行为应改变世界中对象与角色的持久状态; - \*\*世界需依自身逻辑反作用于玩家\*\*:场景规则、角色目标与当前状态应自主决定世界行为,进而约束玩家后续选择与生存状态。 具体而言,论文指出当前方法面临三项核心挑战: 1. \*\*协调双向交互(Coordinating reciprocal interaction)\*\* 响应玩家输入本身无法决定世界如何对玩家采取行动。角色必须依据当前情境与战略目标自主规划行为(例如低血量时撤退),且交互后果需实时更新玩家状态(如生命值、弹药量)并限制后续可执行动作。 2. \*\*有限历史(Limited history)\*\* 视频模型的上下文窗口有限,而超出窗口的早期事件仍可能对当前交互产生持久影响。例如,在持续战斗中,累积伤害必须决定下一次攻击是否致命,即使先前攻击已不在视听上下文中。 3. \*\*部分可观察状态(Partially observed state)\*\* 生命值、弹药量等关键游戏变量往往无法直接从可见场景中恢复,但这些变量决定了动作是否可执行及其应产生的结果。 为应对上述挑战,论文提出 \*\*RIGen\*\*(Reciprocally Interactive Generation)框架,通过耦合以下三个层次形成闭环: - \*\*状态与规则层\*\*:维护持久游戏变量(如血量、弹药、策略模式)与可配置转移规则; - \*\*交互推理层\*\*:基于场景与玩家动作推断事件后果,并规划角色自适应行为; - \*\*音视频生成层\*\*:在玩家控制与交互指导下实现局部动态与视听反馈。 由此,玩家动作改变持久世界状态,世界依据状态与规则自主决策并反向影响玩家,从而在生成式游戏中实现状态依赖的后果、角色行为与视听表现的一致性。 Q2: 有哪些相关研究? 论文在第2节"Related Work"及相关段落中,将现有研究归纳为三个主要方向: ### 1. 视频生成与交互世界(Video Generation and Interactive Worlds) 该方向关注从视频/视听基础模型到可玩游戏环境的演进: - \*\*基础模型\*\*:如 HunyuanVideo、Wan、MM-Diffusion、Movie Gen、LTX-2、OmniForcing 等,为外观与动态提供先验。 - \*\*可控视觉模拟到生成式游戏\*\*:早期 Playable Video Generation/Environments(Kim et al., 2020; Menapace et al., 2021, 2022)发展为生成式交互环境,包括 Genie(Bruce et al., 2024)、GameNGen(Valevski et al., 2025)、DIAMOND(Alonso et al., 2024)、Oasis(Decart et al., 2024)、WHAM/Muse(Kanervisto et al., 2025)、MineWorld(Guo et al., 2025)等。 - \*\*交互式内容创作\*\*:GameGen-X(Che et al., 2025)、GameFactory(Yu et al., 2025)、Hunyuan-GameCraft(Li et al., 2025b)等探索开放世界游戏视频生成。 - \*\*场景扩展与流式控制\*\*:SCOPE(Tong et al., 2026)、LingBot-World(Robbyant Team et al., 2026)、SANA-WM(Zhu et al., 2026a)、SolarWM(Huang et al., 2026a)等扩大场景覆盖并支持流式控制。 - \*\*多人/音视频交互\*\*:Multiplayer Interactive World Models(Hu et al., 2026)、EchoWM(Zhang et al., 2026)加入多人或全模态交互。 - \*\*长程一致性记忆\*\*:WorldMem(Xiao et al., 2025)、Infinite-World(Wu et al., 2026)、WorldPlay(Sun et al., 2025)、Matrix-Game 3.0(Wang et al., 2026b)、AlayaWorld(AlayaWorld Team et al., 2026)等通过记忆机制延长一致性。 \*\*与RIGen的区别\*\*:上述工作主要关注局部动态、导航控制或长程一致性,尚未充分解决"世界依自身状态与规则反作用于玩家"这一双向交互问题。 ### 2. 角色控制与语义交互(Actor Control and Semantic Interaction) 该方向关注在生成环境中引入可控角色与世界事件: - \*\*反应式对手与语言引导控制\*\*:COMBAT(Agarwal et al., 2026)、ReactiveGWM(Wang et al., 2026a)、ShadowDancer(Cao et al., 2026)、HelloWorld(Ouyang et al., 2026)、Hunyuan-GameCraft-2(Tang et al., 2025)等扩展了玩家驱动的生成与社交互动角色。 - \*\*LingBot-World 2.0\*\*(Gao et al., 2026):使用 pilot 与 director agents 进行角色控制和场景事件调度。 - \*\*Incantation\*\*(Zhu et al., 2026b):在生成后追踪伤害,将命中/未命中与死亡响应留给视频模型处理,存在快节奏战斗中死亡响应延迟的风险。 - \*\*EchoWM\*\*(Zhang et al., 2026):支持开放式全模态世界模型。 \*\*与RIGen的区别\*\*:这些系统提供了角色控制或语义事件指导,但缺乏与可执行游戏规则的紧密耦合——即行为规划、状态更新与视觉反馈的闭环 reconciliation。 ### 3. 显式状态与生成式渲染(Explicit State and Generative Rendering) 该方向尝试将结构化状态或游戏逻辑与生成式渲染结合: - \*\*状态+生成渲染\*\*:Unbounded(Li et al., 2025a)、AnimeGamer(Cheng et al., 2025)、Model as a Game(Chen et al., 2025)、WildWorld(Li et al., 2026)、StatePlay(Lin et al., 2026b)、MASS(Cai et al., 2026)等将显式状态或逻辑嵌入生成流程。 - \*\*基于规划的视频生成\*\*:VideoDirectorGPT(Lin et al., 2024)、LLM-grounded Video Diffusion(Lian et al., 2024)在合成前组织内容。 - \*\*几何渲染接口\*\*:Generative World Renderer(Huang et al., 2026c; Lin et al., 2026a)、Marionette(Meng et al., 2026)提供显式空间控制。 - \*\*Programmable World Model\*\*(Huang et al., 2026b):将状态增强的3D边界框编译为渲染控制。 - \*\*Code World Model\*\*(Chen et al., 2026):将可执行状态编译为代理视频(proxy videos)。 - \*\*PixVerse Game Engine\*\*(PixVerse Research, 2026):在概念层面描述了游戏机制、智能体编排与生成式表达的分离。 \*\*与RIGen的区别\*\*:这些系统或通过显式空间轨迹、或通过代理视频/3D表示来桥接状态与生成;而RIGen选择通过\*\*玩家动作\*\*与\*\*局部交互指导\*\*直接连接游戏状态与视听生成,无需构造显式空间轨迹,并强调交互后果、角色行为与视听反馈的实时闭环对齐。 Q3: 论文如何解决这个问题? 论文通过 \*\*RIGen(Reciprocally Interactive Generation)\*\* 框架解决生成式游戏中的双向交互问题。该框架将持久状态管理、交互推理与动作条件化的音视频生成耦合为闭环系统,使玩家动作能够改变持久世界状态,同时世界依据自身逻辑与角色目标反向影响玩家。 核心解决方案可分为以下层面: ### 1. 三层耦合架构 RIGen 由三个紧密协作的层次构成(图 2、图 3): - \*\*状态与规则引擎(State and Rule Engine)\*\*:维护持久的游戏变量(玩家/敌人生命值、弹药量、生死状态、对手策略模式等),并基于可配置规则执行状态转移; - \*\*交互代理(Interaction Agent)\*\*:解析场景与玩家输入,推断动作后果,规划角色自适应行为,并对齐生成视频与状态; - \*\*音视频生成器(Audiovisual Generator)\*\*:在玩家控制与交互指导下建模局部视听动态,并将生成结果作为观测反馈给上层。 ### 2. 持久状态与可执行规则 区别于纯视觉推理,RIGen 将游戏玩法显式建模为事件驱动的状态机。引擎维护数值化状态 s ,并根据交互事件记录 e 按规则计算后继状态:

s’ = Rules(s, e)
例如,射击需满足”玩家存活且弹药充足”的前提,命中后按配置伤害扣减目标生命值,生命值为零时触发死亡;若对手低血量幸存,规则可将其策略从”交战”切换为”撤退”,进而约束下一时刻的交互选择。 ### 3. 交互推理:先验后果预测与视觉校正 在生成第 k 个音视频块 ok 之前,系统先执行**验证与状态校正(reconciliation)**,再**预测当前后果**,而非事后补救: 1. **后验校正**:基于上一块的采样验证帧 I(k-1) 、玩家控制 u(k-1) 与局部指导 p(k-1) ,验证实际生成的视觉结果,并修正计划状态:

f(k-1) = Verify(I(k-1), u(k-1), p(k-1)), quad sk = Correct(s_k^(plan), f(k-1))

  1. **先验推理**:利用校正后的状态 sk 、当前玩家控制 u_k 与反馈 f(k-1) ,交互代理提出当前块的事件 ek 与行为 b_k :
    (e_k, b_k) = Agent(o
    (k-1), sk, u_k, f(k-1))
  2. **规则预演**:引擎计算临时后继状态与局部指导,但不立即覆盖 sk :
    s
    (k+1)^(plan) = Rules(sk, e_k), quad p_k = Prompt(s(k+1)^(plan), ek, b_k)
    生成完成后,系统再次核对 o_k 中的实际观测,将视觉证据优先于冲突预测,从而确定下一决策的真实状态 s
    (k+1) 。 ### 4. 动作条件化与块级指导的音视频生成 生成器 G_θ 基于 LTX-2.3 构建,采用耦合的视频-音频 Transformer 流: - **动作条件**:连续移动/视角控制通过 FiLM 层调制视频特征;离散按钮(射击、跳跃等)通过因果交叉注意力注入; - **块级提示路由(Chunk-wise Prompt Routing)**:全局提示 g 描述场景与角色身份,局部指导 p_k 描述当前块的行为与可观测后果,二者组成文本条件 c_k = (g, p_k) 。每个生成块仅关注自身对应的局部指导,历史块保留其原始条件; - **自回归生成**:给定视听上下文 H_k (首帧锚点、角色参考、近期生成块)与玩家动作 u_k ,生成下一块:

ok sim Gθ(H_k, u_k, c_k)
上下文采用滑动窗口维护,保证持续生成长序列时历史有界。 ### 5. 面向交互的数据集与训练策略 为支撑上述能力,论文构建了 **Xonotic Interaction Dataset**,包含三个互补子集: - **Basic Dynamics**:玩家控制与局部动态; - **Interaction**:受控的交火、命中/未命中/击杀结果与角色行为标注; - **Arena**:持续的一对一战斗,支持长期交互。 训练分为阶段化适配: - **双向预训练**:先学习跟随玩家控制与块级语义指导; - **自回归适配**:在滑动窗口上进行因果下一块预测训练; - **蒸馏加速**:采用 Causal-rCM 两阶段蒸馏(一致性蒸馏 + 分布匹配),实现少步数生成; - **后训练**:在 Arena 数据上进一步微调,得到可交互原型 **RIGen-Arena**; - **视频细化**:训练域 LoRA 恢复高频动态细节,提升视觉保真度。 ### 6. 闭环的双向交互效果 综合以上设计,RIGen 实现了以下双向因果链: - **玩家 to 世界**:玩家射击消耗弹药,命中依据规则扣减敌人生命,致死则改变敌人状态并终止其行动能力; - **世界 to 玩家**:敌人依据自身血量与策略模式(如低血量撤退)自主规划行为;敌人攻击命中依据规则扣减玩家生命,玩家生命归零则进入战败状态,丧失行动能力并触发相应的视听与游戏逻辑反馈。 由此,交互后果与角色行为共同决定生成内容,且游戏状态与视觉结果保持显式一致。 Q4: 论文做了哪些实验? 论文在第6节及附录中开展了**定量对比**、**消融分析**、**定性评估**与**系统原型验证**四类实验,全面检验交互结果可靠性、状态依赖行为与视听生成质量。 —- ### 1. 定量评估(Quantitative Evaluation) #### 实验设置 - **测试数据**:100个 Xonotic 战斗场景(50个要求敌人最终**存活**,50个要求最终**死亡**)。测试地图与敌人外观均排除在所有训练子集之外。 - **受控条件**:敌方运动由源标注预先指定,而非在线行为选择,从而将对比焦点置于“战斗结果实现”与“运动遵循度”上。 - **交互代理**:使用 InternVL3-8B 进行事件推理与视觉状态校正。 #### 对比基线与消融变体 | 类型 | 方法/变体 | 说明 | |:—-|:—-|:—-| | 外部基线 | SCOPE (Tong et al., 2026) | 可玩 FPS 世界模型 | | 外部基线 | LingBot-World 2.0 (Gao et al., 2026) | 支持语义交互的开放世界模型 | | 消融 | Motion guidance only | 仅保留玩家输入与源运动,移除交互指导与在线代理 | | 消融 | Semantic prompting | 代理以文本形式维护状态与规则,无可执行规则引擎,无独立验证校正 | | 消融 | Post-hoc event updates | 交互推理与状态更新移至生成**之后**,仅指导下一块 | | 消融 | Without reconciliation | 保留生成前预测与可执行状态更新,但禁用生成后的视觉校正 | | 参考对照 | RIGen (w/ Reference guidance) | 离线冻结基于源标注的参考事件指导,移除在线代理调用 | #### 评估指标 - **Enemy Behavior (EB)**:敌人在指定时间窗口内是否遵循预定方向运动(由 Qwen3.8-27B 离线评判)。 - **Final-state Agreement (FSA)**:最终可见敌人状态与预设目标(存活/死亡)的一致性。细分为: - FSA-S:50个生存目标中敌人最终可见存活的比率; - FSA-D:50个死亡目标中敌人最终可见死亡的比率。 - **视觉与动态指标**(VBench 维度):Motion Smoothness (MS)、Aesthetic Quality (AQ)、Subject Consistency (SC)、Background Consistency (BC)、Dynamic Degree (DD)。 #### 主要结果 - **交互结果可靠性**:RIGen 的 FSA 达到 **81%**,显著高于 LingBot-World 2.0 的 29% 与 SCOPE 的 4%。 - **消融分析支撑设计选择**: - Motion guidance only:FSA 降至 39%,说明仅有运动指导无法保证战斗结果; - Semantic prompting:FSA 降至 35%,表明文本状态管理不如可执行规则引擎可靠; - Post-hoc event updates:FSA 降至 64%,验证了“当前块生成前即预测后果”的必要性; - Without reconciliation:FSA 降至 76%,说明视觉反馈校正状态对长期一致性仍有贡献; - RIGen (w/ Reference guidance):作为特权离线控制,FSA 达到 89%,接近上限。 - **视觉质量**:各变体的 MS、SC、BC 均保持高位且差异很小,表明交互机制改进不损害视听保真度;AQ 略低于 LingBot-World 2.0,推测与训练数据规模较小有关。 —- ### 2. 定性评估(Qualitative Evaluation) #### (a) 战斗响应对比(图 6) 在相同初始图像与射击请求下,对比 RIGen、SCOPE 与 LingBot-World 2.0 的可见战斗反馈: - **SCOPE 与 LingBot-World 2.0**:尽管接收了预设结果作为文本指导,敌人在连续射击后仍保持直立或活跃状态,未正确呈现受击、倒地与死亡。 - **RIGen**:依次生成受击反应(hit reaction)、倒地(falling)与持续倒地状态(sustained downed state),实现了与预设目标一致的战斗后果。 #### (b) 状态依赖交互分支(图 7) 从同一起始场景出发,在相同玩家输入请求下,改变初始游戏状态,验证世界状态的因果影响: - **Reference**:敌人血量 200,弹药充足,策略为“固守”; - **Low Health**:敌人血量降至 50,射击后敌人直接倒地死亡; - **Low Ammo**:玩家弹药仅余 2 发,第三发射击被抑制(无射击动作); - **Injured / Strategy Switch**:敌人受伤后策略由“固守”切换为“撤退”,表现为向左移动并在掩体后蹲下。 这些分支直观展示了显式状态与策略规则如何塑造可见的游戏行为与生存结果。 —- ### 3. 交互原型验证:RIGen-Arena(图 5) 论文构建了可交互的一对一 FPS 原型 **RIGen-Arena**,验证持续闭环游玩中的双向交互: - **对手被击败**:敌人生命值归零后进入死亡状态; - **重生与搜寻**:按规则重生后,敌人自主搜寻玩家; - **玩家被击败**:敌人攻击命中导致玩家生命归零,进入战败状态,HUD 同步更新。 该原型将少步数音视频生成、状态与规则引擎、交互代理整合为完整系统,支持在持久游戏规则下的连续游玩。 —- ### 4. 附录中的补充实验细节 附录进一步提供了支撑主实验的实现与评估细节: - **训练与蒸馏协议**:包括双向适配、自回归滑动窗口训练、Causal-rCM 两阶段蒸馏(一致性蒸馏 + 分布匹配)及视频细化器的训练目标; - **数据集构建**:Basic Dynamics、Interaction(短/长序列)、Arena 三个子集的采集路线、对齐方式、标注流程与质控标准; - **运行时性能**:在单张 NVIDIA GB300 上,RIGen(含细化)在线交互模式约为 3.93 FPS,离线重放模式约为 4.41 FPS(均低于实时播放目标 20 FPS,说明实时性仍是未来改进方向)。 Q5: 有什么可以进一步探索的点? 基于论文第7节(Conclusion)中”Limitations and future work”的明确阐述,以及全文技术框架所隐含的扩展空间,可从以下几个方向进一步探索: ### 1. 实时推理与生成加速 当前 RIGen-Arena 的在线交互模式在单张 NVIDIA GB300 上的生成吞吐量为 **3.93 FPS**(附录表 5),尚低于实时播放所需的 20 FPS。未来的关键工作包括: - 压缩交互代理(VLM)的推理延迟; - 进一步减少生成器去噪步数或采用更激进的蒸馏策略; - 优化滑动窗口历史缓存与键值复用机制,降低自回归生成的计算开销。 ### 2. 从单对手 FPS 扩展到多实体与复杂游戏类型 现有原型仅支持**一对一**第一人称射击(FPS)场景。拓展空间包括: - **多对手协同与对抗**:引入多智能体交互,处理群体战术、仇恨值分配与团队行为; - **游戏类型泛化**:将状态-规则-推理三层架构迁移至动作角色扮演(ARPG)、生存建造、潜行或叙事驱动类游戏,支持更复杂的机制(如资源合成、任务状态机、对话分支)。 ### 3. 视觉基础能力与推理可靠性的提升 当前系统依赖**VLM(InternVL3-8B)**进行事件推理、行为规划与视觉验证,存在感知错误与推理不稳定的风险。可探索: - 针对游戏交互的专用视觉基础模型,提升对血量变化、受击动画、遮挡与距离判定的细粒度识别; - 引入不确定性量化或置信度阈值,在视觉证据不足时回退到规则引擎的保守估计; - 结合显式追踪(如持久化角色重识别与空间定位)辅助 VLM,降低纯视觉推理的负担。 ### 4. 规则与状态表示的自动学习 当前状态机与转移规则由人工设计与配置。长期可探索: - **从游戏日志与视频中自动抽取状态转移规则**,减少手工工程; - **可微分或神经符号状态机**:将规则引擎与生成器联合优化,使状态转移逻辑能够适应新游戏领域; - 学习隐式状态表示,作为显式数值状态的补充,以处理难以规则化的微妙交互(如士气、警觉度)。 ### 5. 更长程因果链与跨会话记忆 论文指出视频模型的有限历史上下文会导致早期事件后果难以持续(第1节)。尽管已有滑动窗口,但以下问题仍需深化: - **跨小时级甚至跨会话的持久世界状态**:如角色关系演进、世界地形破坏、剧情标志位的长期保留; - 分层记忆架构:将近期感知细节与长期抽象知识分离存储,避免上下文膨胀。 ### 6. 动作空间与输入模态的扩展 当前动作接口仅包含移动/视角控制与射击/跳跃/换弹。未来可支持: - 更丰富的离散与连续动作组合(如蹲伏、瞄准、投掷、载具驾驶); - 自然语言指令作为玩家输入模态,与底层物理控制融合,实现高层策略与低层执行的统一生成。 ### 7. 多玩家与社交交互 现有数据集与原型仅涉及单玩家与 bot 对手。引入真实多玩家或社交 NPC 将带来新的挑战: - 多视角一致性(multiplayer observations of shared state); - 实时同步下的生成一致性,避免不同客户端间状态漂移; - 社交行为建模(如合作、交易、欺骗)与对应的多模态反馈(语音、文字、表情)。 ### 8. 自动化评估与基准建设 当前离线评估依赖 Qwen3.8-27B 对生成视频进行事后判断,主观性较强。未来需要: - 针对生成式游戏交互的**细粒度、可复现的自动化基准**,涵盖规则遵循率、状态一致性、行为合理性等多维指标; - 建立长期可玩性(playability over extended sessions)的量化标准,超越单片段的 FSA 与 EB 指标。 Q6: 总结一下论文的主要内容 这篇论文提出了 **RIGen(Reciprocally Interactive Generation)**,一个用于生成式游戏玩法中**双向交互(reciprocal interaction)**的框架,解决了现有视频生成模型仅被动响应玩家动作、难以维持持久游戏状态与自主世界行为的局限。 ### 核心问题 生成式游戏不仅需要视听反馈跟随玩家控制,更要求: - 玩家动作对世界产生**持久后果**(如消耗弹药、扣减生命值); - 世界依据自身状态与规则**反作用于玩家**(如敌人低血量时撤退、击败玩家后终止其行动)。 现有基于纯视频预测的方法面临三大挑战:协调双向因果交互、历史上下文有限、关键游戏状态仅部分可观测。 ### 方法框架 RIGen 通过三个耦合层次形成闭环: 1. **状态与规则引擎(State and Rule Engine)** 以事件驱动状态机维护玩家/敌人生命值、弹药、策略模式等持久变量,并基于可配置规则执行确定性的状态转移。 2. **交互推理层(Interaction Agent)** 在生成每个音视频块**之前**,基于历史视频、玩家动作与当前状态,先验预测交互事件与角色行为;生成后,通过视觉验证校正状态,使后续决策基于真实观测而非仅依赖预测。 3. **音视频生成器(Audiovisual Generator)** 基于 LTX-2.3 构建,支持动作条件化(FiLM + 交叉注意力)与块级语义指导路由,将行为规划与战斗后果转化为连贯的视听输出。 ### 数据集与原型 - 构建了 **Xonotic Interaction Dataset**(含基础动态、受控交互与持续战斗子集),提供同步的音视频、玩家控制、姿态与交互标注。 - 经后训练得到交互原型 **RIGen-Arena**,支持在持久规则下的一对一 FPS 持续游玩(含死亡、重生、行为策略切换)。 ### 实验结果 - **定量对比**:在 100 个战斗场景测试中,RIGen 的最终状态一致率(FSA)达 **81%**,显著高于 SCOPE(4%)与 LingBot-World 2.0(29%)。 - **消融分析**:验证了“生成前预测后果”、“可执行规则引擎”与“生成后视觉校正”各组件对可靠实现战斗结果的必要性。 - **定性结果**:展示了状态依赖的分支行为(低血量死亡、弹药耗尽抑制射击、受伤后撤退)与正确的受击-倒地反馈。 ### 局限与未来方向 当前系统尚未达到实时(~4 FPS,目标 20 FPS),仅支持单对手 FPS,且依赖 VLM 进行视觉推理。未来工作将聚焦生成与推理加速、多对手/多类型游戏泛化、视觉基础能力增强与长程跨会话记忆。 Q7: 想要进一步了解论文 以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文,请点击 这里 为你跳转Kimi AI网页版,并启动一个与该论文相关的新会话。

Authors: Mengfei Li, Yi Zhang, Xingtong Ge, Lunjie Zhu, Fangyu Lin, Peng Li, Guanglu Song, Yu Liu, Wei Xue, Wenhan Luo, Yike Guo

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20811.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20811

Published: 2026-10-19T01:40:24.655Z


9. Intern Lumina U2: A Multi-Codebook Diffusion Large Language Model for Omni-Visual Understanding and Image Generation

Abstract:Unified multimodal generation covers discrete, continuous, and hybrid paradigms. Discrete approaches with single-codebook visual tokenizers align naturally with text modeling but have limited representational capacity. Continuous and hybrid approaches retain richer visual representations, but they model text and images in heterogeneous spaces with different training objectives, potentially introducing optimization conflicts between modalities with misaligned representation spaces. These choices expose a central question: how to efficiently model high-capacity visual representations alongside textual content in a shared optimization space within a single backbone. We present Intern Lumina U2, to our knowledge the first diffusion language model to represent multi-codebook visual tokens in a fully discrete space. Each visual position is represented by a group of codes from complementary codebooks, increasing its representational capacity while preserving compatibility with discrete text modeling. At the input, the embeddings are concatenated at each position; at the output, a lightweight multi-codebook autoregressive head predicts the codes within each position sequentially, while the diffusion backbone processes all positions in parallel. Built on the unified 4D representation of AToken, Intern Lumina U2 supports image, video, and 3D inputs within a single backbone and achieves leading performance among unified models in visual understanding, image generation, and image editing. We further build a full-stack training, inference, and evaluation system on both Ascend and NVIDIA platforms to verify the cross-platform effectiveness of our model. We also improve the inference efficiency through a stability-aware sampling strategy, which combines confidence-plateau-guided unmasking approach with drift-triggered cache refreshing mechanisms, accelerating image generation by 4.45x.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

Q: 这篇论文试图解决什么问题? 这篇论文旨在解决\*\*统一多模态理解与生成框架中视觉表征容量与优化空间对齐之间的根本矛盾\*\*,具体可概括为以下三个层面: ### 1. 核心问题:共享优化空间中的高容量视觉表征建模 现有统一多模态模型在融合视觉理解与生成时,面临一个关键抉择: - \*\*离散范式\*\*(如 Chameleon、Emu3):采用单码本视觉 tokenizer,将视觉信息压缩为单个离散代码,天然兼容文本的离散建模方式,但每个空间位置的表征容量严重不足,难以忠实重建图像或覆盖复杂视觉任务; - \*\*连续/混合范式\*\*(如 Transfusion、BAGEL、InternVL-U):在主干网络或独立解码器中保留连续视觉状态,虽能获得更高的视觉保真度,但文本与图像在异构空间中被建模,且往往采用不同的训练目标(如文本的 next-token loss 与图像的 diffusion/flow matching loss),导致模态间的优化冲突和表征空间不对齐。 由此,论文提出一个核心研究问题:\*\*如何在单一主干网络的共享优化空间内,以单一训练目标高效地建模高容量视觉表征与文本内容?\*\* ### 2. 现有离散统一模型的表征瓶颈 现有基于离散掩码扩散的统一模型(如 MMaDA、Lumina-DiMOO、LLaDA2.0-Uni)受限于单码本 tokenization,将每个视觉位置的信息压缩为单个离散代码 c ∈ 1, dots, V 。这种紧凑表征对于以下两方面均显不足: - \*\*忠实重建\*\*:单一代码难以编码精细的低层视觉细节; - \*\*语义理解\*\*:高层语义信息与视觉生成所需的细节之间存在张力。 论文指出,这种表征容量的局限是离散统一模型在视觉任务上性能受限的主要原因。 ### 3. 论文提出的解决方案 为突破上述瓶颈,论文提出 \*\*Intern Lumina U2\*\*,其设计目标是在\*\*完全离散空间\*\*中实现高容量视觉表征与文本的统一建模: - \*\*多码本离散视觉表征\*\*:采用 AToken 将图像、视频、3D 资产编码到统一的 4D 潜在空间,每个视觉位置由 K=8 个互补码本的代码组 c = (c_1, dots, c_K) 表示,每个码本大小为 4096。这种表示在不增加空间网格长度的前提下,将视觉信息容量扩展了 K 倍。 - \*\*输入接口的兼容性保持\*\*:在输入端,每个码本拥有独立的嵌入表 E_k ,通过拼接与线性投影将 K 个代码嵌入合并为单一主干 token:

x = W(∈) · [E_1(c_1); E_2(c_2); dots; E_K(c_K)]
这使得序列长度与码本数量 K 解耦,同时避免了将视觉代码混入文本词表。 - **输出接口的依赖建模**:在输出端,轻量级的多码本自回归头(multi-codebook AR head)沿码本维度自回归地建模代码间的依赖关系:
p(c mid h) = prod
(k=1)^(K) p(ck mid h, c(<k))
而扩散主干则并行处理所有空间位置,兼顾了视觉生成对双向上下文的需求与代码间的条件依赖。 - **统一训练目标**:所有模态(文本、图像、视频、3D)在统一的 packed-sequence 格式下,以单一的掩码扩散交叉熵目标进行联合训练,避免了异构优化目标带来的冲突。 ### 4. 延伸的系统与效率问题 除表征与优化空间的统一外,论文还针对实际部署中的两个关键问题给出了解决方案: - **跨平台有效性**:在华为 Ascend 与 NVIDIA 双平台上完成了全栈训练、推理与评估系统的适配,验证了架构与训练 recipe 的跨平台一致性; - **推理加速**:针对视觉 token 置信度低、演化缓慢的特性,提出 **Plateau-Halton sampling with V-verify (PHV)** 策略,通过置信度 plateau 指导的解掩与隐藏状态漂移触发的缓存刷新,在保持生成质量的前提下将图像生成端到端延迟降低 **4.45×**。 综上,Intern Lumina U2 试图在**不引入连续生成目标或独立视觉解码器**的前提下,首次实现基于多码本离散状态的全模态(图像/视频/3D)理解与图像生成、编辑的统一,推动离散扩散语言模型(diffusion LLM)向高容量视觉表征与 omni-visual 范式发展。 Q2: 有哪些相关研究? 与 Intern Lumina U2 相关的研究可归纳为以下四条主线,涵盖从统一架构、扩散语言模型到多码本表征与推理加速的多个维度: —- ### 1. 统一多模态理解与生成模型(Unified Multimodal Understanding and Generation) 这类研究致力于在单一模型内同时实现视觉感知与视觉合成,依据视觉表征的离散/连续属性可进一步分为三类: **(1)纯离散范式:视觉 token 与文本共享离散空间** - **自回归生成**:Chameleon 与 Emu3 采用共享词表,将图像与文本均视为 next-token 预测任务;Janus-Pro 则解耦理解编码器与生成 tokenizer,分别处理两类需求;VAR 与 OneCAT 提出由粗到细的 next-scale 预测,逐步提升分辨率。 - **掩码扩散生成**:MMaDA、Lumina-DiMOO 与 Show-o 将掩码扩散目标同时应用于文本与离散视觉 token,允许视觉内容在双向上下文下并行解码;Emu3.5 进一步将 next-token 图像解码转化为双向并行预测以加速推理。 **(2)连续/混合范式:保留连续视觉状态** - **主干内连续目标**:Transfusion 在统一 Transformer 内对文本使用 next-token loss、对图像使用扩散 loss;UniFluid 为每个空间位置配备独立的扩散头。 - **分离式连续解码器**:BAGEL、InternVL-U 采用 ViT 特征进行理解,再通过 VAE 空间的 flow matching 生成图像;SenseNova-U1 将 flow matching 直接置于像素空间。 - **语义-连续混合**:LLaDA2.0-Uni 在主干中对文本与视觉 token 使用共享掩码预测目标,但图像合成仍依赖一个额外的连续扩散解码器,以预测的离散视觉 token 为条件。 **与 Intern Lumina U2 的区别**:上述离散方法受限于单码本表征容量;连续/混合方法则引入与文本分类目标异构的连续生成目标,甚至需要独立的视觉解码器,导致优化空间不一致。Intern Lumina U2 首次在**完全离散空间**内对**多码本视觉状态**进行建模,无需连续生成器即可直接通过主干 token 解码图像。 —- ### 2. 扩散语言模型(Diffusion Language Models, dLLMs) 该方向关注如何用扩散/掩码过程替代自回归进行语言与多模态建模: - **基础语言模型**:LLaDA 首次证明从零训练的掩码扩散语言模型可媲美强自回归基线;LLaDA2.0 将预训练的 AR MoE 模型转换为 blockwise 扩散 LM,在保留知识的同时实现并行解码。 - **向视觉的扩展**: - **理解**:LLaDA-V 将视觉编码器特征馈入扩散 LM 完成视觉问答。 - **生成**:LLaDA-o 对文本采用离散掩码扩散、对图像采用连续扩散,形成非统一的生成机制。 - **统一**:MMaDA 与 Lumina-DiMOO 将掩码预测扩展到文本与图像 token;LLaDA2.0-Uni 基于 LLaDA2.0 主干采用 block-wise 掩码预测。 **与 Intern Lumina U2 的区别**:现有扩散语言模型(包括统一版本)均采用单码本视觉表征或连续视觉分支。Intern Lumina U2 以 LLaDA2.0-mini 为起点,沿两个轴扩展:视觉状态升级为**多码本离散 token**,且同一模型覆盖图像、视频与 3D 三种视觉模态的理解。 —- ### 3. 多码本视觉表征与深度解码(Multi-Codebook Representations) 为在不增加空间网格长度的前提下提升离散视觉表征容量,研究者探索了残差与因子化量化: - **RQ-Transformer**:使用“空间 + 深度”双层 Transformer 自回归地解码残差代码。 - **VILA-U**:将多码本深度解码引入自回归统一模型。 - **UniTok**:将 H × W × K 的视觉代码折叠为 H × W 个主干位置,但未显式建模同位置内多码本间的依赖。 - **TokenBridge**:记录了当单个位置的各码本代码被独立预测时所丢失的依赖关系。 - **AToken**:提供跨图像、视频、3D 的共享多码本表征,是 Intern Lumina U2 的视觉 tokenizer 基础。 **与 Intern Lumina U2 的区别**:上述工作均在**自回归系统**中建立多码本建模。Intern Lumina U2 首次将多码本视觉 token 作为**掩码扩散语言模型**的视觉状态:扩散主干并行细化所有空间位置,而多码本 AR head 仅在每个位置内部沿码本维度自回归建模依赖,从而将空间并行性与码本序列性解耦。 —- ### 4. 掩码扩散解码加速(Accelerating Masked-Diffusion Decoding) 针对掩码生成模型迭代推理成本高的问题,现有加速方法分为两路: **(1)基于学习的方法** - 训练辅助模型或学习潜在受控动态(latent controlled dynamics),以少数大步长跳跃替代多步细化。 **(2)免训练的方法** - **固定/阈值调度**:MaskGIT 预设每轮解掩位置数量;Fast-dLLM 对概率超过绝对阈值的 token 立即提交。 - **揭示顺序优化**:低差异序列(如 Halton 序列)用于均匀分散剩余不确定性,改善空间覆盖。 - **缓存策略**:固定间隔刷新缓存以减少冗余主干计算;dLLM-Cache 提出基于 V-verify 分数的漂移检测,选择性刷新缓存位置。 **与 Intern Lumina U2 的关系**:该模型提出 **Plateau-Halton sampling with V-verify (PHV)**,属于免训练加速路线。由于视觉 token 置信度绝对值较低且演化缓慢,PHV 放弃固定阈值,转而利用置信度**时间 plateau** 作为稳定性信号进行早提交,并结合隐藏状态漂移触发缓存刷新,在 1024×1024 生成中实现 **4.45×** 端到端加速。 —- ### 总结 如论文 Table 1 所示,在现有的代表性统一多模态模型中,Intern Lumina U2 是唯一同时满足以下五项能力的系统:图像理解、图像生成、视频理解、3D 理解、基于多码本离散状态的掩码扩散解码,且无需独立的连续生成器。这一定位使其与上述四条研究脉络中的既有工作形成明确区分。 Q3: 论文如何解决这个问题? 论文通过 **Intern Lumina U2** 系统性地解决了统一多模态框架中视觉表征容量不足与优化空间异构的矛盾。其核心思路是:在完全离散空间内,以多码本视觉 token 扩展每个空间位置的表征容量,并通过掩码扩散语言模型(dLLM)以单一目标统一处理文本与视觉。具体解决方案涵盖架构设计、训练范式、数据工程与推理系统四个层面。 —- ### 1. 多码本离散视觉表征(Multi-Codebook Discrete Visual Tokens) 为突破单码本离散表征的信息瓶颈,论文采用 **AToken** 作为冻结的视觉 tokenizer,将图像、视频与 3D 资产统一编码至共享的 4D 潜在空间。每个视觉位置由 K=8 个互补码本的代码共同表示:

c = (c_1, c_2, dots, c_K), quad c_k ∈ 1, dots, 4096
相比单码本,这种表示在不扩展空间网格长度的前提下,将视觉信息的编码容量提升了多个数量级,同时保持离散性以兼容文本的分类式训练目标。 —- ### 2. 统一的模型架构 #### 2.1 输入接口:码本独立嵌入与投影 为避免视觉代码混入文本词表、同时保证主干序列长度与码本数量 K 解耦,论文为每个码本 k 配备独立的嵌入表 E_k 。对于视觉位置 c = (c_1, dots, c_K) ,输入向量通过拼接与线性投影得到:

x = W(∈) · [ E_1(c_1); E_2(c_2); dots; E_K(c_K) ], quad W(∈) ∈ R^(d × Kd)
其中 d=2048 为主干隐藏维度。该设计既保留了码本身份(codebook identity),又将多码本信息压缩为单个主干 token。 #### 2.2 输出接口:多码本自回归头(Multi-Codebook AR Head) 在输出端,论文不假设各码本条件独立,而是通过轻量级的 6 层因果 Transformer 沿码本维度自回归建模依赖关系:

p(c mid h) = prod(k=1)^(K) p(c_k mid h, c(<k))
其中 h ∈ R^d 为扩散主干输出的位置隐藏状态。该头部在训练时采用 **teacher forcing**,一次性并行获得所有 K 个码本的 logits;在推理时,每个视觉位置内部的 K 个代码按序采样,不同空间位置则始终保持并行处理。这种“空间并行、码本串行”的因子化策略,既保留了视觉生成所需的双向上下文(通过扩散主干),又显式捕获了同位置多码本间的联合分布。 #### 2.3 任务感知注意力掩码(Task-Aware Attention Mask) 由于模型需同时处理理解、生成、编辑等多类任务,论文设计了基于元数据的统一序列格式与动态注意力掩码: - **条件(conditions)** 对同轮及后续轮次可见; - **文本答案** 采用块级扩散(block diffusion):块内双向、块间因果,并引入 clean copy 机制支持多轮上下文; - **视觉答案** 采用空间掩码扩散:在自身答案内全双向,可读取条件及前序文本答案的 clean copy,但无 clean copy 自身。 通过将图像、视频、3D 与文本统一打包为 1D token 序列,所有任务共享同一套位置编码与掩码规则。 —- ### 3. 统一的训练目标 所有模态在共享的离散空间中优化**同一个掩码扩散交叉熵目标**。 - **文本掩码调度**:按块独立采样掩码比例 rho = cos((π) / (2) u) ,其中 u sim U(0,1) ,以匹配推理时的块扩散行为。 - **视觉掩码调度**:所有空间位置共享同一掩码比例,同位置的 K 个代码全部被监督。位置 j 的损失为:

ell(vis)(j) = (1) / (K) ∑(k=1)^(K) -log p(ck^(gt) ,|, h, c( fp ,则该位置被视为已进入稳定 plateau,可提前提交。 - **Halton 揭示顺序**:未达 plateau 的位置按低差异 Halton 序列排序提交,使剩余不确定性均匀分散在画布上。 - **V-verify 缓存刷新**:通过监测隐藏状态归一化漂移 D_t 决定是否刷新主干缓存,避免过时表征的累积。 配合 CFG 双分支批处理、AR Head KV 缓存与隐藏状态缓存复用,PHV 将端到端图像生成延迟从 159.9 秒降至 35.9 秒(**4.45× 加速**),同时提升生成质量。 —- ### 6. 跨平台全栈适配与验证 论文在 **1,024 张华为 Ascend 910B NPU** 上完成主模型训练,并在 **NVIDIA GPU** 上以相同架构与训练 recipe 进行小规模交叉验证。通过重写融合 MoE 内核(基于 `npu_grouped_matmul`)、适配注意力掩码语义与内存管理策略,实现了从训练到推理的全栈国产化部署,证明了该方法在不同硬件平台上的有效性。 —- ### 总结 Intern Lumina U2 的解决方案可概括为:**以多码本离散 token 扩展视觉表征容量,以拼接-投影输入接口保持主干兼容性,以多码本 AR Head 捕获码本间依赖,以统一掩码扩散目标对齐文本与视觉的优化空间,以渐进课程训练稳定扩展至多模态,以稳定性感知采样突破推理效率瓶颈**。这一系列设计首次在完全离散空间内实现了图像、视频、3D 理解与图像生成、编辑的统一建模。 Q4: 论文做了哪些实验? 论文在 **Intern Lumina U2** 上开展了覆盖 omni-visual 理解、图像生成与编辑、科学图表生成、推理加速及架构消融的系统性实验。所有实验均对比了 Ascend 平台训练的主模型与 NVIDIA 平台训练的参考模型,并在统一协议下与专用及统一基线进行比较。 —- ### 1. Omni-Visual 理解实验 #### 1.1 图像理解与推理 在 10 个图像理解基准上评估,涵盖文档图表、感知幻觉与知识数学推理三类任务: - **文档与图表**:ChartQA、CharXiv(DQ) - **感知与幻觉**:HallusionBench、MME-C、RealWorldQA - **知识与数学推理**:MMMU、MMMU-Pro、MathVista、MathVision、DynaMath 与 Qwen2.5-VL(7B,理解专用)、LLaDA-V、BAGEL、InternVL-U、Lumina-DiMOO、LLaDA-o 及同骨干的 LLaDA2.0-Uni 对比。Ascend 模型在 6/10 项上取得最优,在 CharXiv、HallusionBench、MathVision、DynaMath 上领先明显;与同骨干 LLaDA2.0-Uni 相比,8/10 项提升,验证了多码本表征对细粒度视觉细节的增益。 #### 1.2 视频与 3D 理解 - **视频基准**:Video-MME(无字幕)、MLVU(m-avg)、MVBench - **3D 基准**:3D MM-Vet(Rec / Know / Gen / Spat / Emb / Total) 对比 Video-LLaVA-7B、VideoLLaMA2-7B、LLaVA-Mini、LLaDA-V 等视频/3D 专用或理解模型。Ascend 模型在 Video-MME、MLVU、MVBench 三项视频基准上均取得最高得分;在 3D MM-Vet 上超过 ShapeLLM 与 PointLLM 等 3D 专用模型。 —- ### 2. 图像生成与编辑实验 #### 2.1 文本到图像生成 在多个生成基准上评估整体与维度级性能: - **整体对齐**:GenEval、DPG-Bench、TIIF-Bench(长短提示分别评测)、WISE - **维度细分**:UniGenBench(Sty / Wor / Att / Act / Rel / Log / Gra / Com / Lay / Txt 十维度) - **综合能力**:OneIG-Bench(Alignment / Text / Reasoning / Style / Diversity) 对比生成专用模型 FLUX.1、Qwen-Image,以及统一模型 Janus-Pro、Show-o2、BAGEL、OmniGen2、InternVL-U、MMaDA、Lumina-DiMOO、LLaDA-o、LLaDA2.0-Uni。Intern Lumina U2 在统一模型中取得 DPG-Bench、UniGenBench、TIIF-Bench 最优,GenEval 与 LLaDA2.0-Uni 持平;UniGenBench 中逻辑推理(Log)与语法(Gra)优势显著。 #### 2.2 图像编辑 评测指令驱动的图像编辑: - **ImgEdit**:总体得分(1–5 分制) - **GEdit-Bench**:英文与中文提示的总体 GO 得分(0–10 分制) 对比编辑专用 Step1X-Edit、Qwen-Image-Edit,以及统一模型 BAGEL、OmniGen2、InternVL-U、Lumina-DiMOO、LLaDA2.0-Uni。Ascend 模型在 ImgEdit 与 GEdit 英文基准上接近最优统一模型,优于 BAGEL 与 OmniGen2。 —- ### 3. 科学图表生成实验 针对物理正确性,在两个专用基准上评估: - **GenExam Physics**:严格(strict)与宽松(relaxed)的 MLLM-judge 分数 - **VeriphyT2IBench**:局部属性库(Local)与全局整图库(Global)的问题级准确率,并进一步按电磁学(E&M)、力学(Mechanics)、光学(Optics)、量子力学、热力学、声学六领域及易/中/难三难度级细分 对比生成专用模型(Seedream 4.0、FLUX.1、Gemini 2.5 Flash Image、GPT-Image-1、Qwen-Image 等)、统一模型(BAGEL、Lumina-DiMOO、Show-o2、SenseNova-U1 等)及物理专用变体(Princigram-DiMOO、Princigram-BAGEL)。Intern Lumina U2 在一般用途统一模型中 GenExam 宽松得分最高(44.8),VeriphyT2IBench 局部最优(76.58)、全局第二(81.22);在难度最高的 E&M 提示上达 81.30。 —- ### 4. 推理加速与采样策略实验 #### 4.1 系统级加速消融(Table 4) 在固定 1024 × 1024 文本到图像协议下,逐层叠加优化并测量端到端延迟与 GenEval: - **基础路径**:96 步,159.9 s - **+ CFG 双分支批处理**:117.2 s(1.36×) - **+ AR Head KV 缓存**:84.2 s(1.90×) - **+ 隐藏状态缓存**:64.1 s(2.49×,命中率 65%) - **+ Halton 揭示顺序(HVV)**:66.5 s(2.40×) 进一步引入自适应步预算: - **HVV 固定 60 步**:42.6 s(3.75×) - **Plateau + V-verify**:34.8 s(4.59×) - **PHV(Plateau-Halton + V-verify)**:35.9 s(4.45×),GenEval 从 86.6 提升至 89.3 #### 4.2 采样参数网格搜索(Figure 11) 对 PHV 的两个核心超参——plateau 变化阈值 δ_p 与置信度下限 f_p ——进行 6 × 6 网格消融(553 提示/格)。结果表明二者形成串行门控: δ_p 决定早期提交资格, f_p 限制实际提交数量。选定的质量优先配置 δ_p = 0.18, f_p = 0.005 位于速度-质量前沿,以 35.9 s/张达到 GenEval 89.3。 —- ### 5. 架构消融实验 #### 5.1 视觉输入接口消融(Table 15, Figure 10) 比较三种输入设计(决定嵌入表组织与分类器宽度): - **D1**:单一共享大词表(文本+视觉代码共用) - **D2**:单一视觉嵌入表,通过索引偏移区分码本 - **D3(本文)**:每码本独立嵌入表,拼接后投影 在约 500 万监督代码的 held-out 集上,D3 的 next-code top-1 准确率显著高于 D1/D2(差距超过 10 个标准误),且随码本深度增加优势扩大;图像损失收敛曲线亦优于 D2。D3 的峰值内存与 D2 相当(18.62 GB vs 18.68 GB),远低于 D1(43.79 GB)。 #### 5.2 视觉输出头消融(Table 16) 在 512 × 512 分辨率、相近 token 预算下对比三种输出头: - **Independent parallel**:假设码本条件独立,并行预测 - **Q-Former**:单轮学习查询预测所有代码 - **Multi-codebook AR(本文)**:沿码本维度自回归建模 p(c mid h) = prod(k=1)^(K) p(ck mid h, c(<k)) 多码本 AR Head 在 GenEval(82.78)、DPG-Bench(83.27)、TIIF(73.2)、UniGenBench(62.1)四项上全面最优。Parallel Head 在 DPG 上接近,但在 GenEval 落后 14 分,尤其在计数(counting)、位置(position)与属性绑定(attribute binding)上差距显著,验证了显式建模码本间依赖的必要性。Q-Former 未优于 Parallel Head,说明增益来源于依赖建模而非更复杂的条件化。 —- ### 6. 跨平台一致性验证 论文在 Ascend 910B(1,024 NPUs,完整预算) Q5: 有什么可以进一步探索的点? 基于论文结论与整体架构设计,以下方向值得进一步探索: —- ### 1. 向视频与 3D 生成的模态扩展 当前 Intern Lumina U2 已统一了图像、视频、3D 的**理解**,但生成与编辑能力仍局限于图像。由于 AToken 已将三种模态编码至共享 4D 潜在空间,且视觉接口对模态差异透明,将多码本离散扩散框架扩展至**视频生成**(帧间时序一致性建模)与**3D 资产生成**(稀疏体素或点云的空间结构化生成)是自然的下一步。核心挑战在于设计适配时序与三维结构的任务感知掩码机制,以及处理视频/3D 场景下更长的视觉序列。 —- ### 2. 更紧密交织的多模态生成(Interleaved Multimodal Generation) 现有工作主要处理“文本→图像”或“图像→文本”的单向转换,未来可探索**交错式多模态生成**,即模型在单一对话中交替输出文本推理与视觉内容(如生成带注释的科学图表、逐步绘制故事板)。这要求在 packed-sequence 格式中进一步细化文本块与视觉块间的动态依赖关系,并设计支持流式交错的推理调度策略。 —- ### 3. 强化学习后训练(RL Post-Training) 论文目前依赖监督微调(SFT)对齐指令遵循行为。引入**基于人类反馈或模型反馈的强化学习**(如 RLHF、DPO、GRPO 等)有望显著提升生成图像的美学质量、提示遵循度以及物理正确性。特别是在科学图表生成(GenExam、VeriphyT2IBench)等需要精确结构化约束的任务上,强化学习可通过细粒度奖励信号优化离散视觉 token 的采样策略。 —- ### 4. 骨干网络规模的扩展(Backbone Scaling) 当前模型基于 16B 总参数 / 1.4B 激活参数的 MoE 主干。论文指出未来计划**扩大骨干规模**。鉴于多码本视觉序列在 1024 × 1024 分辨率下已达 4,160 个位置,扩展专家数量、隐藏维度或 Transformer 层数,并配合更高效的 MoE 负载均衡策略,可能进一步缩小与专用大规模生成模型(如 Qwen-Image、Seedream)的保真度差距。 —- ### 5. 解码效率与长视觉序列优化 多码本 AR Head 在每个视觉位置内部引入 K=8 步串行解码,高分辨率下视觉序列长度可达数千,造成显著推理开销。未来可从以下角度优化: - **子采样或渐进式码本解码**:是否所有位置都必须在每一步推断全部 K 个代码,或可采用早期退出(early exiting)机制? - **更激进的缓存策略**:当前 V-verify 缓存刷新基于隐藏状态漂移 Dt ,可探索基于梯度或信息论的缓存更新准则,进一步减少主干网络前向传播次数。 —- ### 6. 多码本表征与依赖建模的深化 论文验证了沿码本索引 k=1,dots,K 自回归建模 p(c_k mid h, c(<k)) 优于条件独立假设。但码本间的依赖结构可能存在更优的分解方式: - **可学习的码本解码顺序**:固定顺序 1 to K 未必最优,可引入排列自回归或图神经网络建模码本间依赖。 - **动态码本深度**:不同视觉复杂度区域可能需要不同数量的有效码本,可探索自适应深度选择(类似 RQ-Transformer 的提前停止)以降低平均解码成本。 —- ### 7. 跨模态编辑能力的统一扩展 当前图像编辑数据涵盖十种编辑类型,但视频编辑与 3D 编辑尚未纳入。将编辑框架扩展至视频(时序一致性编辑)和 3D(部件级几何/材质编辑)需要解决源模态与目标模态在离散空间中**变化区域定位**与**未变化区域保持**的精细控制问题,可能需在注意力掩码中显式引入编辑掩码(edit mask)token。 —- ### 8. 物理 grounded 生成的细粒度提升 尽管 Intern Lumina U2 在 VeriphyT2IBench 上表现优异,但在 GenExam 的 strict 评分(2.7)上仍有较大提升空间。未来可探索: - **显式物理约束注入**:将物理引擎或符号约束融入离散扩散采样过程,而非仅依赖数据驱动学习。 - **专用科学图表预训练**:构建更大规模的物理、化学、工程图表-文本对,并引入领域特定的结构验证器作为辅助监督。 —- ### 9. 数据构建与课程学习的改进 论文采用五阶段渐进课程,但阶段间的数据混合比例(Table 2)主要基于经验设定。未来可研究: - **基于能力的自动课程(Automatic Curriculum Learning)**:根据模型在各能力维度的实时评估分数动态调整数据采样权重,而非固定比例。 - **合成数据的质量-多样性权衡**:当前使用 FLUX、Ideogram 等模型合成图像,如何抑制合成 artifacts 在多轮训练中的累积放大,是一个开放问题。 —- ### 10. 与连续表征范式的深度融合 论文坚持完全离散空间以统一优化目标,但连续扩散在极高分辨率生成上仍具优势。未来可探索**混合目标训练**:在主干保持离散分类目标的同时,为视觉输出头引入连续辅助目标(如预测代码嵌入而非仅分类分布),或研究离散-连续双空间对齐的蒸馏框架,以兼顾优化一致性与生成保真度上限。 Q6: 总结一下论文的主要内容 **Intern Lumina U2** 提出了一种基于**多码本离散视觉 token**的统一扩散语言模型,旨在以单一训练目标和共享优化空间,同时实现高保真视觉生成与 omni-visual(图像/视频/3D)理解。以下从问题背景、核心方法、系统实现与实验验证四个层面进行总结。 —- ### 1. 研究背景与核心问题 统一多模态模型需在单一框架内整合视觉理解与生成,但现有方案存在根本权衡: - **离散范式**(如 Chameleon、Emu3、LLaDA2.0-Uni)将视觉压缩为单码本离散 token,天然兼容文本建模,但每个空间位置的表征容量严重受限,难以支持精细重建与复杂视觉任务; - **连续/混合范式**(如 Transfusion、BAGEL、InternVL-U)保留连续视觉状态以获得高保真生成,但文本与图像在异构空间中被优化,往往依赖不同的训练目标(如 next-token loss vs. diffusion/flow matching),导致模态间优化冲突与表征空间不对齐。 因此,论文瞄准的核心问题是:**如何在单一主干网络的共享离散优化空间内,高效建模高容量视觉表征与文本内容?** —- ### 2. 核心方法:完全离散空间中的多码本建模 论文提出 **Intern Lumina U2**,以 AToken 为视觉 tokenizer、LLaDA2.0-mini 为扩散语言模型主干,实现了视觉与文本在完全离散空间中的统一。 #### 2.1 多码本视觉表征 - **统一 4D 代码空间**:AToken 将图像、视频、3D 资产编码为共享的离散潜在空间。每个视觉位置由 K=8 个互补码本的代码组表示:

c = (c_1, dots, c_K), quad c_k ∈ 1, dots, 4096
这种表示在不增加空间网格长度的前提下,将视觉信息容量扩展数倍。 #### 2.2 输入接口:码本身份保持 - 为每个码本配备独立的嵌入表 E_k ,通过拼接与线性投影将 K 个代码合并为单一主干 token:

x = W_(∈) · [E_1(c_1); dots; E_K(c_K)]
该设计保持序列长度与 K 无关,避免视觉代码混入文本词表。 #### 2.3 输出接口:多码本自回归头 - 轻量级的 6 层因果 Transformer(multi-codebook AR head)沿码本维度自回归建模依赖:

p(c mid h) = prod(k=1)^(K) p(c_k mid h, c(<k))

  • **训练**:采用 teacher forcing,单次前向传播并行获得所有 K 个 logits; - **推理**:每个位置内部 K 步顺序采样,不同空间位置始终并行处理。 - 扩散主干负责空间位置的全局双向并行细化,AR head 负责单点内部码本依赖,二者职责解耦。 #### 2.4 统一数据格式与任务感知注意力 - 所有任务(理解、生成、编辑、纯文本)被格式化为统一对话序列,并标注五类元数据(类型、轮次、噪声状态、块索引、位置 ID)。 - **任务感知注意力掩码**精确控制不同模态与任务间的信息流动: - 文本答案采用**块级扩散**(块内双向、块间因果,带 clean copy 机制); - 视觉答案采用**空间掩码扩散**(答案内全双向,无前向 clean copy)。 #### 2.5 统一训练目标 所有模态共享单一的掩码扩散交叉熵损失:

L = ∑(i ∈ T) ell(txt)(i) + ∑(j ∈ V) ell(vis)(j)|T| + |V| + 10^(-3) L_(balance)
其中视觉位置损失为 K 个码本交叉熵的均值。文本与视觉在相同归一化下联合优化,避免异构目标冲突。 —- ### 3. 训练与推理系统 #### 3.1 五阶段渐进训练 | 阶段 | 分辨率 | 关键内容 | |:—-|:—-|:—-| | Vision–Language Alignment | 256^2 | 视觉模块对齐与初步图文能力 | | Perception Pretraining | 512^2 | 细粒度感知(OCR、文档、图表) | | Omni-Visual Mid-Training I | 1024^2 | 隔离分辨率过渡,稳定现有能力 | | Omni-Visual Mid-Training II | 1024^2 | 引入编辑、视频理解、3D 理解 | | Instruction SFT | 1024^2 | 高质量指令微调,对齐任务遵循行为 | #### 3.2 跨平台全栈实现 - **主模型**:在 1,024 张华为 Ascend 910B NPU 上完成全栈训练、推理与评估; - **参考模型**:在 NVIDIA GPU 上以相同架构与优化 recipe 进行小规模交叉验证,验证跨平台有效性。 #### 3.3 稳定性感知推理加速(PHV) 针对视觉 token 置信度绝对值低、演化缓慢的特点,提出 **Plateau-Halton sampling with V-verify (PHV)**: - **Plateau 检测**:通过置信度指数移动平均的连续低变化,识别已达稳定的 token 并提前提交; - **Halton 揭示顺序**:对未稳定位置按低差异序列均匀分散提交,优化空间覆盖; - **V-verify 缓存刷新**:基于隐藏状态归一化漂移 D_t 触发主干缓存刷新,避免过时表征累积。 该策略将 1024 × 1024 图像生成的端到端延迟从 159.9 秒降至 **35.9 秒(4.45× 加速)**,同时提升生成质量。 —- ### 4. 实验验证与关键结论 #### 4.1 Omni-Visual 理解 - **图像理解**:在 10 项基准中 6 项最优,CharXiv、HallusionBench、MathVision、DynaMath 领先显著;较同骨干 LLaDA2.0-Uni(单码本)8/10 项提升。 - **视频理解**:Video-MME、MLVU、MVBench 三项均优于现有基线(包括专用视频模型)。 - **3D 理解**:3D MM-Vet 上超过 ShapeLLM、PointLLM 等 3D 专用模型。 #### 4.2 图像生成与编辑 - 在 DPG-Bench、UniGenBench、TIIF-Bench 上取得统一模型最优,GenEval 与最优持平; - UniGenBench 中逻辑推理(Log)与语法(Gra)维度优势突出; - 图像编辑(ImgEdit、GEdit-Bench)接近最优统一模型。 #### 4.3 科学图表生成 - GenExam Physics 宽松得分在通用统一模型中最高(44.8); - VeriphyT2IBench 局部最优(76.58)、全局第二(81.22),在困难电磁学提示上达 81.30。 #### 4.4 消融实验 - **输入接口**:独立嵌入表设计(D3)在 next-code 准确率与收敛速度上显著优于共享大词表(D1)与偏移共享表(D2); - **输出头**:显式建模码本依赖的多码本 AR Head,在 GenEval 上比条件独立并行头高 14 分,尤其在计数、位置与属性绑定任务上优势明显; - **采样策略**:PHV 超参 δ_p=0.18, f_p=0.005 位于速度-质量前沿。 —- ### 5. 总结 Intern Lumina U2 首次在完全离散空间内实现了基于多码本视觉状态的统一扩散语言建模,通过“空间并行、码本串行”的因子化设计,以单一训练目标兼容了高容量视觉表征与文本建模,并在一系列理解与生成任务中验证了该范式的有效性与广度。 Q7: 想要进一步了解论文 以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文,请点击 这里 为你跳转Kimi AI网页版,并启动一个与该论文相关的新会话。

Authors: Qian Yu, Zongyang Qiu, Juncheng Yan, Hongliang Li, Qi Qin, Keqi Wang, Siqi Luo, Junkui Ju, Yapeng Mi, Tiantian Lu, Tian Liu, Pengyu Xie, Chenhao Ji, Minghui Zhang, Yiang Zhao, Siyu Yan, Haiying Xu, Yanbo Wang, Jianghan Shen, Yuandong Pu, Jiayi Lei, Borui Zhang, Jinxin Shi, Bo Zhang, Yuhang Zang, Shenglong Ye, Lixin Gu, Hongjie Zhang, Linghan Liu, Xinyuan Chen, Yuewen Cao, Lihe Zhang, Xiaohong Liu, Junjun He, Qiangwei Liu, Qi Zhang, Yu Cheng, Tianfan Xue, Hui Xiong, Chang Xu, Yu Qiao, Yihao Liu, Bin Fu

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20818.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20818

Published: 2026-10-19T01:40:24.655Z


10. MARS - A foundational map auto-regressor

Abstract:Map generation tasks feature extensive non-structural vectorized data (e.g., points, polylines, and polygons) and thus pose significant challenges to common pixel-wise generative models. Conventional approaches use multiple stages, first segmenting these features at the pixel level and then performing vectorized post-processing, with errors and complexity compounding at each stage. Motivated by the recent success of auto-regressive language modeling, we propose the first map foundation model, named Map Auto-Regressor (MARS), that is capable of generating both multi-polyline road networks and polygon buildings in a unified manner. For training MARS, we collected to our knowledge the largest multi-class map extraction dataset totaling 3.4M examples, which we call MAP-3M. Across four road and building datasets, MARS outperforms or matches the performance of multistage baselines. Additionally, we develop a “Chat with MARS” capability that enables interactive human-in-the-loop map generation and correction, supported by the auto-regressive nature of our end-to-end approach. We release our MAP-3M dataset and project demo page at (1) this https URL and (2) this https URL, respectively.

中文摘要

翻译失败: Translation limit exceeded. Please try it again later.
Response status: 401 (Unauthorized)
Response body : “{\”ShowCaptcha\”:false}”

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.20843 (HTTP 429)

Authors: Qi Zhang, Suvam Bag, Rupanjali Kukal, Lucas Thai, Jacob Kovarskiy, Mike Bianco, Xiaojie Guo, Raphael Tang, Mikael Figueroa, Rishi Madhok, Nikolaos Karianakis, Fuxun Yu

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.20843.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.20843

Published: 2026-10-19T01:40:24.655Z