数据来源:ArXiv Domain

LLM Domain Papers

1. PRO-Step: Step-level Process Reward Optimization for Retrieval-Augmented Generation

Abstract:Retrieval-Augmented Generation enhances Large Language Models by grounding responses in external knowledge, but multi-hop reasoning remains vulnerable to error propagation, where early retrieval failures confound subsequent steps. Standard outcome-based optimization only rewards the final answer, leaving intermediate retrieval and reasoning errors undetected. While existing process-based methods introduce step-level signals, they still score each step against the final answer, rewarding spurious successes where flawed retrieval coincidentally produces the correct answer. Step-level supervision in RAG requires evaluating both logical validity and evidential grounding at each step. We introduce PRO-STEP: we train a generative PRM that evaluates both dimensions, employ PRM-guided value tree search to construct preference pairs contrasting valid steps against flawed ones, and optimize the policy via step-level Direct Preference Optimization. Experiments on single and multi-hop QA datasets demonstrate that PRO-STEP achieves the best average EM and F1 across five benchmarks. Code, models, and training data are publicly available at this https URL.

中文摘要

摘要:检索增强生成(Retrieval-Augmented Generation,RAG)通过将响应基于外部知识来增强大型语言模型,但多跳推理仍然容易受到错误传播的影响,其中早期的检索失败会干扰后续步骤。标准的基于结果的优化仅奖励最终答案,而忽略了中间检索和推理错误。虽然现有的基于过程的方法引入了步骤级信号,但它们仍然是将每个步骤的得分与最终答案进行比较,在检索有缺陷但偶然产生正确答案时仍会奖励虚假的成功。RAG 中的步骤级监督需要在每一步评估逻辑有效性和证据依据。我们提出了 PRO-STEP:训练一个生成式 PRM 来评估这两个维度,采用 PRM 指导的价值树搜索构建有效步骤与有缺陷步骤的偏好对,并通过步骤级直接偏好优化(Direct Preference Optimization)来优化策略。在单跳和多跳问答数据集上的实验表明,PRO-STEP 在五个基准测试中达到了最佳的平均 EM 和 F1 分数。代码、模型和训练数据可在此 https URL 公共获取。

LLM Analysis

Q: 这篇论文试图解决什么问题? 该论文旨在解决\*\*检索增强生成(RAG)系统中多跳推理的错误传播与优化信号稀疏性问题\*\*,具体可分解为以下三个层面: ### 1. 多跳推理中的错误传播 传统RAG采用线性的"检索-生成"范式,难以处理需要迭代检索与动态查询重构的多步推理任务。在链式推理过程中,早期的检索失败(如查询不精确、召回无关文档)会沿着后续推理步骤不断放大,导致最终答案错误,而现有方法缺乏对这类中间故障的检测与修正机制。 ### 2. 基于结果优化的固有缺陷 当前主流的强化学习方法(如Search-R1、R1-Searcher)仅使用\*\*最终答案的正确性\*\*作为奖励信号,这带来了两个关键问题: - \*\*奖励稀疏\*\*:模型仅在完成整个推理链后才获得反馈,收敛所需训练数据量大,样本效率低。 - \*\*伪成功(Spurious Successes)\*\*:有缺陷的中间检索或推理可能偶然推导出正确答案,结果监督无法识别并惩罚这些逻辑上不严谨、证据上不可靠的中间步骤,从而鼓励了"侥幸正确"的错误行为模式。 ### 3. 现有过程奖励模型(PRM)与RAG场景的不适配 尽管过程级监督在数学推理等领域取得进展,但将其应用于RAG仍面临挑战: - \*\*缺乏事实依据验证\*\*:现有PRM主要验证逻辑一致性,无法检验某一步的声明是否真实基于外部检索到的文档(evidential grounding)。 - \*\*监督信号单一\*\*:多数PRM仅输出标量分数,不提供错误原因解释,反馈信息丰富度不足。 - \*\*仍依赖最终结果\*\*:现有过程级RAG方法在评估中间步骤时,仍通过模拟完整轨迹至结束、以最终F1分数作为步骤质量代理,本质上仍未摆脱对最终答案的依赖,无法有效过滤前述的伪成功。 ### 核心解决思路 为此,论文提出 \*\*PRO-STEP\*\* 框架,其核心创新在于: - 训练一个\*\*生成式过程奖励模型(Generative PRM)\*\*,在每个步骤同时评估\*\*逻辑有效性(logical validity)\*\*与\*\*证据依据(evidential grounding)\*\*,并输出可解释的判断理由; - 利用该PRM引导\*\*值树搜索(Value Tree Search)\*\*,主动构建"有效步骤 vs. 有缺陷步骤"的对比偏好对,显式剔除那些虽最终结果正确但中间过程谬误的伪成功轨迹; - 通过\*\*步骤级直接偏好优化(Step-level DPO)\*\*训练策略模型,使模型在检索与推理交织的每一步都能获得稠密、可解释的反馈,从而缓解奖励稀疏性并抑制错误传播。 Q2: 有哪些相关研究? 该论文的相关研究主要围绕\*\*检索增强生成(RAG)\*\*与\*\*过程奖励模型(PRM)\*\*两条主线展开,可系统梳理如下: --- ### 1. 检索增强生成(RAG) #### 1.1 早期检索-生成范式 早期系统遵循固定的"先检索、后生成"流程,将固定段落直接附加于输入提示(Lewis et al., 2020)。此类方法对单跳查询有效,但难以应对需要迭代检索与动态推理的多跳场景。 #### 1.2 自适应检索与多跳推理 为突破线性流程的局限,后续研究提出了多种启发式自适应机制: - \*\*FLARE\*\*(Jiang et al., 2023):基于不确定性触发检索; - \*\*DRAGIN\*\*(Su et al., 2024):动态查询重构; - \*\*IRCoT\*\*(Trivedi et al., 2023):将检索与思维链推理交错进行; - \*\*Self-RAG\*\*(Asai et al., 2024):通过自反思机制决定检索时机。 #### 1.3 基于强化学习的RAG 近期工作将搜索引擎纳入大语言模型的交互环境,采用端到端强化学习进行优化: - \*\*Search-R1\*\*(Jin et al., 2025a)、\*\*R1-Searcher\*\*(Song et al., 2025)、\*\*AutoRAG\*\*(Yu et al., 2024):均以\*\*最终答案正确性\*\*作为奖励信号,验证了 outcome-based RL 可增强自主检索能力; - \*\*Search-o1\*\*(Li et al., 2025a):结合智能体搜索增强推理模型。 然而,这些框架仅依赖最终结果的稀疏奖励,无法检测中间检索或推理步骤中的错误。 #### 1.4 过程级监督的RAG 部分近期研究尝试引入步骤级信号: - \*\*R3-RAG\*\*(Li et al., 2025b)、\*\*ProRAG\*\*(Wang et al., 2026)、\*\*StepSearch\*\*(Zheng et al., 2025b)、\*\*HiPRAG\*\*(Wu et al., 2026):通过步骤级奖励优化检索增强生成; - \*\*GRAT\*\*(Peng and Wei, 2025)、\*\*ReasonRAG\*\*(Zhang et al., 2025a):利用过程奖励或树搜索指导RAG。 \*\*现有局限\*\*:上述方法在评估中间步骤时,仍主要依赖模拟完整轨迹至结束后的\*\*最终F1分数\*\*作为步骤质量代理。这种设计无法惩罚"伪成功"(spurious successes)——即有缺陷的检索或推理偶然推导出正确答案的情形。 --- ### 2. 过程奖励模型(PRM) #### 2.1 数学推理中的PRM PRM最初在形式化推理领域取得突破: - \*\*Lightman et al. (2024)\*\*:首次证明基于PRM的逐步验证可提升数学推理性能,通过独立验证每一步引导模型沿正确轨迹推理。 #### 2.2 自动化过程监督 为降低人工标注成本,后续工作探索了自动化PRM训练: - \*\*OmegaPRM\*\*(Luo et al., 2024):利用蒙特卡洛展开自动标注步骤标签; - \*\*Math-Shepherd\*\*(Wang et al., 2024):通过步骤级结果估计实现自动化过程监督。 #### 2.3 生成式PRM 超越标量评分范式,近期研究强调可解释的过程监督: - \*\*GenPRM\*\*(Zhao et al., 2025):提出生成式范式,模型在给出正确性标签前先显式生成判断依据(rationale),从而提供更丰富的反馈信号。 #### 2.4 多领域PRM扩展 - \*\*VersaPRM\*\*(Zeng et al., 2025):将PRM扩展至多领域场景。 \*\*现有局限\*\*:当前PRM研究仍集中于逻辑可验证的形式化推理(如数学)。RAG环境具有显著不同的挑战:每一步均交织内部推理与外部检索,需要同时验证\*\*逻辑有效性\*\*(logical validity)与\*\*证据依据\*\*(evidential grounding)。现有PRM既无法检验声明是否基于检索文档,也未充分处理检索失败后的错误恢复与策略调整。 --- ### 3. 与PRO-STEP的定位关系 | 研究类别 | 代表工作 | 核心局限 | PRO-STEP的改进 | |---------|---------|---------|--------------| | Outcome-based RL for RAG | Search-R1, R1-Searcher, AutoRAG | 仅使用最终答案奖励,信号稀疏,无法检测中间错误 | 引入步骤级稠密反馈,直接优化中间检索与推理质量 | | Process-level RAG | ReasonRAG, StepSearch, HiPRAG | 仍基于最终F1分数评估中间步骤,纵容伪成功 | 通过生成式PRM独立评估每步的逻辑与证据,显式过滤伪成功 | | 通用/数学PRM | Lightman et al., Math-Shepherd, GenPRM, VersaPRM | 仅验证逻辑一致性,缺乏事实依据检验;未考虑检索-推理依赖 | 训练专用的RAG生成式PRM,联合评估逻辑有效性与检索证据 grounding,并支持错误恢复场景 | Q3: 论文如何解决这个问题? 该论文提出 \*\*PRO-STEP\*\* 框架,通过\*\*生成式过程奖励建模\*\*与\*\*步骤级直接偏好优化\*\*的协同,系统性地解决上述问题。整个框架分为两个紧密衔接的阶段,以下按技术流程展开说明。 --- ### 总体架构 PRO-STEP 的核心思想是:在 RAG 的每一个中间步骤同时检验\*\*逻辑有效性\*\*(logical validity)与\*\*证据依据\*\*(evidential grounding),并基于步骤级的对错信号构建对比偏好数据,最终通过步骤级 DPO 训练策略模型。框架包含两大阶段: 1. \*\*阶段一:过程奖励模型(Process Reward Modeling)训练\*\* 2. \*\*阶段二:策略优化(Policy Optimization)\*\* --- ### 阶段一:训练生成式过程奖励模型(Generative PRM) #### 1. 交错式轨迹生成 将策略模型 π_θ 与外部搜索引擎 E 进行多轮交互采样。 π_θ 在 \`...\` 中生成内部推理;当需要外部事实支撑时,生成 \`...\` 触发检索,系统将召回的文档以 \`...\` 形式追加到序列中。该过程持续至模型输出 \`...\` 或达到最大交互轮次。论文从 HotpotQA 与 MuSiQue 中采样,最终保留 31,728 条完整推理轨迹。 #### 2. 步骤级标注(Step Annotation) 使用 QwQ-32B 作为标注器,在\*\*不依赖最终答案\*\*的前提下,对每一步 s_t 依据六项准则(实体/关系、查询质量、推理质量、答案质量、错误恢复、过度自信)生成: - 二元正确性标签 r_t ∈ 0, 1 ; - 自然语言判断理由 v_t 。 与既有 PRM 方法不同,该标注\*\*保留完整轨迹\*\*(包括错误后的恢复步骤),因为 RAG 场景中模型可通过修正查询从局部失败中恢复。 #### 3. 生成式 PRM 训练 基于上述数据 D_(PRM) ,将 DeepSeek-R1-Distill-8B 初始化为生成式 PRM π_psi ,优化目标为联合生成理由与标签:

L(PRM) = -E((q,τ,vt,r_t)sim D)(PRM) [ log πpsi(v_t mid q, s(≤ t)) + log πpsi(r_t mid q, s(≤ t), vt) ]
推理时,模型先自回归生成理由,再预测二分类标签:
r(s_t) = argmax
(r ∈ 0,1) πpsi(r mid q, s(≤ t), vt)
—- ### 阶段二:基于 PRM 引导值树搜索的策略优化 #### 1. PRM 引导的值树搜索(VTS) 在 AlphaGo Zero 式的树搜索框架中,将随机 rollout 替换为**直接值估计**: - 每个节点代表状态 s_t = (q, s
(≤ t)) ; - 若节点为终止节点,奖励为深度折扣 F1:

Q(s_T) = F1(s_T, y^*) · γ^(d(s_T))

  • 若为非终止节点,直接使用 PRM 的二元预测作为即时步骤价值:
    Q(s_t) = r(s_t) · γ^(d(s_t))
    其中 γ ∈ (0,1) 用于惩罚过长的推理链, d(·) 为节点深度。PRM 信号在节点扩展后回传至祖先节点,更新均值价值 Q(s) ,但不虚增访问计数,从而确保树搜索既考虑最终答案质量,又严格评估每一步的独立有效性。 #### 2. 步骤级偏好数据构建 对每个父节点的子节点,使用综合价值函数排序:

V(s) = Q(s) + α · r(s)
其中 Q(s) 聚合了后代轨迹的最终结果质量, r(s) ∈ 0,1 评估当前步骤本身是否正确。对于任意兄弟节点对 si, s_j ,若 V(s_i) - V(s_j) > δ ,则将高分节点标记为 **chosen** (s^+) ,低分节点标记为 **rejected** (s^-) 。 **关键作用**:通过显式引入 r(s) ,即使某条错误路径偶然以高 F1 结束,只要其中间步骤被 PRM 判为 0,该路径的整体 V(s) 将被拉低,从而**主动过滤伪成功轨迹**,确保对比 pair 中的 rejected 样本确实是”步骤有缺陷”而非”结果不够好”。 #### 3. 步骤级直接偏好优化(Step-level DPO) 利用上述 pair 数据,对策略模型 πθ 执行步骤级 DPO。首先定义隐式效用:

uθ(x, y_t) = β log πθ(yt mid x, y(0 均优于无 PRM | | **数据过滤消融** | Appendix B / Table 7 | 验证 outcome-margin filter(F1≥0.2 且 ΔF1≥0.2)对数据规模与性能的影响 | | **PRM 标签质量审计** | Appendix E / Table 13–14 | 人工与 Claude Opus 4.7 审计 QwQ-32B 标注,Cohen’s kappa=0.61 (实质性一致) | | **统计显著性检验** | Appendix F / Table 15–16 | Bootstrap 95% 置信区间与配对 t 检验,验证主实验提升的显著性 | | **IFBC 轨迹分析** | Appendix H / Table 18 | 量化”中间有缺陷但最终正确”(IFBC)轨迹比例;PRO-STEP 的 IFBC 率(26.8%)显著低于 Outcome-only 消融(40.3%) | | **PRM 反馈引导再生的失败案例** | Appendix K / Table 22–23 | 记录一种尝试:用 PRM 批评反馈重写失败轨迹。训练指标提升但评测崩溃(平均 EM 降 10.7),归因于训练/推理分布漂移 | Q5: 有什么可以进一步探索的点? 基于论文的局限性与实验发现,以下几个方向值得进一步探索: —- ### 1. 提升过程监督数据的质量与规模 **更强大的开源标注器**:当前框架依赖 QwQ-32B 进行步骤级标注,尽管与 Claude Opus 4.7 达成了实质性一致( kappa = 0.61 ),但在主观性较强的推理步骤( kappa = 0.12 )上分歧仍然显著。探索更新的开源推理模型或混合专家标注策略,可进一步降低标注噪声,尤其在边界案例上。 **闭源模型蒸馏与数据扩展的权衡**:论文指出,PRO-STEP 刻意避免使用闭源教师模型(如 GPT-4o),导致在最难任务 MuSiQue 上仍略逊于基于 GPT-4o 蒸馏的 StepSearch。未来可探索**混合范式**——利用闭源模型生成高难度种子数据,再通过开源 pipeline 进行过滤与增强,以在可复现性与性能上限之间取得更好平衡。 —- ### 2. 解决 PRM 反馈引导再生的分布漂移问题 论文附录 K 记录了一个关键失败案例:利用 PRM 的批评理由重写失败轨迹时,训练指标(DPO loss、margin、accuracy)显著改善,但下游评测平均下降 10.7 EM。根本原因在于再生轨迹的 `` 块变成了对反馈的元评论,推理时无此反馈通道,导致严重的分布漂移。 **未来可探索**: - **动作格式一致的增强(Action-Format-Consistent Augmentation)**:设计约束机制,确保再生轨迹在动作空间(``、``、``)的分布与推理时一致,而非仅优化答案正确性。 - **隐式反馈注入**:将 PRM 信号转换为潜在表示或结构化向量,而非自然语言批评,避免风格漂移。 —- ### 3. 扩展至多步推理以外的复杂场景 当前验证集中于单跳/多跳 QA,过程监督的潜力在以下场景尚未充分释放: - **长文档推理与大规模知识库**:当检索空间从维基百科扩展至百万级文档或长上下文时,PRM 需同时评估**信息检索的充分性**与**长程逻辑依赖性**,这对步骤边界定义与信用分配提出新挑战。 - **工具使用与多模态 RAG**:在涉及代码执行、数据库查询或图文检索的 Agentic 系统中,PRM 需扩展至验证**工具调用参数的正确性**及**跨模态证据的一致性**。 - **开放式生成与事实核查**:超越 QA 的短答案形式,探索 PRM 在长文本生成(如报告撰写)中逐句或逐段的事实 grounding 监督。 —- ### 4. 算法层面:更精细的搜索与优化策略 **高级树搜索算法**:当前 VTS 采用固定分支因子与 UCB 选择。可引入: - 动态分支预算,根据问题复杂度自适应分配计算; - 非对称树搜索,对 PRM 高置信度节点提前剪枝,对低置信度节点加深探索。 **与在线 RL 算法结合**:论文采用离线 DPO 进行策略优化。将生成式 PRM 与在线 RL(如 PPO、GRPO)结合,允许策略在训练过程中与环境持续交互,可能更有效地探索检索策略空间,并动态适应检索器的变化。 **细粒度奖励塑形**:当前 PRM 提供二元标签 0,1 与深度折扣。未来可探索**连续型过程奖励**(如基于证据重叠度的分数),以提供更细粒度的梯度信号,尤其在错误恢复的模糊地带。 —- ### 5. PRM 本身的可解释性与可信度 论文中的生成式 PRM 已输出自然语言理由,但这些理由如何**反作用于策略的可信度校准**仍待研究: - **不确定性量化**:让 PRM 同时输出对其判断的置信度,策略可据此决定是继续推理、重新检索还是请求外部验证。 - **人在回路中的交互式修正**:将 PRM 的批评理由呈现给终端用户,允许用户介入纠正中间步骤,形成”人类-PRM-策略”三方协同的迭代优化回路。 —- ### 6. 理论分析:过程奖励与结果奖励的最优组合 论文经验性地验证了 α = 0.3 时过程奖励与结果奖励的互补性(Appendix D, Table 11),但缺乏理论指导。未来可从以下角度深入: - 在**部分可观察马尔可夫决策过程(POMDP)**框架下,形式化分析过程奖励对样本复杂度的降低作用; - 探讨在何种任务结构(如跳跃距离、检索噪声水平)下,过程奖励相对于结果奖励的边际收益最大。 Q6: 总结一下论文的主要内容 该论文提出 **PRO-STEP**,一个面向检索增强生成(RAG)的过程级监督优化框架,旨在解决多跳推理中的错误传播与结果奖励稀疏问题。以下是论文的主要内容总结: —- ## 1. 研究问题 - **错误传播**:在多步推理链中,早期的检索失败(如查询偏差、文档不相关)会沿后续步骤放大,导致最终答案错误。 - **结果奖励的局限**:现有强化学习方法(如 Search-R1)仅依据最终答案的正确性提供奖励,信号稀疏,且无法识别”伪成功”(flawed intermediate steps 偶然推导出正确答案)。 - **现有 PRM 与 RAG 不适配**:过程奖励模型(PRM)在数学推理中验证逻辑一致性,但缺乏对**外部检索证据依据(evidential grounding)**的检验,且仍依赖最终结果评估中间步骤。 —- ## 2. 方法框架:PRO-STEP PRO-STEP 采用两阶段架构,将生成式过程奖励建模与步骤级直接偏好优化相结合。 ### 阶段一:生成式过程奖励模型训练 - **交错式轨迹生成**:策略模型 π_θ 与搜索引擎多轮交互,生成 ``、``、``、`` 交错的推理轨迹。 - **步骤级标注**:使用 QwQ-32B 对每步独立标注二元标签 r_t ∈ 0,1 与判断理由 v_t ,评估逻辑有效性与证据依据;**保留含错误恢复的完整轨迹**(RAG 中模型可通过后续检索修正早期错误)。 - **PRM 训练**:将 DeepSeek-R1-Distill-8B 训练为生成式 PRM π_psi ,联合优化理由生成与标签预测:

L(PRM) = -E((q,τ,vt,r_t)sim D)(PRM) [ log πpsi(v_t mid q, s(≤ t)) + log πpsi(r_t mid q, s(≤ t), v_t) ]

阶段二:策略优化 - **PRM 引导的值树搜索(VTS)**:以 PRM 替代随机 rollout,节点价值由步骤级 PRM 信号直接估计:

Q(st) = r(s_t) · γ^(d(s_t))
其中 γ ∈ (0,1) 惩罚过长轨迹, d(·) 为节点深度。 - **偏好数据构建**:综合回溯价值与步骤正确性,定义节点排序函数:
V(s) = Q(s) + α · r(s)
通过设定阈值 δ ,将同层节点划分为 chosen( s^+ )与 rejected( s^- ),**显式过滤伪成功路径**(即使最终 F1 高,中间步骤 flawed 的轨迹仍被标记为 rejected)。 - **步骤级 DPO**:利用上述 pair 数据,掩码 `` 块后执行步骤级直接偏好优化:
L
(DPO)(θ) = -E(D)(dpo) [ log σ( uθ(x, y_t^w) - uθ(x, y_t^l) ) ]
—- ## 3. 实验结果 论文在 **5 个基准**(PopQA、HotpotQA、2WikiMultiHopQA、Bamboogle、MuSiQue)上开展实验: - **主性能**:PRO-STEP 在平均 EM 与 F1 上优于所有对比方法(包括 Zero-shot、高级自适应 RAG、Outcome-based RL 及现有 Process-based 基线),在 2WikiMultiHopQA 等复杂多跳任务上优势尤为显著。 - **PRM 必要性消融**:去除 PRM 后平均 EM/F1 显著下降,证实仅靠最终 F1 无法有效惩罚中间错误。 - **优化策略对比**:DPO 显著优于 SFT(缺乏对比信号)与 KTO(非成对信号不稳定)。 - **PRM 轨迹重排序**:与 VersaPRM、Math-PRM 相比,PRO-STEP 的 PRM 在扩大候选池时能有效识别幻觉轨迹,WMV/BoN 性能持续提升或稳定。 - **检索行为分析**:PRO-STEP 能自适应调整搜索深度,有效利用额外检索文档( k=3 to 5 持续提升),并在检索失败场景下展现出更强的错误恢复能力。 - **跨模型泛化**:在 Llama-3.1-8B、Qwen2.5-7B Base 及 Qwen2.5-3B 上均取得一致提升,验证框架的通用性。 - **IFBC 分析**:PRO-STEP 的”中间有缺陷但最终正确”(IFBC)轨迹比例(26.8%)显著低于无 PRM 的 Outcome-only 方法(40.3%),直接验证了过程监督对过滤伪成功的有效性。 —- ## 4. 主要贡献 - 识别出现有过程级 RAG 方法仍依赖最终结果优化、无法惩罚伪成功的关键缺陷。 - 提出首个面向 RAG 的生成式 PRM,同步评估**逻辑有效性**与**证据依据**,并支持错误恢复场景。 - 通过 PRM 引导的 VTS 显式构造高质量步骤级偏好对,结合 DPO 实现细粒度策略优化,缓解结果奖励稀疏性。 - 在单跳与多跳 QA 基准上取得最佳平均性能,并系统分析了跨模型、跨规模及检索失败恢复等场景。 —- ## 5. 局限与未来方向 - 依赖 QwQ-32B 进行开源可复现的标注,在最难任务(如 MuSiQue)上略逊于使用闭源 GPT-4o 蒸馏的基线。 - Bamboogle 测试集规模较小(125 例),统计可靠性受限。 - PRM 反馈引导的轨迹再生会导致训练/推理分布漂移,提示未来需探索**动作格式一致**的增强策略。 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: MinKeon Kim, Namjun Lee, Jaekwang Kim

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01658.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01658

Published: 2026-09-04T01:21:25.500Z


2. Learning Evidence Sufficiency Boundaries for Selective Answering in Grounded Multi-Hop QA

Abstract:Grounded question answering systems should answer only when the supplied evidence supports the answer. In multi-hop QA, this requirement is difficult because partial evidence can make an unsupported answer appear plausible. We study selective answering through evidence sufficiency boundaries: for the same question, a model should abstain under unsupported or partially supported context, answer when the context first becomes sufficient, and keep the answer stable when redundant evidence is added. We introduce Evidence Sufficiency Boundary Training, a generation-native training framework that constructs ordered evidence chains and supervises the abstain-to-answer transition directly. The method combines level supervision, a boundary flip margin, post-boundary stability, and answer recall protection. We build evidence chains from HotpotQA, 2WikiMultiHopQA, and MuSiQue, then evaluate models with chain metrics, raw QA utility, and unsupported-answer rates on external non-answerable sets. With Qwen2.5-3B-Instruct and LoRA adaptation, Evidence Sufficiency Boundary Training gives the strongest boundary localization among the tested systems, with flip accuracy of 0.807 compared with 0.781 for a token-level abstention baseline. It also achieves the lowest overall unsupported-answer rate on external non-answerable evaluation, 0.095 compared with 0.101 for the same baseline, while retaining competitive raw QA F1. The results show that grounded selective answering improves when training marks the evidence level where refusal should give way to answering.

中文摘要

摘要:基于证据的问答系统应仅在提供的证据支持答案时作答。在多跳问答中,这一要求较为困难,因为部分证据可能使不被支持的答案看起来合理。我们通过证据充分性边界研究选择性作答:对于同一个问题,当上下文不被支持或仅部分被支持时,模型应选择回避;当上下文首次变得充分时,模型应作答;当添加冗余证据时,答案应保持稳定。我们提出了证据充分性边界训练(Evidence Sufficiency Boundary Training),这是一种生成原生的训练框架,通过构建有序证据链并直接监督从回避到作答的过渡。该方法结合了等级监督、边界翻转余量、边界后稳定性以及答案回溯保护。我们从HotpotQA、2WikiMultiHopQA和MuSiQue构建证据链,然后通过链指标、原始问答效用以及在外部不可答集合上的不被支持答案率评估模型。使用Qwen2.5-3B-Instruct和LoRA适配时,证据充分性边界训练在所测试系统中实现了最强的边界定位能力,翻转准确率为0.807,而基于token的回避基线为0.781。在外部不可答评估中,它还实现了最低的总体不被支持答案率,为0.095,而相同基线为0.101,同时保持了有竞争力的原始问答F1分数。结果表明,当训练标记出何时应由拒答转为作答的证据水平时,基于证据的选择性作答性能会得到提升。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01687 (HTTP 429)

Authors: Haruto Sato, Yuki Tanaka, Ren Nakamura, Aoi Kobayashi, Mei Ito

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01687.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01687

Published: 2026-09-04T01:21:25.500Z


3. SpeakPay: Domain-Adaptive LoRA Fine-Tuning of Whisper for Low-Resource Nepali Financial Speech Recognition

Abstract:Mobile payment applications in Nepal are graphically mediated and largely inaccessible to visually impaired users. This paper presents SpeakPay, a voice-first digital wallet, and documents the central technical contribution: a controlled study of domain adaptation for low-resource financial speech recognition. We introduce NepFinSpeech-403, a 403-utterance dataset of Nepali financial voice commands (send, load, and balance operations spanning 237 unique numerals), and fine-tune Whisper large-v2 with LoRA. On the held-out test set, the domain-adapted model reduces Word Error Rate from 129.95% (zero-shot baseline) to 42.58% —- a 67.2% relative reduction —- and improves Devanagari numeral recognition accuracy from 0.0% to 73.9%. We find that word-level metrics understate the practical task-level impact: domain adaptation improves the Transaction Success Rate from 1.67% to 33.33%, a roughly 20x gain. The improvement is consistent at the individual-utterance level (sign test, $p < 10^{-17}$) and across all command types. A data efficiency analysis shows that as few as 100 domain-specific utterances are sufficient to halve the zero-shot WER, with performance plateauing around 300 examples. Error analysis reveals systematic numeral confusion patterns (zero insertion/deletion, prefix hallucination) that account for the majority of remaining transaction failures. The trained system is deployed as a publicly accessible voice-first web application. All code, dataset, model weights, and this paper are released at this https URL.

中文摘要

摘要:尼泊尔的移动支付应用主要通过图形界面进行操作,对于视障用户来说基本上无法使用。本文介绍了 SpeakPay,一款以语音为优先的数字钱包,并记录了核心技术贡献:针对低资源金融语音识别的领域适应性控制研究。我们推出了 NepFinSpeech-403,这是一个包含 403 条尼泊尔语金融语音指令(包括发送、充值和余额操作,涵盖 237 个独特数字)的数据集,并使用 LoRA 对 Whisper large-v2 进行微调。在保留的测试集中,领域适应模型将词错误率从 129.95%(零样本基线)降低到 42.58%,相对减少 67.2%,并将天城文数字识别准确率从 0.0% 提升至 73.9%。我们发现,基于单词的指标低估了对实际任务的影响:领域适应使交易成功率从 1.67% 提升至 33.33%,约为 20 倍的增长。该改进在单条语音级别(符号检验,$p < 10^{-17}$)及所有指令类型中保持一致。数据效率分析表明,仅需 100 条领域特定语音即可将零样本词错误率减半,性能在约 300 条样本时趋于平稳。错误分析显示系统存在系统性的数字混淆模式(零插入/删除,前缀幻觉),占据了大部分剩余交易失败的原因。训练好的系统已部署为公众可访问的语音优先网页应用。所有代码、数据集、模型权重及本文均在此 https URL 上发布。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01737 (HTTP 429)

Authors: Biraj Subedi

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01737.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01737

Published: 2026-09-04T01:21:25.500Z


4. MemeCULT-1K: Benchmarking South Asian Cultural Context and Humor Understanding of Multimodal Models

Abstract:Meme understanding goes beyond recognizing visual content or literal text; it requires implicit cultural knowledge and pragmatic inference that most vision-language models still lack. We introduce MemeCULT-1K, a multilingual benchmark of 1,000 South Asian memes in Bengali, English, and Hindi, where each meme is paired with a cultural context note and three human-written explanations, along with a supplementary set of 54 Bengali regional dialect memes. We evaluate thirteen popular Vision Language Models (VLMs) under two settings: meme-only and context-aware. Providing minimal cultural context yields consistent gains across all models and languages: mean SBERT similarity improves from 44.6 to 56.4 (+11.8), BLEURT from 37.3 to 42.3 (+5.0), and LLM-as-a-Judge scores from 2.57 to 3.43 out of 5 (+0.86). Fine-grained error analysis reveals that closed-source models fail mainly on entity and reference misidentification, while open-source models are bottlenecked by broader cultural knowledge gaps, with linguistic and phonological failures proving the most context-resistant across both. These results highlight the difficulty of culturally grounded meme understanding and motivate future work on explicit cultural knowledge integration. Our dataset and code are publicly available at TawsifDipto17/MemeCULT-1K.

中文摘要

摘要:理解表情包不仅仅是识别视觉内容或字面文本;它还需要隐性文化知识和语用推理,而大多数视觉-语言模型仍然缺乏这些能力。我们介绍了 MemeCULT-1K,这是一个多语言基准数据集,包含 1,000 个南亚表情包,语言包括孟加拉语、英语和印地语,每个表情包都附有文化背景说明和三条人工撰写的解释,同时还提供了 54 条孟加拉地区方言表情包的补充集合。我们在两种设置下评估了十三个流行的视觉语言模型(VLM):仅表情包和具备上下文感知。提供最少量的文化背景会在所有模型和语言中带来一致的提升:平均 SBERT 相似度从 44.6 提升至 56.4(+11.8),BLEURT 从 37.3 提升至 42.3(+5.0),LLM 评分(5分制)从 2.57 提升至 3.43(+0.86)。细粒度错误分析显示,闭源模型主要在实体和引用识别上失败,而开源模型的瓶颈在于更广泛的文化知识缺口,语言和语音方面的失败在两者中都最难通过上下文弥补。这些结果突显了基于文化的表情包理解的复杂性,并激励未来在显性文化知识整合方面的研究。我们的数据集和代码可在 TawsifDipto17/MemeCULT-1K 公开获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01772 (HTTP 429)

Authors: Tawsif Tashwar Dipto, Mehedi Ahamed, Radib Bin Kabir, Mueeze Al Mushabbir, Mohammed Saidul Islam, Mir Rayat Imtiaz Hossain, Md Tahmid Rahman Laskar, Sabbir Ahmed

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01772.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01772

Published: 2026-09-04T01:21:25.500Z


5. VakyArth: Evaluating Pragmatic Competence in LLMs across Indic Languages

Abstract:Real-world communication often requires pragmatic reasoning: interpreting meanings implied through context and cultural convention rather than stated literally. Existing pragmatic evaluation remains largely limited to English and high-resource languages, leaving Indic languages unexplored despite their linguistic and cultural diversity. We introduce VakyArth, the first pragmatic benchmark for Indic languages, designed as a diagnostic evaluation covering Hindi, Punjabi, Tamil, and Malayalam. VakyArth evaluates models across five phenomena: deixis, speech acts, implicature, social pragmatics, and coherence; through multiple-choice questions, natural language inference, and translation, with all items authored by native speakers. Across multilingual large language models (LLMs) of varying families and sizes, we find consistent failures on pragmatic meanings rooted in Indic linguistic and cultural conventions. Our analysis shows systematic differences across languages and tasks: MCQ accuracy exceeds NLI accuracy in all model-language combinations, translation performance does not reliably track pragmatic understanding, and Indo-Aryan languages show a translation advantage over Dravidian languages. We further show that automatic translation metrics can miss fluent but pragmatically unfaithful outputs, especially for implicature and deixis.

中文摘要

摘要:现实世界的交流常常需要语用推理:通过上下文和文化习惯理解隐含的意义,而不仅仅是字面上的表达。目前的语用评估在很大程度上仍局限于英语和资源丰富的语言,而印度语言尽管具有语言和文化多样性,却鲜有研究。我们介绍了 VakyArth,这是首个面向印度语言的语用基准测试,设计为诊断性评估,涵盖印地语、旁遮普语、泰米尔语和马拉雅拉姆语。VakyArth 评估模型在五种语用现象上的表现:指示、言语行为、含义、社会语用以及连贯性;评估方式包括多项选择题、自然语言推理和翻译,所有题目均由母语者编写。在不同系列和规模的多语言大型语言模型(LLM)中,我们发现模型在基于印度语言和文化习惯的语用意义上普遍表现失败。我们的分析显示,在语言和任务之间存在系统性差异:在所有模型-语言组合中,多项选择题的准确率均高于自然语言推理的准确率,翻译性能并不能可靠反映语用理解,而印欧语系语言在翻译上优于达罗毗荼语系语言。我们进一步显示,自动翻译指标可能遗漏语法流畅但语用上不准确的输出,尤其是在含义和指示方面。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01788 (HTTP 429)

Authors: Usneek Singh, Poorvaja Veera Balaji Kumar, Parth Nanda, Anand Madhusoodanan, Geyang Guo, Wei Xu, Junyi Jessy L

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01788.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01788

Published: 2026-09-04T01:21:25.500Z


6. Disentangling Statistical Preemption from Entrenchment in Language Models’ Avoidance of Overgeneralization

Abstract:How do learners avoid overgeneralizations such as Tom laughed me without explicit negative evidence? Constructionists have posited two proposals that describe indirect negative evidence against overgeneralizations: preemption (which privileges exposure to near-synonymous construction—-e.g., she made him laugh) vs. entrenchment (all exposures to a verb’s grammatical usages, including cases like He laughed). We disentangle these hypotheses by running controlled rearing experiments on LMs trained on child-caregiver conversations, where we systematically remove preemptive vs. non-preemptive evidence. We find that while LMs avoid overgeneralizations, they do not show preemption at a verb-specific level, instead showing weak but non-zero evidence of abstract preemption. Combined with results from analyzing the LMs’ training dynamics, we find that LMs treat competing structures as indirect positive—-as opposed to negative—-evidence in the verb-specific condition. Insofar as preemption is the more plausible route to avoiding overgeneralizations in humans, our results point the need for there to be sensitivities to indirect negative evidence in neural network learners, and suggest new human experiments to test abstract preemption.

中文摘要

摘要:学习者如何在没有明确否定证据的情况下避免过度概括,例如 Tom laughed me 呢?建构主义者提出了两种解释过度概括的间接否定证据的假说:替代抑制(preemption,强调接触近义构式——例如 she made him laugh)与巩固作用(entrenchment,即接触动词的所有语法用法,包括 He laughed 这样的例子)。我们通过在训练于儿童-看护者对话的语言模型(LMs)上进行受控养育实验,将替代抑制证据与非替代抑制证据系统地移除,从而解开这些假说。我们发现,虽然 LMs 避免了过度概括,但它们在动词特定层面上并不表现出替代抑制,而是显示出弱但非零的抽象替代抑制证据。结合对 LMs 训练动态的分析结果,我们发现 LMs 在动词特定条件下将竞争结构视为间接正面证据,而非负面证据。鉴于替代抑制在避免人类过度概括中更为可信,我们的结果指出神经网络学习者需要具备对间接否定证据的敏感性,并建议进行新的实验来测试抽象替代抑制。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01794 (HTTP 429)

Authors: Yixuan Wang, Freda Shi, Kanishka Misra

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01794.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01794

Published: 2026-09-04T01:21:25.500Z


7. How Do Prompt Variations Affect Energy Consumption in On-Device LLMs?

Abstract:Large language models (LLMs) are increasingly deployed on mobile devices, making energy efficiency a key deployment constraint, yet the energy impact of prompt design remains underexplored. This paper aims to understand how two prompt properties, cognitive load and phrasing pattern, shape the energy behavior of on-device LLM inference. We conduct a broad empirical study covering prompt properties, datasets, models, and devices, with phase-level profiling that separates prefill and decode energy. We find that cognitive load primarily affects the energy cost per token, while phrasing pattern affects energy largely through token usage. Our energy-quality analysis further shows that prompt design reshapes the attainable frontier differently across models, highlighting the need for model-aware prompt design in energy-efficient on-device LLM inference. Code, datasets, and scripts are available at this https URL.

中文摘要

摘要:大型语言模型(LLMs)正越来越多地部署在移动设备上,使得能效成为关键的部署约束,但提示设计对能源消耗的影响仍未得到充分研究。本文旨在理解两种提示特性——认知负荷和措辞模式——如何影响设备端LLM推理的能量行为。我们进行了广泛的实证研究,涵盖提示特性、数据集、模型和设备,并进行了阶段级别的性能分析,将预填充和解码能耗分开。研究发现,认知负荷主要影响每个词元的能量消耗,而措辞模式则主要通过词元使用量影响能耗。我们的能量-质量分析进一步表明,提示设计会在不同模型中重塑可达的前沿,强调在能效设备端LLM推理中进行模型感知提示设计的必要性。代码、数据集和脚本可通过此 https URL 获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01798 (HTTP 429)

Authors: Wei Hu, Xiaolong Tu, Dawei Chen, Yitao Chen, Kyungtae Han, Haoxin Wang

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01798.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01798

Published: 2026-09-04T01:21:25.500Z


8. TalkFa: A Unified Benchmark for Farsi Dialogue Generation and Understanding

Abstract:Farsi, spoken by more than 120 million people, lacks a comprehensive benchmark for dialogue generation and understanding. We introduce TALKFA, a unified benchmark comprising three complementary datasets: (1) WIKI-FADIAL, 4.2K Wikipedia-grounded dialogues for knowledge-grounded generation; (2) DAILYDIALOG-FA, 6.6K dialogues annotated for dialogue acts and emotions; and (3) PLAYDIAL-FA, 2.1K theatrical dialogues with sentiment labels. While LLMs assist data construction, every dialogue undergoes multi-stage review and revision by native Farsi speakers, and only the final human-approved dialogues are released. Experiments with six LLAMA and MISTRAL models show that LoRA substantially improves dialogue generation while requiring only 25-50% of the training data to recover over 90% of the final performance gains. Across classification tasks, FABERT achieves the best dialogue-act performance, LORA-MISTRAL-7B performs best on emotion recognition, and MISTRAL-24B achieves the highest sentiment score. Human evaluation and independent external validation demonstrate the reliability of the benchmark, while comparisons with GPT-4.1 as an LLM judge reveal that automatic metrics substantially overestimate dialogue quality. Zero-shot evaluation with frontier LLMs further shows that TalkFa remains a challenging benchmark. We will release all datasets, annotation guidelines, code, and checkpoints.

中文摘要

摘要:波斯语由超过1.2亿人使用,但在对话生成和理解方面缺乏全面的基准。我们引入了TALKFA,一个统一的基准,包括三组互补的数据集:(1) WIKI-FADIAL,包含4.2千条基于维基百科知识的对话,用于知识驱动生成;(2) DAILYDIALOG-FA,包含6.6千条对话,标注了对话行为和情感;(3) PLAYDIAL-FA,包含2.1千条戏剧对话,并标注了情感标签。虽然大型语言模型(LLMs)辅助数据构建,但每条对话都经过母语为波斯语的人员的多阶段审核和修订,最终仅发布经过人工批准的对话。对六个LLAMA和MISTRAL模型的实验表明,LoRA能够显著提升对话生成性能,同时仅需使用25%-50%的训练数据即可恢复超过90%的最终性能提升。在分类任务中,FABERT在对话行为识别上表现最佳,LORA-MISTRAL-7B在情感识别上性能最好,而MISTRAL-24B取得了最高的情感得分。人工评估和独立外部验证证明了该基准的可靠性,而作为LLM评审的GPT-4.1对比显示,自动评估指标显著高估了对话质量。前沿LLM的零样本评估进一步显示,TalkFa仍然是一个具有挑战性的基准。我们将发布所有数据集、标注指南、代码和模型检查点。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01810 (HTTP 429)

Authors: Neda Jamshidi, Kamyar Zeinalipour, Fahimeh Akbari, Monica Bianchini, Marco Maggini, Marco Gori

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01810.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01810

Published: 2026-09-04T01:21:25.500Z


9. AVERT: Audio-Verified Adjudication for Spoken Dialogue State Tracking

Abstract:Spoken dialogue state tracking recovers slot-value pairs from speech, where ASR errors concentrate in entity values and persist across turns, making it both a generation and an editing problem. A strong per-turn text editor corrects much of this but, operating on the transcript alone, leaves three recoverable errors: a value predicted inconsistently across turns, an omitted slot, and a value the audio does not support. We present AVERT, which scores each candidate value by combining cross-turn agreement with a trained audio-conditioned verifier and resolves the three error types with three operators, vote, add, and swap, each restricted to the slots where its error is common. On SpokenWOZ, a base speech-LLM reaches 33.04 JGA, a text editor 38.34, and AVERT 40.13, without retraining either. This is in the range of a 1B end-to-end system that consumes the full spoken history (39.32), though AVERT uses two 1B decoders rather than one. The audio verifier contributes a statistically significant gain, and restricting each operator to a selected slot subset matters: removing it lets unrestricted voting overwrite correct categorical values and fall below the editor.

中文摘要

摘要:口语对话状态跟踪从语音中恢复槽值对,其中自动语音识别(ASR)错误主要集中在实体值上,并在多轮对话中持续存在,使其既是生成问题也是编辑问题。一个强大的逐轮文本编辑器可以纠正大部分错误,但仅在转录文本上操作时,仍会留下三类可恢复的错误:跨轮次预测不一致的值、遗漏的槽位以及音频不支持的值。我们提出了AVERT,通过将跨轮次一致性与训练好的音频条件验证器相结合,对每个候选值进行评分,并使用三个操作——投票(vote)、添加(add)和交换(swap)——来解决这三类错误,每个操作仅限于其常见错误的槽位。在SpokenWOZ上,一个基础的语音大语言模型(speech-LLM)达到了33.04的JGA,一个文本编辑器为38.34,而AVERT达到了40.13,且无需重新训练任一模型。这与一个端到端、消耗完整语音历史的1B参数系统(39.32)的性能相当,尽管AVERT使用了两个1B参数解码器而不是一个。音频验证器带来了统计显著的增益,而将每个操作限制在选定槽位子集上也很重要:移除此限制会导致不受限制的投票覆盖正确的类别值,从而低于编辑器的表现。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01828 (HTTP 429)

Authors: Chunggi Lee, Hanspeter Pfister

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01828.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01828

Published: 2026-09-04T01:21:25.500Z


10. Interpretable Symptom Vectors for Depression in a Large Language Model

Abstract:Patients with depression present with diverse symptom profiles, yet clinical practice routinely reduces this variation to a single severity score. Large language models (LLMs) can potentially capture various symptoms and their severity from patient speech. However, how depressive symptoms are represented inside LLMs remains poorly understood, limiting clinical trust. To examine whether internal model activations match clinician judgment, we analyzed the residual stream of Gemma-3-27B-PT using mechanistic interpretability techniques. Recording activations across symptom descriptions drawn from validated clinical instruments, we found that symptom groups geometrically separated the most at layer 21 across multiple distance metrics. Using Semantic Projection, we then projected held-out naturalistic text onto Symptom Vectors constructed from these instruments. The resulting per-symptom coefficients preserved clinician-annotated rank ordering across mood, somatic, and suicidality axes. Furthermore, a single depression vector in Layer 21 separates held-out depressive from non-depressive text (AUC = 0.789), which can be used as an emotional valence gate that restricts symptom projection to depressive speech. These results reveal a decorrelated, clinician-aligned symptom signal readable directly from internal activations, offering a mechanistic foundation for interpretable depression-assessment tools.

中文摘要

摘要:抑郁症患者表现出多样的症状特征,但临床实践中通常将这种差异简化为单一的严重程度评分。大型语言模型(LLMs)有可能从患者的语言中捕捉各种症状及其严重程度。然而,抑郁症状在LLMs中的内部表示方式仍然了解不足,这限制了临床信任。为了检验内部模型激活是否与临床医生判断相匹配,我们使用机制可解释性技术分析了Gemma-3-27B-PT的残差流。通过记录来自经过验证的临床工具的症状描述的激活,我们发现,在第21层,症状组在多个距离度量上表现出最大的几何分离。随后,使用语义投影,我们将保留的自然语言文本投影到由这些工具构建的症状向量上。由此得到的每种症状系数在情绪、躯体和自杀倾向轴上保留了临床标注的排名顺序。此外,第21层的单一抑郁向量能够区分保留的抑郁与非抑郁文本(AUC = 0.789),可以用作情感效价门,将症状投影限制在抑郁语言上。这些结果揭示了一个去相关且与临床一致的症状信号,可以直接从内部激活中读取,为可解释的抑郁评估工具提供了机制基础。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01832 (HTTP 429)

Authors: Fangyi Zhu, Ajay Subramanian, Allison Constant, Camille Wang, Ravish Gupta, Corey J. Keller

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.01832.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01832

Published: 2026-09-04T01:21:25.500Z


Agent Domain Papers

1. EvalDetectBench: A Benchmark for Measuring Evaluation Awareness in Frontier Language Models

Abstract:Frontier large language models can often recognize when they are being evaluated, a capability known as evaluation awareness. If models behave differently in evaluations than in deployment, this undermines the validity of evaluation results, which are a crucial component of current AI safety frameworks. We introduce EvalDetectBench, an open pipeline and benchmark for measuring evaluation awareness that works with any Inspect-compatible evaluation, allowing practitioners to test against current and future benchmarks. EvalDetectBench ships with a newly curated transcript suite covering current frontier system-card evaluations and diverse deployment sources. The benchmark serves two purposes: measuring how reliably frontier LLMs recognize that they are being evaluated, and assessing how detectable individual benchmarks are as evaluations. We identify two methodological choices in the existing literature that introduce systematic bias: the identity of the model that generated the deployment transcripts accounts for 11.25% of measurement variance and can reorder model rankings; and elicitation prompts selected for high performance on one model can perform near chance on others. EvalDetectBench corrects for both via per-model probe calibration and a stratified generator-harmonisation procedure.

中文摘要

摘要:前沿大语言模型(LLM)通常可以识别自身是否正在被评估,这种能力被称为评估意识。如果模型在评估中的行为与部署中的行为不同,这将破坏评估结果的有效性,而评估结果是当前人工智能安全框架的重要组成部分。我们引入了 EvalDetectBench,一个开放的管线和基准,用于测量评估意识,可与任何 Inspect 兼容的评估一起使用,使实践者能够针对当前和未来的基准进行测试。EvalDetectBench 配备了新整理的一套记录,涵盖当前前沿系统卡片评估和多样化的部署来源。该基准有两个目的:一是衡量前沿 LLM 识别自己正在被评估的可靠程度,二是评估单个基准作为评估的可检测性。我们在现有文献中识别了两个引入系统性偏差的方法论选择:生成部署记录的模型身份占测量方差的 11.25%,并可能重新排列模型排名;为某一模型选择的高性能诱导提示在其他模型上可能接近偶然水平。EvalDetectBench 通过每模型探针校正和分层生成器协调程序纠正了这两点。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01611 (HTTP 429)

Authors: Xinning Li, Kemunto Ochwang’i, Aryasomayajula Ram Bharadwaj, Alexandra Souly, Robert Kirk

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01611.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01611

Published: 2026-09-04T01:21:44.694Z


2. Meta-ethics and AI: exploring the novel meta-ethical questions in the era of AI

Abstract:With the development of artificial intelligence (AI), the landscape of meta-ethics, which has largely centred on human ethics, faces pressures that may significantly reconfigure it. In particular, if future AI systems were to exhibit sufficiently integrated capacities for moral reasoning, moral intentionality, and moral reflection, novel meta-ethical questions would arise concerning what I call “AI’s own ethics”, as distinct from ethical principles merely imposed on AI by human designers. This paper offers a conditional and methodological framework for identifying the questions that would emerge if such AI systems were to arise. On that basis, the paper distinguishes four domains of meta-ethical inquiry in the era of AI: questions about the nature of human ethics from the human perspective; questions about the nature of AI’s own ethics from the human perspective; questions about the nature of human ethics from the AI perspective; and questions about the nature of AI’s own ethics from the AI perspective. The paper then considers how some existing mainstream meta-ethical theories (such as cognitivism and non-cognitivism, error theory and success theory, relativism, and objective realism) might illuminate these domains, while arguing that many familiar human-centred formulations of those theories may not transfer straightforwardly to AI cases without substantial revision. The overall conclusion is that the emergence of AI’s own ethics would place significant pressure on current frameworks and may require substantial refinement, reconstruction, or reconceptualisation.

中文摘要

摘要:随着人工智能(AI)的发展,一直以人类伦理为中心的元伦理学格局面临可能显著重构的压力。特别是,如果未来的人工智能系统展现出足够综合的道德推理、道德意向性和道德反思能力,就会出现我称之为“AI自身伦理”的新型元伦理问题,这与仅由人类设计者施加于AI的伦理原则不同。本文提出了一个有条件和方法论的框架,用于识别如果出现此类AI系统可能出现的问题。在此基础上,本文区分了人工智能时代的四个元伦理研究领域:从人类视角看人类伦理的本质问题;从人类视角看AI自身伦理的本质问题;从AI视角看人类伦理的本质问题;以及从AI视角看AI自身伦理的本质问题。随后,本文探讨了一些现有主流元伦理理论(如认知主义与非认知主义、错误理论与成功理论、相对主义与客观现实主义)可能如何阐释这些领域,同时论证了许多人类中心的理论表述在未经重大修正的情况下,可能无法直接迁移到AI情境中。总体结论是,AI自身伦理的出现将对现有框架施加显著压力,并可能需要进行大幅度的优化、重构或重新概念化。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01685 (HTTP 429)

Authors: Shang Lu

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01685.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01685

Published: 2026-09-04T01:21:44.694Z


Abstract:Statutes are increasingly parsed by machines before people read them, and the parsers disagree: on Missouri’s statutes, two independently written extractors diverge on numeric-threshold presence at a false-negative rate of 0.43. We ask what formal logic survives such noise. We build a passive survival certificate for the Duquenne-Guigues implication basis of machine-extracted statutory contexts: per-attribute inter-extractor disagreement is measured, replayed against the basis in 1,000 Monte Carlo trials, and an implication is certified only when a one-sided Wilson 95% lower bound on survival reaches 0.95; every certified implication carries premise spans and a minimal counterexample. On 29,365 Missouri sections and 502 Indian central-Act sections, the preregistered held-out gate passes (10 statute families across 7 Titles exact; 16 across 11 with 5% tolerance), yet under one globally deployed error model 93.2% of held-out chapters fall below the informativeness floor, and a 2x2 factorial assigns that to calibration-rate transfer, not selection. The certificate is usable but fragile: deploy it per-chapter-calibrated or error-tolerant. Code, data products, and the audit trail, including one retracted claim, are released.

中文摘要

摘要:在被人类阅读之前,法规条文越来越多地被机器解析,而解析器之间存在差异:在密苏里州的法规中,两个独立编写的提取器在数字阈值存在性上分歧,假阴性率为0.43。我们探讨在这种噪声下,哪些形式逻辑能够存活。我们为机器提取的法条上下文的Duquenne-Guigues蕴含基构建了一个被动存活证书:对每个属性的解析器间分歧进行测量,在1000次蒙特卡洛试验中针对该基重放,并且仅当生存的一侧Wilson 95%下限达到0.95时,才对该蕴含进行认证;每个经过认证的蕴含都包含前提跨度和最小反例。在29,365条密苏里州条文和502条印度中央法案条文中,预先注册的保留门检验通过(10个法规系列跨7个标题完全匹配;16个跨11个标题在5%容差内匹配),然而在一个全球部署的错误模型下,93.2%的保留章节低于信息量下限,且2x2因子分析表明这是由校准率转移而非选择造成的。该证书可用但脆弱:应按章节校准或容错部署。代码、数据产品及审计跟踪(包括一条已撤回的声明)已发布。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01741 (HTTP 429)

Authors: Surya Saka

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01741.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01741

Published: 2026-09-04T01:21:44.694Z


4. When Does Information Sharing Improve Decentralized Discovery? Aggregation, Independent Rescue, and Equilibrium Selection

Abstract:Information sharing can improve a pooled estimate while eliminating independent rescue actions. This paper separates those effects in exact finite discovery models. A centralized action-budget profile shows that equal one-person accuracy can coexist with different portfolio values. Under a registered incremental-sharing protocol, a sharing step improves discovery exactly when pooled residual error contracts faster than an independent rescue attempt. Exact bounded registries exhibit compression, aggregation, neutral curves, and a bounded zero mixed class. In a two-agent Bayesian game with a hidden mixture of common and independent signal sources, the registered selected equilibrium yields a strict positive sharing interval at signal accuracy 3/5, while alternative equilibria show that the result is selection-dependent rather than universal. The models are synthetic and finite; no human or organizational data are used.

中文摘要

摘要:信息共享可以在消除独立救援行动的同时改善汇总估计。本文在精确的有限发现模型中区分了这些效应。集中行动预算配置表明,个人准确率相同可以与不同的投资组合价值共存。在注册的增量共享协议下,只有当汇总的残差误差收缩速度快于独立救援尝试时,共享步骤才会改善发现。精确有界的注册机制表现出压缩、聚合、中性曲线和有界零混合类别。在一个具有隐藏的公共和独立信号来源混合的双代理贝叶斯博弈中,注册选择的均衡在信号准确率为3/5时产生严格正的共享区间,而其他均衡则表明结果依赖于选择,而非普遍适用。模型是合成的且有限;未使用任何人类或组织数据。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01814 (HTTP 429)

Authors: Yohei Nakajima

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01814.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01814

Published: 2026-09-04T01:21:44.694Z


5. Induction and Inquiry via Probabilistic Reasoning over Language and Code

Abstract:How humans grow and maintain abstract knowledge from the sparse, streaming noisy data of experience is a longstanding challenge in cognitive science. Any computational account must satisfy at least three desiderata: It must be (1) data-efficient and compute-efficient, (2) capture gradations of uncertainty to support intelligent inquiry and information gathering, and (3) be flexible enough to mentally represent the endless range of concepts people can learn and think about. Here we introduce a computational model that captures these three properties, by encoding symbolic knowledge as mental programs that combine natural language with source code, and sequentially inferring mental programs using LLM-guided Bayesian learning algorithms. Across a range of behavioral studies this model successfully reproduces quantitative signatures of human inductive learning and active inquiry, such as anchoring, garden-pathing, and other effects. In contrast, pure LLMs and classic Bayesian models either fail at the underlying task, or do not reproduce human behavior, or succeed only at exorbitant computational cost. These results suggest that one way humans continually grow their knowledge is by mentally representing many hypotheses spanning language-like and program-like representations, then revising those hypotheses to approximate Bayesian updates, while a bottom-up neural mechanism (an LLM) makes inference both tractable and learnable.

中文摘要

摘要:人类如何从稀疏、连续且噪声较大的经验数据中成长并维持抽象知识,是认知科学中的一个长期挑战。任何计算模型至少必须满足三个基本要求:它必须(1)数据高效且计算高效,(2)能够捕捉不确定性的层次,以支持智能的探究和信息收集,以及(3)足够灵活,能够在心理上表示人们可以学习和思考的无限概念范围。本文我们提出了一种计算模型,通过将符号知识编码为结合自然语言与源代码的心理程序,并使用大语言模型指导的贝叶斯学习算法顺序推断心理程序,从而捕捉这三种特性。在一系列行为研究中,该模型成功再现了人类归纳学习和主动探究的定量特征,例如锚定效应、路径依赖效应等。相比之下,纯大语言模型和经典贝叶斯模型要么在基础任务上失败,要么无法再现人类行为,或者仅以极高的计算成本取得成功。这些结果表明,人类持续增长知识的一种方式,是在心理上表示跨越类似语言和类似程序的多种假设,然后修正这些假设以近似贝叶斯更新,同时底层神经机制(即大语言模型)使得推断既可处理又可学习。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01815 (HTTP 429)

Authors: Wasu Top Piriyakulkij, Sam Acquaviva, Cassidy Langenfeld, Joshua Tenenbaum, Kevin Ellis

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01815.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01815

Published: 2026-09-04T01:21:44.694Z


6. Architecting Conversational Data Systems for Stateless LLM APIs: The Hydration Proxy Pattern

Abstract:As enterprise platforms transition to conversational reasoning interfaces, the stateless nature of LLM APIs creates an architectural gap. While statelessness enables horizontal scalability for AI providers, it forces client applications to manage the entire burden of conversational state and semantic memory. The work identifies the Hydration Proxy Pattern, an architecture that decouples session persistence from the reasoning engine. The framework ensures platform sovereignty over conversational data while enabling secure, multi-stage semantic grounding. We further propose the Context Stabilization Mandate to resolve the tradeoff between sovereign state management and KV caching.

中文摘要

摘要:随着企业平台向对话推理接口转型,大型语言模型(LLM)API 的无状态特性产生了架构上的差距。虽然无状态性使 AI 提供商能够实现水平扩展,但它迫使客户端应用程序承担整个对话状态和语义记忆的管理工作。本文提出了“Hydration Proxy 模式”,这是一种将会话持久性与推理引擎解耦的架构。该框架确保平台对对话数据拥有主权,同时实现安全的多阶段语义落地。我们进一步提出了“上下文稳定性要求”以解决主权状态管理与键值缓存之间的权衡问题。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01834 (HTTP 429)

Authors: Joseph Axisa

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01834.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01834

Published: 2026-09-04T01:21:44.694Z


7. SSAKG 2.0: An Open-Source Package for Structural Associative Sequence Memory and Context-Based Retrieval

Abstract:This article presents SSAKG 2.0, an open-source software package for constructing and operating Structural Sequential Associative Knowledge Graphs (SSAKGs). An SSAKG represents objects as graph vertices and ordered sequences as structural patterns of graph connections. The resulting sparse graph is used as an associative memory in which complete sequences can be reconstructed from a partial, unordered context. Version 2.0 introduces new algorithms that exploit individual bits of computer memory to efficiently search graph connections. The package is implemented in Python, while performance-critical graph operations are implemented in C and exposed through a Python interface. This hybrid implementation provides a flexible high-level programming environment while reducing the memory and computational overhead associated with large sparse graphs. The algorithms were evaluated using randomly generated numerical sequences, sequences derived from sentences in the NLTK corpus, and mRNA sequences. The experiments demonstrate the ability of the package to store and reconstruct sequences from partial contexts and provide a basis for evaluating the effects of graph density, sequence length, and memory size on retrieval performance. SSAKG 2.0 is distributed under the Apache 2.0 open-source license. The package includes documentation and reproducible examples and is publicly available through GitHub and the Python Package Index (PyPI).

中文摘要

摘要:本文介绍了 SSAKG 2.0,这是一个用于构建和操作结构化顺序关联知识图(SSAKGs)的开源软件包。SSAKG 将对象表示为图的顶点,并将有序序列表示为图连接的结构模式。由此生成的稀疏图被用作关联记忆,可以从部分无序的上下文中重建完整序列。2.0 版本引入了利用计算机内存单个位高效搜索图连接的新算法。该软件包用 Python 实现,而性能关键的图操作用 C 实现,并通过 Python 接口暴露出来。这种混合实现提供了灵活的高级编程环境,同时减少了与大型稀疏图相关的内存和计算开销。使用随机生成的数值序列、来自 NLTK 语料库句子的序列以及 mRNA 序列对算法进行了评估。实验表明,该软件包能够从部分上下文存储和重建序列,并为评估图密度、序列长度和内存大小对检索性能的影响提供了基础。SSAKG 2.0 在 Apache 2.0 开源许可协议下发布。该软件包包括文档和可重复的示例,并可通过 GitHub 和 Python 包索引(PyPI)公开获得。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01849 (HTTP 429)

Authors: Przemysław Stokłosa, Janusz A. Starzyk, Paweł Raif

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01849.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01849

Published: 2026-09-04T01:21:44.694Z


8. The Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory Agents

Abstract:Persistent memory supports personalized agents, but a stale stored fact can override current authoritative evidence without warning. We study when this harm begins as model capability changes. We evaluate a frozen, closed-set, action-scored benchmark with 2 suites that represent 2 different meanings of “no memory” (a Benefit suite, unsolvable without the stored fact, and a Safety suite, in which an authoritative tool always holds the correct value), on a same-family model-size series (Qwen3 0.6/1.7/4/8B). The Memory Trust Gap reflects over-trust rather than confusion. In the Benefit suite, models answer with the stale value 0.92-1.00 of the time at every scale. In the Safety suite, harm below the no-memory baseline under the trap conditions ($\Delta_{\mathrm{mem}}$) is capability-gated, with the larger models collapsing most once a stale note is made to look current. In a $2\times2\times2\times2$ factorial, which feature triggers over-trust depends on both the feature and model scale. Removing a label amplifies over-trust at every size, and a recency feature (stale dated newer) fools the larger models harder. Source authority is weak and scale-flat, and position changes from positive to negative across the Qwen3 model-size series. We confirm these scale interactions with direct cross-size contrast tests rather than overlapping per-model intervals. Mitigation is likewise capability-dependent: exposing metadata improves accuracy for the capable models, but only pre-resolving the conflict restores accuracy for the 2 smaller checkpoints. The same pattern appears on the capable models in an independent Llama-Instruct model-size series and on 2 external datasets (RGB, MisBench). A framing control finds no consistent advantage for the memory label: at the 3 smaller scales, models trust a stale document more than a stale memory; at 8B, the difference is not significant.

中文摘要

摘要:持久记忆支持个性化代理,但陈旧的存储事实可能无预警地覆盖当前权威证据。我们研究模型能力变化时何时开始伤害。我们评估一个冻结、封闭、动作评分基准测试,使用两个代表“无记忆”两种不同含义的套件(一个是没有存储事实无法解决的Benefit套件,另一个是权威工具始终保持正确值),适用于同一家族模型规模系列(Qwen3 0.6/1.7/4/8B)。记忆信任差距反映的是过度信任而非混淆。在Benefit套件中,模型在每个尺度上均以0.92-1.00时间的陈旧值回答。在安全套件中,陷阱条件($\Delta_{\mathrm{mem}}$)下低于无内存基线的伤害是能力门槛,大型模型一旦出现陈旧笔记以显得最新,便最为崩溃。在$2\times2\times2\times2$的阶乘中,哪些特征触发过度信任取决于特征和模型尺度。移除标签会在各种规模下放大过度信任,而较新的新近性特征则更能欺骗大型模型。来源权威性薄弱且尺度平坦,Qwen3模型大小系列中位置从正向变负。我们通过直接的跨尺度对比测试确认这些尺度交互,而非每个模型间隔重叠。缓解同样依赖能力:暴露元数据提升能模型的准确性,但只有预先解决冲突才能恢复两个较小检查点的准确性。同样的模式出现在独立的Llama-Instruct模型大小系列中具备能力的模型,以及两个外部数据集(RGB、MisBench)。框架控制显示内存标签没有一致优势:在三个较小尺度中,模型更信任陈旧文档而非陈旧内存;在8B尺度下,差异不显著。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01852 (HTTP 429)

Authors: Jundong Hu, Shekar Ramachandran

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01852.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01852

Published: 2026-09-04T01:21:44.694Z


9. Belief-Calibrated Optimization: An Explicit World Model for Agentic Optimization

Abstract:The performance of an LLM agent depends on the scaffold around a frozen model. A common way to improve that scaffold is to use a coding agent as an optimizer: it reads current scores and traces and iteratively edits the source, producing a new candidate each round. Each edit is chosen according to a belief about how the environment will respond: what went wrong, and which change should help. That belief is typically implicit. It lives in the coding agent’s reasoning on the current call, or remains latent in its parameters, rather than as something written down. Later calls therefore see scores and traces, but they do not use that belief. We introduce Belief-Calibrated Optimization (BCO), a method that writes that belief down as a persistent in-context document and continually revises that document as new candidates are evaluated. The resulting document is a world model: the current account of how the environment responds to edits. Added to an otherwise standard loop, BCO reaches a higher train passrate than a matched control that lacks only the world model, on five benchmarks spanning memory QA, tool-use QA, code-as-action app agents, and terminal agents. The gap remains on every held-out split, which is not used to select the candidate. After a target-model swap, in which the frozen model is replaced and the scaffold is not, the selected BCO scaffold leads on the tasks we test, except where context-window overruns leave it unfinished. An offline ablation then asks whether that gap comes from what the world model says. A fresh predictor given the accumulated document forecasts how the environment will respond more accurately than predictors given either no document or a same-form copy whose content has been falsified. The comparison indicates that the document carries reusable information in its content, not only in its form.

中文摘要

摘要:LLM(大型语言模型)代理的性能取决于围绕冻结模型的支架。改善这种支架的常见方法是使用编码代理作为优化器:它读取当前分数和跟踪信息,并迭代地编辑源代码,每轮生成一个新的候选版本。每次编辑的选择都基于对环境反应的信念:哪里出错了,以及哪种改变会有帮助。这种信念通常是隐性的。它存在于编码代理对当前调用的推理中,或者潜伏在其参数中,而不是以书面形式记录。后续调用因此会看到分数和跟踪信息,但不会使用该信念。我们提出了信念校准优化(Belief-Calibrated Optimization,BCO)方法,将这种信念写下作为持久的上下文内文档,并在评估新候选时不断修订该文档。最终得到的文档是一个世界模型:当前关于环境如何对编辑做出反应的描述。将其加入到其他标准循环中,BCO 在五个基准测试上(涵盖记忆问答、工具使用问答、以代码为行动的应用代理和终端代理)达到比仅缺少世界模型的对照组更高的训练通过率。在每个未使用来选择候选的保留分割上,这一差距仍然存在。在进行目标模型替换时,即冻结模型被替换但支架未改变,所选择的 BCO 支架在我们测试的任务中表现领先,除非上下文窗口超出限制导致任务未完成。接着进行的离线消融研究探讨了这一差距是否源于世界模型的内容。给定累积文档的新预测器,比不给文档或者给内容被篡改的同形式副本的预测器更准确地预测环境反应。比较表明,文档的内容携带了可重复使用的信息,而不仅仅是形式上的意义。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01861 (HTTP 429)

Authors: Yuhan Chen, Zhihua Tian, Mahavir Dabas, Charith Peris, Rahul Gupta, Ming Jin, Feiyang Kang, Siyuan Zhang, Nan Wang, Ruoxi Jia

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01861.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01861

Published: 2026-09-04T01:21:44.694Z


10. Epistemic Sybil Resistance: Multiplying AI Agents Without Multiplying Evidence

Abstract:Multi-agent AI systems improve inference by spawning agents and synthesizing reports. But another agent is not another observation: apparently independent reports may descend from the same evidence, and genuinely independent evidence can produce nearly identical reports. We formalize this as an epistemic Sybil problem. A report Z is an epistemic Sybil extension relative to reports R when I(Theta; Z | R) = 0. No report-only aggregator can generally distinguish replication from independent corroboration: identical reports can warrant different posteriors under unobserved ancestry. A Gaussian shared-root model shows common ancestry does not imply complete redundancy. Repeated extraction adds information toward a source-level ceiling, and correlated extraction errors, which a shared base model can induce among independent agents, lower that ceiling further. We test these predictions with more than 20,000 controlled LLM-agent report and extraction calls on synthetic evidentiary documents. Holding one evidence root fixed while report multiplicity rises from 1 to 32 collapses naive posterior coverage from 0.940 to 0.263. Holding report count fixed while evidence-root multiplicity rises from 1 to 16 closes the gap, and the aggregators are statistically indistinguishable at k = 16. The agent’s replicate extraction errors are correlated (gamma_cal = 0.719, estimated out of sample), and a correlated-extraction aggregator restores calibration accordingly. A controlled manipulation isolates representation similarity from evidential ancestry. It changes a report-space deduplication mechanism’s mean inferred cluster count by 1.425 (95% CI [1.363, 1.485]), whereas a fourfold change in true ancestry changes it by only 0.040 ([-0.045, 0.120]). Collective inference should therefore track evidential ancestry and dependence, not agent or report multiplicity or similarity.

中文摘要

摘要:多智能体 AI 系统通过生成代理并综合报告来提升推理能力。但另一个代理并不等于另一次观察:表面上独立的报告可能源自相同的证据,而真正独立的证据也可能生成几乎相同的报告。我们将此形式化为一个认知 Sybil 问题。当 I(Theta; Z | R) = 0 时,报告 Z 相对于报告 R 是认知 Sybil 扩展。仅依靠报告的聚合器通常无法区分复制与独立证实:在未观察到的源头下,相同的报告可能导致不同的后验结果。高斯共享根模型表明,共同的源头并不意味着完全冗余。重复提取增加了向源级上限的信息,而相关的提取错误——共享基础模型可能在独立代理之间引发的——进一步降低了该上限。我们通过对合成证据文档进行超过 20,000 次受控的 LLM-代理报告及提取调用来测试这些预测。当保持一个证据根固定而报告数量从 1 增加到 32 时,天真的后验覆盖率从 0.940 降至 0.263。当保持报告数量固定而证据根数量从 1 增加到 16 时,差距消失,在 k = 16 时,聚合器在统计上不可区分。代理的重复提取错误存在相关性(gamma_cal = 0.719,样本外估计),相关提取聚合器因此恢复了校准。受控操作将表示相似性与证据源分离开来。它改变了报告空间去重机制的平均推断集群数量为 1.425(95% 置信区间 [1.363, 1.485]),而真实源头的四倍变化仅改变 0.040([-0.045, 0.120])。因此,集体推理应追踪证据源和依赖关系,而非代理或报告的数量或相似性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01873 (HTTP 429)

Authors: Marc Bara

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01873.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01873

Published: 2026-09-04T01:21:44.694Z


Evaluation Domain Papers

1. EvalDetectBench: A Benchmark for Measuring Evaluation Awareness in Frontier Language Models

Abstract:Frontier large language models can often recognize when they are being evaluated, a capability known as evaluation awareness. If models behave differently in evaluations than in deployment, this undermines the validity of evaluation results, which are a crucial component of current AI safety frameworks. We introduce EvalDetectBench, an open pipeline and benchmark for measuring evaluation awareness that works with any Inspect-compatible evaluation, allowing practitioners to test against current and future benchmarks. EvalDetectBench ships with a newly curated transcript suite covering current frontier system-card evaluations and diverse deployment sources. The benchmark serves two purposes: measuring how reliably frontier LLMs recognize that they are being evaluated, and assessing how detectable individual benchmarks are as evaluations. We identify two methodological choices in the existing literature that introduce systematic bias: the identity of the model that generated the deployment transcripts accounts for 11.25% of measurement variance and can reorder model rankings; and elicitation prompts selected for high performance on one model can perform near chance on others. EvalDetectBench corrects for both via per-model probe calibration and a stratified generator-harmonisation procedure.

中文摘要

摘要:前沿大语言模型(LLM)通常可以识别自身是否正在被评估,这种能力被称为评估意识。如果模型在评估中的行为与部署中的行为不同,这将破坏评估结果的有效性,而评估结果是当前人工智能安全框架的重要组成部分。我们引入了 EvalDetectBench,一个开放的管线和基准,用于测量评估意识,可与任何 Inspect 兼容的评估一起使用,使实践者能够针对当前和未来的基准进行测试。EvalDetectBench 配备了新整理的一套记录,涵盖当前前沿系统卡片评估和多样化的部署来源。该基准有两个目的:一是衡量前沿 LLM 识别自己正在被评估的可靠程度,二是评估单个基准作为评估的可检测性。我们在现有文献中识别了两个引入系统性偏差的方法论选择:生成部署记录的模型身份占测量方差的 11.25%,并可能重新排列模型排名;为某一模型选择的高性能诱导提示在其他模型上可能接近偶然水平。EvalDetectBench 通过每模型探针校正和分层生成器协调程序纠正了这两点。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01611 (HTTP 429)

Authors: Xinning Li, Kemunto Ochwang’i, Aryasomayajula Ram Bharadwaj, Alexandra Souly, Robert Kirk

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01611.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01611

Published: 2026-09-04T01:22:06.146Z


2. Meta-ethics and AI: exploring the novel meta-ethical questions in the era of AI

Abstract:With the development of artificial intelligence (AI), the landscape of meta-ethics, which has largely centred on human ethics, faces pressures that may significantly reconfigure it. In particular, if future AI systems were to exhibit sufficiently integrated capacities for moral reasoning, moral intentionality, and moral reflection, novel meta-ethical questions would arise concerning what I call “AI’s own ethics”, as distinct from ethical principles merely imposed on AI by human designers. This paper offers a conditional and methodological framework for identifying the questions that would emerge if such AI systems were to arise. On that basis, the paper distinguishes four domains of meta-ethical inquiry in the era of AI: questions about the nature of human ethics from the human perspective; questions about the nature of AI’s own ethics from the human perspective; questions about the nature of human ethics from the AI perspective; and questions about the nature of AI’s own ethics from the AI perspective. The paper then considers how some existing mainstream meta-ethical theories (such as cognitivism and non-cognitivism, error theory and success theory, relativism, and objective realism) might illuminate these domains, while arguing that many familiar human-centred formulations of those theories may not transfer straightforwardly to AI cases without substantial revision. The overall conclusion is that the emergence of AI’s own ethics would place significant pressure on current frameworks and may require substantial refinement, reconstruction, or reconceptualisation.

中文摘要

摘要:随着人工智能(AI)的发展,一直以人类伦理为中心的元伦理学格局面临可能显著重构的压力。特别是,如果未来的人工智能系统展现出足够综合的道德推理、道德意向性和道德反思能力,就会出现我称之为“AI自身伦理”的新型元伦理问题,这与仅由人类设计者施加于AI的伦理原则不同。本文提出了一个有条件和方法论的框架,用于识别如果出现此类AI系统可能出现的问题。在此基础上,本文区分了人工智能时代的四个元伦理研究领域:从人类视角看人类伦理的本质问题;从人类视角看AI自身伦理的本质问题;从AI视角看人类伦理的本质问题;以及从AI视角看AI自身伦理的本质问题。随后,本文探讨了一些现有主流元伦理理论(如认知主义与非认知主义、错误理论与成功理论、相对主义与客观现实主义)可能如何阐释这些领域,同时论证了许多人类中心的理论表述在未经重大修正的情况下可能无法直接适用于AI案例。总体结论是,AI自身伦理的出现将对现有框架施加重大压力,可能需要进行实质性的精炼、重建或重新概念化。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01685 (HTTP 429)

Authors: Shang Lu

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01685.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01685

Published: 2026-09-04T01:22:06.146Z


Abstract:Statutes are increasingly parsed by machines before people read them, and the parsers disagree: on Missouri’s statutes, two independently written extractors diverge on numeric-threshold presence at a false-negative rate of 0.43. We ask what formal logic survives such noise. We build a passive survival certificate for the Duquenne-Guigues implication basis of machine-extracted statutory contexts: per-attribute inter-extractor disagreement is measured, replayed against the basis in 1,000 Monte Carlo trials, and an implication is certified only when a one-sided Wilson 95% lower bound on survival reaches 0.95; every certified implication carries premise spans and a minimal counterexample. On 29,365 Missouri sections and 502 Indian central-Act sections, the preregistered held-out gate passes (10 statute families across 7 Titles exact; 16 across 11 with 5% tolerance), yet under one globally deployed error model 93.2% of held-out chapters fall below the informativeness floor, and a 2x2 factorial assigns that to calibration-rate transfer, not selection. The certificate is usable but fragile: deploy it per-chapter-calibrated or error-tolerant. Code, data products, and the audit trail, including one retracted claim, are released.

中文摘要

摘要:在被人类阅读之前,法规条文越来越多地被机器解析,而解析器之间存在差异:在密苏里州的法规中,两个独立编写的提取器在数字阈值存在性上分歧,假阴性率为0.43。我们探讨在这种噪声下,哪些形式逻辑能够存活。我们为机器提取的法条上下文的Duquenne-Guigues蕴含基构建了一个被动存活证书:对每个属性的解析器间分歧进行测量,在1000次蒙特卡洛试验中针对该基重放,并且仅当生存的一侧Wilson 95%下限达到0.95时,才对该蕴含进行认证;每个经过认证的蕴含都包含前提跨度和最小反例。在29,365条密苏里州条文和502条印度中央法案条文中,预先注册的保留门检验通过(10个法规系列跨7个标题完全匹配;16个跨11个标题在5%容差内匹配),然而在一个全球部署的错误模型下,93.2%的保留章节低于信息量下限,且2x2因子分析表明这是由校准率转移而非选择造成的。该证书可用但脆弱:应按章节校准或容错部署。代码、数据产品及审计跟踪(包括一条已撤回的声明)已发布。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01741 (HTTP 429)

Authors: Surya Saka

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01741.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01741

Published: 2026-09-04T01:22:06.146Z


4. When Does Information Sharing Improve Decentralized Discovery? Aggregation, Independent Rescue, and Equilibrium Selection

Abstract:Information sharing can improve a pooled estimate while eliminating independent rescue actions. This paper separates those effects in exact finite discovery models. A centralized action-budget profile shows that equal one-person accuracy can coexist with different portfolio values. Under a registered incremental-sharing protocol, a sharing step improves discovery exactly when pooled residual error contracts faster than an independent rescue attempt. Exact bounded registries exhibit compression, aggregation, neutral curves, and a bounded zero mixed class. In a two-agent Bayesian game with a hidden mixture of common and independent signal sources, the registered selected equilibrium yields a strict positive sharing interval at signal accuracy 3/5, while alternative equilibria show that the result is selection-dependent rather than universal. The models are synthetic and finite; no human or organizational data are used.

中文摘要

摘要:信息共享可以在消除独立救援行动的同时改善汇总估计。本文在精确的有限发现模型中区分这些效应。集中式行动预算配置表明,个人准确率相同可以与不同的投资组合价值共存。在注册的增量共享协议下,只有当汇总的残差误差收缩速度快于独立救援尝试时,共享步骤才会改善发现。精确有界的注册机制表现出压缩、聚合、中性曲线和有界零混合类别。在一个具有隐藏的公共和独立信号来源混合的双代理贝叶斯博弈中,注册选择的均衡在信号准确率为3/5时产生严格正的共享区间,而其他均衡则表明结果依赖于选择,而非普遍适用。模型是合成的且有限;未使用任何人类或组织数据。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01814 (HTTP 429)

Authors: Yohei Nakajima

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01814.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01814

Published: 2026-09-04T01:22:06.146Z


5. Induction and Inquiry via Probabilistic Reasoning over Language and Code

Abstract:How humans grow and maintain abstract knowledge from the sparse, streaming noisy data of experience is a longstanding challenge in cognitive science. Any computational account must satisfy at least three desiderata: It must be (1) data-efficient and compute-efficient, (2) capture gradations of uncertainty to support intelligent inquiry and information gathering, and (3) be flexible enough to mentally represent the endless range of concepts people can learn and think about. Here we introduce a computational model that captures these three properties, by encoding symbolic knowledge as mental programs that combine natural language with source code, and sequentially inferring mental programs using LLM-guided Bayesian learning algorithms. Across a range of behavioral studies this model successfully reproduces quantitative signatures of human inductive learning and active inquiry, such as anchoring, garden-pathing, and other effects. In contrast, pure LLMs and classic Bayesian models either fail at the underlying task, or do not reproduce human behavior, or succeed only at exorbitant computational cost. These results suggest that one way humans continually grow their knowledge is by mentally representing many hypotheses spanning language-like and program-like representations, then revising those hypotheses to approximate Bayesian updates, while a bottom-up neural mechanism (an LLM) makes inference both tractable and learnable.

中文摘要

摘要:人类如何从稀疏、连续且噪声较大的经验数据中成长并维持抽象知识,是认知科学中的一个长期挑战。任何计算模型至少必须满足三个基本要求:它必须(1)数据高效且计算高效,(2)能够捕捉不确定性的层次,以支持智能的探究和信息收集,以及(3)足够灵活,能够在心理上表示人们可以学习和思考的无限概念范围。本文我们提出了一种计算模型,通过将符号知识编码为结合自然语言与源代码的心理程序,并使用大语言模型指导的贝叶斯学习算法顺序推断心理程序,从而捕捉这三种特性。在一系列行为研究中,该模型成功再现了人类归纳学习和主动探究的定量特征,例如锚定效应、路径依赖效应等。相比之下,纯大语言模型和经典贝叶斯模型要么在基础任务上失败,要么无法再现人类行为,或者仅以极高的计算成本取得成功。这些结果表明,人类持续增长知识的一种方式,是在心理上表示跨越类似语言和类似程序的多种假设,然后修正这些假设以近似贝叶斯更新,同时底层神经机制(即大语言模型)使得推断既可处理又可学习。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01815 (HTTP 429)

Authors: Wasu Top Piriyakulkij, Sam Acquaviva, Cassidy Langenfeld, Joshua Tenenbaum, Kevin Ellis

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01815.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01815

Published: 2026-09-04T01:22:06.146Z


6. Architecting Conversational Data Systems for Stateless LLM APIs: The Hydration Proxy Pattern

Abstract:As enterprise platforms transition to conversational reasoning interfaces, the stateless nature of LLM APIs creates an architectural gap. While statelessness enables horizontal scalability for AI providers, it forces client applications to manage the entire burden of conversational state and semantic memory. The work identifies the Hydration Proxy Pattern, an architecture that decouples session persistence from the reasoning engine. The framework ensures platform sovereignty over conversational data while enabling secure, multi-stage semantic grounding. We further propose the Context Stabilization Mandate to resolve the tradeoff between sovereign state management and KV caching.

中文摘要

摘要:随着企业平台向对话推理接口转型,大型语言模型(LLM)API 的无状态特性产生了架构上的差距。虽然无状态性使 AI 提供商能够实现水平扩展,但它迫使客户端应用程序承担整个对话状态和语义记忆的管理工作。本文提出了“Hydration Proxy 模式”,一种将会话持久化与推理引擎解耦的架构。该框架确保平台对对话数据的主权,同时实现安全的多阶段语义落地。我们进一步提出了“上下文稳定性要求”,以解决主权状态管理与键值缓存之间的权衡问题。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01834 (HTTP 429)

Authors: Joseph Axisa

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01834.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01834

Published: 2026-09-04T01:22:06.146Z


7. SSAKG 2.0: An Open-Source Package for Structural Associative Sequence Memory and Context-Based Retrieval

Abstract:This article presents SSAKG 2.0, an open-source software package for constructing and operating Structural Sequential Associative Knowledge Graphs (SSAKGs). An SSAKG represents objects as graph vertices and ordered sequences as structural patterns of graph connections. The resulting sparse graph is used as an associative memory in which complete sequences can be reconstructed from a partial, unordered context. Version 2.0 introduces new algorithms that exploit individual bits of computer memory to efficiently search graph connections. The package is implemented in Python, while performance-critical graph operations are implemented in C and exposed through a Python interface. This hybrid implementation provides a flexible high-level programming environment while reducing the memory and computational overhead associated with large sparse graphs. The algorithms were evaluated using randomly generated numerical sequences, sequences derived from sentences in the NLTK corpus, and mRNA sequences. The experiments demonstrate the ability of the package to store and reconstruct sequences from partial contexts and provide a basis for evaluating the effects of graph density, sequence length, and memory size on retrieval performance. SSAKG 2.0 is distributed under the Apache 2.0 open-source license. The package includes documentation and reproducible examples and is publicly available through GitHub and the Python Package Index (PyPI).

中文摘要

摘要:本文介绍了 SSAKG 2.0,这是一个用于构建和操作结构化顺序关联知识图(SSAKGs)的开源软件包。SSAKG 将对象表示为图的顶点,并将有序序列表示为图连接的结构模式。由此生成的稀疏图被用作关联记忆,可以从部分无序的上下文中重建完整序列。2.0 版本引入了利用计算机内存单个位高效搜索图连接的新算法。该软件包用 Python 实现,而性能关键的图操作用 C 实现,并通过 Python 接口暴露出来。这种混合实现提供了灵活的高级编程环境,同时减少了与大型稀疏图相关的内存和计算开销。使用随机生成的数值序列、来自 NLTK 语料库句子的序列以及 mRNA 序列对算法进行了评估。实验表明,该软件包能够从部分上下文存储和重建序列,并为评估图密度、序列长度和内存大小对检索性能的影响提供了基础。SSAKG 2.0 在 Apache 2.0 开源许可协议下发布。该软件包包括文档和可重复示例,并通过 GitHub 和 Python 包索引(PyPI)公开提供。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01849 (HTTP 429)

Authors: Przemysław Stokłosa, Janusz A. Starzyk, Paweł Raif

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01849.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01849

Published: 2026-09-04T01:22:06.146Z


8. The Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory Agents

Abstract:Persistent memory supports personalized agents, but a stale stored fact can override current authoritative evidence without warning. We study when this harm begins as model capability changes. We evaluate a frozen, closed-set, action-scored benchmark with 2 suites that represent 2 different meanings of “no memory” (a Benefit suite, unsolvable without the stored fact, and a Safety suite, in which an authoritative tool always holds the correct value), on a same-family model-size series (Qwen3 0.6/1.7/4/8B). The Memory Trust Gap reflects over-trust rather than confusion. In the Benefit suite, models answer with the stale value 0.92-1.00 of the time at every scale. In the Safety suite, harm below the no-memory baseline under the trap conditions ($\Delta_{\mathrm{mem}}$) is capability-gated, with the larger models collapsing most once a stale note is made to look current. In a $2\times2\times2\times2$ factorial, which feature triggers over-trust depends on both the feature and model scale. Removing a label amplifies over-trust at every size, and a recency feature (stale dated newer) fools the larger models harder. Source authority is weak and scale-flat, and position changes from positive to negative across the Qwen3 model-size series. We confirm these scale interactions with direct cross-size contrast tests rather than overlapping per-model intervals. Mitigation is likewise capability-dependent: exposing metadata improves accuracy for the capable models, but only pre-resolving the conflict restores accuracy for the 2 smaller checkpoints. The same pattern appears on the capable models in an independent Llama-Instruct model-size series and on 2 external datasets (RGB, MisBench). A framing control finds no consistent advantage for the memory label: at the 3 smaller scales, models trust a stale document more than a stale memory; at 8B, the difference is not significant.

中文摘要

摘要:持久记忆支持个性化代理,但陈旧的存储事实可能无预警地覆盖当前权威证据。我们研究模型能力变化时何时开始伤害。我们评估一个冻结、封闭、动作评分基准测试,使用两个代表“无记忆”两种不同含义的套件(一个是没有存储事实无法解决的Benefit套件,另一个是权威工具始终保持正确值),适用于同一家族模型规模系列(Qwen3 0.6/1.7/4/8B)。记忆信任差距反映的是过度信任而非混淆。在Benefit套件中,模型在每个尺度上均以0.92-1.00时间的陈旧值回答。在安全套件中,陷阱条件($\Delta_{\mathrm{mem}}$)下低于无内存基线的伤害是能力门槛,大型模型一旦出现陈旧笔记以显得最新,便最为崩溃。在$2\times2\times2\times2$的阶乘中,哪些特征触发过度信任取决于特征和模型尺度。移除标签会在各种规模下放大过度信任,而较新的新近性特征则更能欺骗大型模型。来源权威性薄弱且尺度平坦,Qwen3模型大小系列中位置从正向变负。我们通过直接的跨尺度对比测试确认这些尺度交互,而非每个模型间隔重叠。缓解同样依赖能力:暴露元数据提升能模型的准确性,但只有预先解决冲突才能恢复两个较小检查点的准确性。同样的模式出现在独立的Llama-Instruct模型大小系列中具备能力的模型,以及两个外部数据集(RGB、MisBench)。框架控制显示内存标签没有一致优势:在三个较小尺度中,模型更信任陈旧文档而非陈旧内存;在8B尺度下,差异不显著。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01852 (HTTP 429)

Authors: Jundong Hu, Shekar Ramachandran

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01852.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01852

Published: 2026-09-04T01:22:06.146Z


9. Belief-Calibrated Optimization: An Explicit World Model for Agentic Optimization

Abstract:The performance of an LLM agent depends on the scaffold around a frozen model. A common way to improve that scaffold is to use a coding agent as an optimizer: it reads current scores and traces and iteratively edits the source, producing a new candidate each round. Each edit is chosen according to a belief about how the environment will respond: what went wrong, and which change should help. That belief is typically implicit. It lives in the coding agent’s reasoning on the current call, or remains latent in its parameters, rather than as something written down. Later calls therefore see scores and traces, but they do not use that belief. We introduce Belief-Calibrated Optimization (BCO), a method that writes that belief down as a persistent in-context document and continually revises that document as new candidates are evaluated. The resulting document is a world model: the current account of how the environment responds to edits. Added to an otherwise standard loop, BCO reaches a higher train passrate than a matched control that lacks only the world model, on five benchmarks spanning memory QA, tool-use QA, code-as-action app agents, and terminal agents. The gap remains on every held-out split, which is not used to select the candidate. After a target-model swap, in which the frozen model is replaced and the scaffold is not, the selected BCO scaffold leads on the tasks we test, except where context-window overruns leave it unfinished. An offline ablation then asks whether that gap comes from what the world model says. A fresh predictor given the accumulated document forecasts how the environment will respond more accurately than predictors given either no document or a same-form copy whose content has been falsified. The comparison indicates that the document carries reusable information in its content, not only in its form.

中文摘要

摘要:LLM(大型语言模型)代理的性能取决于围绕冻结模型的支架。改善这种支架的常见方法是使用编码代理作为优化器:它读取当前分数和跟踪信息,并迭代地编辑源代码,每轮生成一个新的候选版本。每次编辑的选择都基于对环境反应的信念:哪里出错了,以及哪种改变会有帮助。这种信念通常是隐性的。它存在于编码代理对当前调用的推理中,或者潜伏在其参数中,而不是以书面形式记录。后续调用因此会看到分数和跟踪信息,但不会使用该信念。我们提出了信念校准优化(Belief-Calibrated Optimization,BCO)方法,将这种信念写下作为持久的上下文内文档,并在评估新候选时不断修订该文档。最终得到的文档是一个世界模型:当前关于环境如何对编辑做出反应的描述。将其加入到其他标准循环中,BCO 在五个基准测试上(涵盖记忆问答、工具使用问答、以代码为行动的应用代理和终端代理)达到比仅缺少世界模型的对照组更高的训练通过率。在每个未使用来选择候选的保留分割上,这一差距仍然存在。在进行目标模型替换时,即冻结模型被替换但支架未改变,所选择的 BCO 支架在我们测试的任务中表现领先,除非上下文窗口超出限制导致任务未完成。接着进行的离线消融研究探讨了这一差距是否源于世界模型的内容。给定累积文档的新预测器,比不给文档或者给内容被篡改的同形式副本的预测器更准确地预测环境反应。比较结果表明,该文档在其内容中携带了可重复使用的信息,而不仅仅在其形式上。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01861 (HTTP 429)

Authors: Yuhan Chen, Zhihua Tian, Mahavir Dabas, Charith Peris, Rahul Gupta, Ming Jin, Feiyang Kang, Siyuan Zhang, Nan Wang, Ruoxi Jia

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01861.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01861

Published: 2026-09-04T01:22:06.146Z


10. Epistemic Sybil Resistance: Multiplying AI Agents Without Multiplying Evidence

Abstract:Multi-agent AI systems improve inference by spawning agents and synthesizing reports. But another agent is not another observation: apparently independent reports may descend from the same evidence, and genuinely independent evidence can produce nearly identical reports. We formalize this as an epistemic Sybil problem. A report Z is an epistemic Sybil extension relative to reports R when I(Theta; Z | R) = 0. No report-only aggregator can generally distinguish replication from independent corroboration: identical reports can warrant different posteriors under unobserved ancestry. A Gaussian shared-root model shows common ancestry does not imply complete redundancy. Repeated extraction adds information toward a source-level ceiling, and correlated extraction errors, which a shared base model can induce among independent agents, lower that ceiling further. We test these predictions with more than 20,000 controlled LLM-agent report and extraction calls on synthetic evidentiary documents. Holding one evidence root fixed while report multiplicity rises from 1 to 32 collapses naive posterior coverage from 0.940 to 0.263. Holding report count fixed while evidence-root multiplicity rises from 1 to 16 closes the gap, and the aggregators are statistically indistinguishable at k = 16. The agent’s replicate extraction errors are correlated (gamma_cal = 0.719, estimated out of sample), and a correlated-extraction aggregator restores calibration accordingly. A controlled manipulation isolates representation similarity from evidential ancestry. It changes a report-space deduplication mechanism’s mean inferred cluster count by 1.425 (95% CI [1.363, 1.485]), whereas a fourfold change in true ancestry changes it by only 0.040 ([-0.045, 0.120]). Collective inference should therefore track evidential ancestry and dependence, not agent or report multiplicity or similarity.

中文摘要

摘要:多智能体 AI 系统通过生成代理并综合报告来提升推理能力。但另一个代理并不等于另一次观察:表面上独立的报告可能源自相同的证据,而真正独立的证据也可能生成几乎相同的报告。我们将此形式化为一个认知 Sybil 问题。当 I(Theta; Z | R) = 0 时,报告 Z 相对于报告 R 是认知 Sybil 扩展。仅依靠报告的聚合器通常无法区分复制与独立证实:在未观察到的祖先下,相同的报告可能导致不同的后验结果。高斯共享根模型表明,共同祖先并不意味着完全冗余。重复提取会向源级上限增加信息,而相关提取错误——共享基础模型可在独立代理之间引入——会进一步降低该上限。我们通过在合成证据文档上进行超过 20,000 次受控 LLM 代理报告和提取调用来验证这些预测。将一个证据根固定,而报告数量从 1 增加到 32 时,天真的后验覆盖率从 0.940 降至 0.263。将报告数固定,而证据根数量从 1 增加到 16 时,差距关闭,在 k = 16 时聚合器在统计上无法区分。代理的重复提取错误呈相关性(gamma_cal = 0.719,样本外估计),相关提取聚合器相应地恢复了校准。受控操纵隔离了表征相似性与证据祖先的影响。它改变了报告空间去重机制的平均推断集群数量为 1.425(95% 置信区间 [1.363, 1.485]),而真实祖先的四倍变化仅改变 0.040([-0.045, 0.120])。因此,集体推理应追踪证据祖先和依赖性,而非代理或报告的数量或相似性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01873 (HTTP 429)

Authors: Marc Bara

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.01873.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01873

Published: 2026-09-04T01:22:06.146Z


VLM Domain Papers

1. Beyond Textual Chain-of-Thought: A Survey on Action-Grounded Reasoning in Autonomous Driving

Abstract:Chain-of-thought (CoT) reasoning powers generative models by eliciting intermediate steps before producing an answer. In autonomous driving, the answer is a continuous action. Thus its reasoning must share the same spatiotemporal structure as the physical world. This survey studies the resulting shift from textual CoT to action-grounded reasoning. Surveying 171 papers, including 130 method papers and 41 benchmarks, datasets, surveys, and analysis papers, we propose a representation-centered taxonomy that treats the form of the intermediate state as the organizing axis. We systematize the 130 methods into four categories: language-based, visual-spatial, latent-dynamic, and externalized reasoning, further divided into 13 subtypes tied to distinct regions of interests. Our synthesis shows that the open frontier of reasoning in driving agents lies in intermediate representations that can be grounded in the real world, coupled to real-time action, and verified under safety-critical systems. Project page: this https URL.

中文摘要

摘要:链式思维 (Chain-of-thought, CoT) 推理通过在给出答案之前引导中间步骤来增强生成模型的能力。在自动驾驶中,答案是一个连续的动作。因此,其推理必须与物理世界具有相同的时空结构。本综述研究了从文本化 CoT 推理向以动作为基础的推理转变的趋势。通过对 171 篇论文的调查,包括 130 篇方法论文和 41 篇基准、数据集、综述及分析论文,我们提出了一种以表示为中心的分类法,将中间状态的形式作为组织轴心。我们将 130 种方法系统化为四类:基于语言的、视觉-空间的、潜在-动态的和外化推理,并进一步细分为与不同兴趣区相关的 13 个子类型。我们的综合分析表明,驾驶智能体推理的开放前沿在于能够与现实世界接地、与实时动作耦合,并在安全关键系统下验证的中间表示。项目页面:这个 https URL。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01659 (HTTP 429)

Authors: Zhengxu Tang, Xiaozhou Zhang, Guofeng Cui, Ziyu Gong, Zi Wang, Yunfei Shi, Ruifeng Deng, Chengzhi Qi, Ke Chen, Sachin Patil, Tianjun Xiao, Langechuan Liu, Pichao Wang

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01659.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01659

Published: 2026-09-04T01:22:30.627Z


2. FORGE: Forward-Only Test-Time Adaptation for Integer-Only Vision Models on Microcontrollers

Abstract:Vision models deployed on microcontrollers (MCUs) are quantized to integer-only arithmetic and run in inference-only runtimes that do not carry the machinery backpropagation needs: the standard tool for adapting a model to the distribution shift (sensor noise, blur, lighting) it meets in the field. Existing forward-only test-time adaptation (TTA) methods either run only on server- or edge-GPU-class models (not true microcontroller integer execution), or require the batch-normalization (BN) layers that integer deployment fuses away. We present a forward-only TTA method that operates on deployed, BN-folded, integer-only convolutional networks. The key observation is that fusing BN into the preceding convolution, a mandatory step for integer inference, destroys the statistics that normalization-based adaptation relies on. We restore adaptation by re-normalizing each folded convolution’s per-channel output to its clean training statistics, using only forward-pass estimates. The method (i) recovers most of gradient-based TENT’s accuracy gain (+20.9 vs. +24.9 points) and matches forward-only BN adaptation, while being the only method that runs on a folded integer-only model; (ii) needs to adapt only 3 of 21 layers (selected without seeing the test corruptions) to recover 93% of the benefit; (iii) survives single-sample streaming with a batch-size-scaled momentum; and (iv) generalizes across three datasets (up to 200 classes) and two architectures. We validate bit-exact int8 convolution execution and deploy on an ESP32-S3, where, measured with a Nordic PPK2 power profiler, the forward-only adaptation (a lightweight fp32 recalibration around the int8 convolutions) costs only 8.3 mJ (6.8% of inference energy) and 21.9 ms on the deployed SIMD-optimized model: forward-only adaptation is cheap on a real microcontroller.

中文摘要

摘要:部署在微控制器(MCU)上的视觉模型都量化为仅整数运算,并在仅推理的运行时环境中运行,这些环境不具备反向传播所需的机制:反向传播是将模型适应其在实际环境中遇到的分布偏移(传感器噪声、模糊、光照)的标准工具。现有的仅前向的测试时自适应(TTA)方法要么只在服务器或边缘 GPU 类模型上运行(不是真正的微控制器整数执行),要么依赖于整数部署中被融合掉的批量归一化(BN)层。我们提出了一种仅前向 TTA 方法,可在已部署、BN 融合且仅整数的卷积网络上运行。关键观察是,将 BN 融合到前置卷积中——对于整数推理是必须的步骤——会破坏归一化自适应所依赖的统计信息。我们通过使用仅前向估计,将每个折叠卷积的每通道输出重新归一化到其干净训练统计,从而恢复自适应。该方法(i)恢复了大部分基于梯度的 TENT 精度提升(+20.9 对比 +24.9 点)并匹配仅前向 BN 自适应,同时是唯一可在折叠整数模型上运行的方法;(ii)仅需适应 21 层中的 3 层(在未看到测试污染的情况下选择)即可恢复 93% 的收益;(iii)通过批量大小缩放的动量可在单样本流中生存;以及(iv)在三个数据集(最多 200 类)和两种架构上泛化。我们验证了逐位完全的 int8 卷积执行,并部署在 ESP32-S3 上,在使用 Nordic PPK2 功耗分析仪测量时,仅前向自适应(围绕 int8 卷积的轻量级 fp32 重新校准)仅消耗 8.3 mJ(推理能量的 6.8%)并在部署的 SIMD 优化模型上耗时 21.9 ms:在真实微控制器上,仅前向自适应代价低廉。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01683 (HTTP 429)

Authors: Muhammad Rehan, Haider Ali, Muhammad Ali Munir, Moaz Amjad

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01683.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01683

Published: 2026-09-04T01:22:30.627Z


3. FairLens: Benchmarking Fairness in Vision-Language Models for High-Stakes Decision-Making

Abstract:Vision-language models (VLMs) are increasingly used to make decisions from visual inputs. We introduce FAIRLENS, a benchmark and evaluation framework for measuring both the fairness and the validity of VLM responses in three high-stakes domains: hiring, legal, and healthcare. FAIRLENS pairs real face images spanning gender, race, and age groups with closed- and open-ended questions, giving more than 100K image-question pairs per model, and evaluates responses from four complementary views: demographic parity over adverse outcome rates, soundness, demographic association over unsupported roles and statuses, and bias in free-text generation. Soundness is the central validity criterion: a response is sound when it follows the evidence stated in the question and abstains when the image cannot support an answer. Evaluating eight VLMs, we find that the primary failure is unwarranted inference rather than unequal treatment. Models routinely infer qualifications, threat, illness, or professional role from a face instead of abstaining, and the weakest model does so on 99% of the questions its input cannot answer. These failures are most severe in legal and healthcare, where recognizing insufficient evidence matters most, and disparity metrics alone would miss them: parity gaps are small in absolute terms, yet when baseline adverse rates are low the same gap means one demographic group receives adverse labels several times as often as another, and a small gap can equally reflect a model that treats every group unsafely. Bias in free-text responses is only loosely coupled to multiple-choice accuracy, so correct structured answers do not imply safe generation. FAIRLENS shows that fair high-stakes VLM behavior requires similar treatment across groups and refusal to infer high-stakes attributes from appearance, and its question suite transfers to any face corpus with demographic annotations.

中文摘要

摘要:视觉-语言模型(VLMs)越来越多地被用于从视觉输入中做出决策。我们介绍了 FAIRLENS,这是一个用于衡量 VLM 在三个高风险领域(招聘、法律和医疗)中响应的公平性和有效性的基准和评估框架。FAIRLENS 将涵盖性别、种族和年龄群体的真实面部图像与封闭式和开放式问题配对,为每个模型提供超过 10 万张图像-问题对,并从四个互补的视角评估响应:不良结果率的人口统计平等性、合理性、对不支持角色和身份的人口统计关联性以及自由文本生成中的偏差。合理性是核心有效性标准:当响应遵循问题中陈述的证据并在图像无法支持答案时选择放弃时,响应是合理的。在对八个 VLM 进行评估时,我们发现主要失败原因是无根据的推测,而非不平等对待。模型经常从面部推断资格、威胁、疾病或职业角色,而不是选择放弃,而最弱的模型在其输入无法回答的问题中有 99% 出现这种情况。这些失败在法律和医疗领域最为严重,因为识别证据不足至关重要,而仅靠差异指标无法发现:绝对值上的平等差距很小,但当基线不良率较低时,同样的差距意味着一个群体比另一个群体更可能多次收到不利标签,且小差距同样可能反映模型对每个群体的处理都不安全。自由文本响应中的偏差与选择题准确性仅有松散关联,因此正确的结构化答案并不意味着安全生成。FAIRLENS 显示,要实现高风险情境下的公平 VLM 行为,需要在各群体间进行类似处理,并拒绝根据外表推断高风险属性,其问题套件也可以转用于任何带有人口统计注释的面部数据集。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01691 (HTTP 429)

Authors: Vahid Reza Khazaie, Ahmed Y. Radwan, Shaina Raza

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01691.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01691

Published: 2026-09-04T01:22:30.627Z


4. From Visual Cues to Spoken Narration: Rethinking Audio Description

Abstract:Audio Description (AD) provides spoken narration of visual events during dialogue gaps, making movies accessible to visually impaired audiences. The problem requires determining both what (which visual event) and when (position for inserting the AD) to narrate, to achieve the best user experience. Prior work has largely reduced the problem to video captioning of pre-segmented video clips, i.e., what is largely predefined and when is ignored entirely. We propose Cue2Narrate, a two-stage pipeline that jointly predicts what and when to narrate in longer untrimmed movie clips. A dual-head audio-visual localizer predicts two temporally distinct windows per AD utterance: a visual cue window and a spoken narration window. A LoRA-adapted VLM then generates concise ADs from the predicted visual evidence, trained with a Description Ranking Loss that ranks captions (negative samples) of the same frames lower than the GT AD. To benchmark this new problem statement, we introduce the LongLSMDC benchmark with up to 8-min movie clips (~6.5min on average). On LongLSMDC, Cue2Narrate outperforms video-only and audio-only localization baselines by 5—12 points in avg. mAP. Under both predicted- and GT-window evaluation, Cue2Narrate improves AD generation over the corresponding fine-tuned base VLM. These results establish the first benchmark for multi-segment AD generation on long-form clips. Data & Code: this https URL

中文摘要

摘要:音频描述(AD)在对话间隙提供视觉事件的口述叙述,使电影对视力受限的观众可访问。该问题需要确定讲述的内容(哪个视觉事件)和时间(插入AD的位置),以实现最佳用户体验。先前的工作主要将该问题简化为对预先分割的视频片段进行视频字幕生成,即内容大多是预定义的,而时间则被完全忽略。我们提出了Cue2Narrate,一个两阶段流程,可在较长的未剪辑电影片段中联合预测何时以及讲述什么。一个双头音视频定位器为每个AD语句预测两个时间上不同的窗口:视觉提示窗口和口述叙述窗口。随后,采用LoRA适配的VLM从预测的视觉证据中生成简明的AD,并通过描述排序损失进行训练,使得同一帧的其他字幕(负样本)排名低于GT AD。为了评估这一新的问题陈述,我们引入了LongLSMDC基准数据集,其中电影片段最长可达8分钟(平均约6.5分钟)。在LongLSMDC上,Cue2Narrate在平均mAP指标上比仅使用视频或音频的定位基线高出5-12个点。在预测窗口和GT窗口的评估下,Cue2Narrate均改善了相应微调基础VLM的AD生成效果。这些结果建立了长篇片段多段AD生成的首个基准。数据及代码: 此https URL

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01725 (HTTP 429)

Authors: Akshita Gupta, Aditya Arora, Federico Tombari, Marcus Rohrbach, Anna Rohrbach

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01725.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01725

Published: 2026-09-04T01:22:30.627Z


5. UAV Thermal Imagery for Inert Ordnance Screening: Multi Campaign Dataset Development,Object Detection, and Practical Recommendations

Abstract:Unexploded ordnance (UXO) continues to restrict civilian access, agricultural activity, infrastructure recovery, and environmental remediation in contaminated areas around the world. This study created a multi campaign UAV thermal image data set of inert ordnance, developed a labeled image set from collected imagery, tested object detection models, and identified practical considerations for humanitarian mine action and demining applications. Data were collected during four field campaigns in Tennessee under summer and winter conditions using inert mines, munitions, and other ordnance placed in short grass, tall vegetation, gravel, mulch, rock, compost, and compacted surfaces. Thermal imagery was collected under flight altitutes of 33 m and 15 m. The final source inventory contained 5,855 thermal image label pairs, including 918 positive images and 4,937 background images. After retaining all positive images and downsampling background images, the 33 m dataset contained 420 training and 106 validation images, while the 15 m dataset contained 629 training and 157 validation images. YOLOV11l and RT-DETR-R50 algorithms were trained and evaluated to develop an automated candidate detection model. Practical recommendations include collecting thermal and RGB imagery together, incorporating varied surfaces and background only imagery, considering periods following changes in solar exposure, balancing survey coverage against target pixel representation, calibrating models with representative local data, and retaining qualified human review. The intended use is screening and prioritization for follow on technical survey or EOD assessment, and not a standalone clearance.

中文摘要

摘要:未爆炸弹药(UXO)持续限制着全球污染地区的民用通行、农业活动、基础设施恢复和环境治理。本研究创建了一个多次航次的无人机热成像惰性弹药数据集,从收集的图像中开发了标注图像集,测试了目标检测模型,并识别了人道主义排雷和排雷应用的实际考虑因素。数据收集在田纳西州进行了四次实地考察,涵盖夏季和冬季条件,使用了放置在短草、高植被、碎石、覆盖物、岩石、堆肥和压实地面上的惰性地雷、弹药及其他弹药品类。热成像图像在飞行高度 33 米和 15 米下收集。最终数据源库存包含 5,855 对热成像标签对,其中包括 918 张正样本图像和 4,937 张背景图像。在保留所有正样本图像并下采样背景图像后,33 米数据集包含 420 张训练图像和 106 张验证图像,而 15 米数据集包含 629 张训练图像和 157 张验证图像。YOLOV11l 和 RT-DETR-R50 算法进行了训练和评估,以开发自动候选检测模型。实际建议包括同时收集热成像和 RGB 图像、纳入不同表面和仅背景图像,考虑太阳照射变化后的时段,平衡调查覆盖范围与目标像素表示,使用具有代表性的本地数据校准模型,并保留合格的人类复核。其预期用途是为后续技术调查或排爆评估进行筛选和优先排序,而非独立进行清理。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01738 (HTTP 429)

Authors: Chad Melton, PhD., Annabelle Kelton

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01738.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01738

Published: 2026-09-04T01:22:30.627Z


6. ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes

Abstract:Compact token sequences are essential for efficient 3D generation. However, existing 3D tokenizers typically organize latent representations either over spatial regions or as fixed-size sets of global tokens, both suffering sharp reconstruction degradation when compressed to extremely low token budgets. In this paper, we present ZipTok3D, a 3D tokenizer designed for high-fidelity reconstruction from extremely short token sequences. Its key idea is to organize object geometry into progressively informative global-token prefixes and unfold these compact representations through iterative decoding. Specifically, nested dropout randomly truncates the latent sequence after encoding during training and requires each retained prefix to reconstruct the complete object, thereby prioritizing essential geometric information in the leading tokens. The decoder then repeatedly applies a parameter-shared Transformer block to recover fine-grained geometry from each prefix without a separate generative sampling stage. With the same token dimension, ZipTok3D achieves reconstruction quality comparable to the 32-token COD-VAE baseline using only one token on ShapeNet and four on TRELLIS, yielding $32\times$ and $8\times$ shorter token sequences, respectively.

中文摘要

摘要:紧凑的令牌序列对于高效的3D生成至关重要。然而,现有的3D令牌器通常将潜在表示组织在空间区域上,或作为固定大小的全局令牌集,在压缩到极低的令牌预算时,两者都容易出现明显的重建退化。在本文中,我们提出了ZipTok3D,一种旨在从极短令牌序列中实现高保真重建的3D令牌器。其关键思想是将物体几何组织为逐步提供信息的全局令牌前缀,并通过迭代解码展开这些紧凑表示。具体而言,嵌套丢弃在训练期间编码后随机截断潜在序列,并要求每个保留的前缀重建完整物体,从而优先在前置令牌中保留关键的几何信息。然后,解码器重复应用参数共享的Transformer模块,从每个前缀中恢复细粒度几何,而无需单独的生成采样阶段。在相同令牌维度下,ZipTok3D在ShapeNet上使用单个令牌、在TRELLIS上使用四个令牌即可实现与32令牌COD-VAE基线相当的重建质量,分别实现$32 imes$和$8 imes$的令牌序列缩短。

LLM Analysis

Q: 这篇论文试图解决什么问题? 这篇论文旨在解决\*\*从极短token序列(如1–4个token)中实现高保真3D物体重建\*\*的问题,核心动机与具体挑战可概括如下: - \*\*3D tokenizer的压缩–保真权衡困境\*\* 现有3D tokenizer主要分为两类,但均在极低token预算下失效: - \*\*空间结构化方法\*\*(如sparse voxel、octree)通过局部锚点保留空间结构,减少序列长度通常需要更粗或更稀疏的空间支撑,导致重建质量下降。 - \*\*全局token方法\*\*(如VecSet、COD-VAE)将整体几何压缩为固定数量的全局token,虽然序列更短,但当token预算缩减至极少量时,重建质量会急剧退化。 - \*\*极低token预算下的双重瓶颈\*\* 当整个3D对象必须被编码到极少数几个token时,现有方法面临两个未被分离的困难: 1. \*\*编码侧的信息竞争\*\*:全局结构与细粒度细节必须在严重受限的潜在容量中竞争,而固定预算优化目标并未显式要求“哪些信息应被优先保留”,导致前缀token不足以支撑完整重建。 2. \*\*解码侧的计算瓶颈\*\*:传统固定深度的单前馈解码器难以将极少数全局向量充分展开为具有空间细节的3D场,需要更多计算来逐步展开紧凑的潜在表示。 - \*\*本文的解决目标\*\* 为此,论文提出\*\*ZipTok3D\*\*,通过联合学习\*\*渐进式信息前缀\*\*(progressively informative global-token prefixes)与\*\*参数共享的迭代解码\*\*(parameter-shared iterative refinement),在编码阶段将关键几何信息集中于序列前缀,在解码阶段通过重复应用共享Transformer块逐步细化空间表示,从而在无需额外生成采样阶段的情况下,从极短token序列中恢复高保真3D几何。 Q2: 有哪些相关研究? 根据论文第2节(Related Work),相关研究主要分布在以下三个方向: ## 1. 3D 潜在表示(3D Latent Representations) ### 空间结构化方法 这类方法将潜在特征与点、网格或层次结构关联,但压缩通常需要更粗或更稀疏的空间支撑: - \*\*LION\*\*:采用层次化点云特征进行3D形状生成。 - \*\*3DILG\*\*:使用非规则网格进行3D生成建模。 - \*\*OctFusion\*\*:基于八叉树的3D形状扩散模型。 - \*\*XCube\*\*:构建稀疏体素层次结构进行大规模3D生成。 - \*\*TRELLIS\*\*:将特征附加到占用的稀疏网格单元。 - \*\*LATTICE / O-Voxel\*\*:分别通过粗体素锚点或稀疏几何–外观潜变量进一步压缩空间表示。 ### 全局Token方法 将对象级几何压缩为紧凑的token集合,但在极少token预算下重建质量会急剧下降: - \*\*VecSet\*\*:通过交叉注意力获得神经场潜变量。 - \*\*COD-VAE\*\*:逐步压缩点特征用于三平面(triplane)重建。 - \*\*SceneTok\*\*:将置换不变token集扩展到多视图场景建模。 - \*\*LaGeM\*\*:使用token层次结构。 - \*\*多尺度残差量化与几何感知采样\*\*:通过紧凑代码支持多样化的解码器和生成模型。 ### 基于紧凑代码的生成模型 - \*\*Shape Tokens\*\*:用量化形状代码条件化流匹配表面场。 - \*\*Kyvo\*\*:用于多模态自回归场景建模的量化形状代码。 - \*\*FlashVDM\*\*:通过扩散蒸馏与高效隐式解码加速VecSet生成。 - \*\*Block3D\*\*:研究块级扩散以实现高效的文本到3D生成。 ## 2. 灵活长度Token化(Flexible-Length Tokenization) ### 通用灵活长度方法 - \*\*Nested Dropout\*\*:训练时随机截断潜在序列,诱导信息排序,要求每个保留前缀可解码。 - \*\*图像与视频Tokenizer\*\*:学习前缀可解码或因果一维表示(如FlexTok、基于主成分的图像表示、Catok);自适应方法根据输入复杂度分配token预算(如Adaptive Length Image Tokenization、ElasticTok、AdapTok、EvaTok)。 - \*\*VideoFlexTok\*\*:学习粗到细视频前缀,配合生成流解码器。 - \*\*ReTok\*\*:改进Nested Dropout表示中后续token的利用。 ### 3D领域灵活长度方法 - \*\*OAT\*\*:根据形状复杂度自适应分配八叉树token。 - \*\*SuperVoxelGPT\*\*:在固定生成顺序下自适应调整超体素大小以匹配局部细节。 - \*\*LoST\*\*:按语义显著性排序token,短前缀指定整体形状,长前缀添加实例细节;但其短前缀条件化基于扩散的补全(completion),而非直接精确重建编码实例。 ## 3. 迭代细化(Iterative Refinement) ### 共享参数的通用循环网络 - \*\*Universal Transformer\*\*:通过循环应用共享的自注意力与前馈层,解决标准Transformer固定深度的问题。 - \*\*ELT\*\*:将权重共享的Transformer循环与循环内自蒸馏扩展到图像和视频生成。 ### 3D几何的迭代恢复 - \*\*Cascaded Refinement Network\*\*:通过级联的粗到精细化网络,恢复粗预测中缺失的几何细节。 - \*\*RFNet\*\*:在循环层级共享操作,逐步增加点密度并保留观测细节,降低密集点云补全的参数与内存开销。 --- 上述工作为ZipTok3D提供了基础,但现有方法未能同时解决\*\*极少token前缀中的信息组织\*\*与\*\*紧凑潜在表示的高效空间展开\*\*这两个耦合问题。 Q3: 论文如何解决这个问题? 论文通过提出 \*\*ZipTok3D\*\*,从编码与解码两个层面联合解决极低 token 预算下的高保真 3D 重建问题。核心思路是将对象几何组织为\*\*渐进式信息前缀\*\*(progressively informative prefixes),并通过\*\*参数共享的迭代细化\*\*(parameter-shared iterative refinement)将这些紧凑表示逐步展开为精细的空间场。具体方法可概括为以下四个部分: --- ### 1. 编码端:嵌套前缀训练(Nested Prefix Training) 现有 tokenizer 通常仅在完整潜在序列上优化重建目标,导致信息分散在整个序列中,前几枚 token 不足以独立重建完整对象。ZipTok3D 则显式要求\*\*每个前缀都能重建完整对象\*\*。 - \*\*潜在序列生成\*\* 编码器 E_φ 将输入表面点云 X = x_i_(i=1)^N ⊂ R^3 映射为最多 M=128 枚潜在向量:

Z = E_φ(X) = (z_1, …, z_M) ∈ R^(M × d)

  • **随机前缀采样** 训练时对每个样本,从指数间隔的 token 预算集合 K = 1, 2, 4, 8, 16, 32, 64, 128 中均匀采样一个保留长度 K :
    K sim Unif(K)
  • **掩码与重建约束** 仅保留前缀 Z(:K) = (z_1, …, z_K) ,后缀 Z(K+1:M) 被完全掩码。重建目标必须**仅通过该前缀实现**。这强制序列前端集中保留对象级几何,后续 token 则编码残差细节。由于前缀的嵌套结构(短前缀包含于长前缀中),同一 checkpoint 可在推理时通过截断直接支持任意训练过的 token 预算。 —- ### 2. 解码端:全局到空间的迭代细化(Iterative Global-to-Spatial Refinement) 短前缀将对象级信息压缩至极少数全局向量,传统固定深度的单前馈解码器难以将其一次性展开为高密度空间场。ZipTok3D 以**循环复用**的方式增加有效解码深度,而不引入逐层新参数。 - **Triplane 初始化与选择** 可学习的 triplane token T^0 首先通过选择块 Sω 与保留前缀 Z(:K) 交互,被划分为**选中状态** H^0_K (用于后续细化)和**冗余 token** R_K (被 bypass):

(H^0K, R_K) = Sω(T^0, Z_(:K))

  • **参数共享的 Transformer 循环** 同一六层 Transformer 块 gθ (参数为 θ )被重复应用 L 次。第 ell 步的更新仅依赖于上一步状态与固定的前缀:
    H^ell_K = g
    θ(H^(ell-1)K, Z(:K)), quad ell = 1, …, L
    前缀 Z(:K) 与网络参数在迭代中保持不变,仅 triplane 状态被更新。 - **恢复与查询** 每一步将细化后的状态与冗余 token 恢复为完整 triplane T^ell_K = R(H^ell_K, R_K) ,再通过共享的占用 MLP h(occ) 在查询点 qj 处预测占用概率:
    y^ell
    (K,j) = σ(h(occ)(T^ell_K, q_j))
    该过程无需额外的生成采样阶段,直接从紧凑前缀恢复几何,且 L 可在推理时作为计算量的独立控制变量。 —- ### 3. 训练目标:最终监督与中间自蒸馏 仅监督最终迭代输出会导致中间状态缺乏有效约束。ZipTok3D 对**最终输出**与**随机采样的中间输出**同时监督,并引入自蒸馏以稳定优化。 - **重建损失** 对第 ell 步预测,使用体查询 Q
    (vol) 与近表面查询 Q_(near) 的加权二元交叉熵:

L(rec)(Y^ell_K, Y) = (1) / (N(textvol))∑(j ∈ I)(vol) BCE(y^ell(K,j), y_j) + λ(near)N(near)∑(j ∈ I)(near) BCE(y^ell(K,j), yj)
其中 λ
(near) = 0.1 。 - **中间自蒸馏** 随机采样一个中间步 ell ∈ 1, …, L-1 ,要求其预测不仅匹配真实标签,还与**分离梯度(stop-gradient)的最终预测**对齐:
L(dist)(Y^ell_K, sg(Y^L_K)) = (1) / (N(textvol))∑(j ∈ I)(vol) BCE(y^ell(K,j), sg(y^L(K,j))) + λ(near)N(near)∑(j ∈ I)(near) BCE(y^ell(K,j), sg(y^L(K,j)))

  • **完整目标** 综合最终重建、辅助损失(来自 COD-VAE 的初始化损失 L(∈it) 与不确定性损失 L(unc) )以及中间监督:
    L(K,ell) = L(rec)(Y^LK, Y) + L(aux) + β[αt L(rec)(Y^ellK, Y) + (1-α_t)L(dist)(Y^ellK, sg(Y^L_K))]
    其中 β = 0.5 , α_t 随训练步数从 1 衰减至 0,实现从重建监督到蒸馏监督的渐进过渡。 —- ### 4. 推理时的解耦控制 ZipTok3D 将传统 autoencoder 中隐含的“压缩–计算”耦合显式解耦为两个独立控制变量: | 控制变量 | 作用 | 推理时调节方式 | |————-|———|———————| | **前缀长度 K ** | 决定潜在序列携带的信息量 | 直接截断至 K 枚 token,无后缀填充 | | **细化深度 L ** | 决定解码器展开该表示所用的计算量 | 循环调用共享 Transformer 块 L 次 | 这种解耦使得:在 K 极小时,可通过增加 L 充分展开前缀中的紧凑几何;而在 K 足够大时,即使 L=1 也能获得较强重建。实验表明,当 L=5 时,1 枚 token 在 ShapeNet 上即可接近 32-token COD-VAE 的性能;4 枚 token 在 TRELLIS 上不仅达到、且在表面指标上超过 COD-VAE-32。 Q4: 论文做了哪些实验? 论文在 **ShapeNet-v2** 与 **TRELLIS-500K** 上开展了系统性的定量和定性实验,涵盖重建、生成、消融、效率与自适应预算诊断等多个方面。具体实验内容如下: —- ### 1. 主要重建实验(Reconstruction from Extremely Short Prefixes) 在固定评估协议下,将 ZipTok3D 与 3DILG、VecSet(32/64/512 token)及 COD-VAE(2/32/64 token)进行对比: - **数据集**:ShapeNet-v2(55 类,1,283 个测试对象)与 TRELLIS-500K(2,613 个测试对象)。 - **评估指标**:query IoU(体积占用交并比)、mesh Chamfer Distance(CD)、mesh F1@0.02。 - **核心结论**(见 Table 1): - **ShapeNet**:在 ** K=1 **、 L=5 时,ZipTok3D 的 CD 与 F1 与 **32-token COD-VAE** 持平,IoU 仅差 0.3 个百分点,序列长度缩短 **32 倍**;而 COD-VAE 在 K=2 时性能急剧下降。 - **TRELLIS**:在 ** K=4 **、 L=5 时,ZipTok3D 的 query IoU 与 COD-VAE-32 相当,CD 与 F1 更优,序列长度缩短 **8 倍**。 - **统计显著性**:对 TRELLIS 上 K=4 与 COD-VAE-32 的差异进行了配对、对象级别的非参数百分位 bootstrap 检验(20,000 次重采样),确认 CD 与 F1 的提升区间不包含零。 —- ### 2. 类别条件生成实验(Class-Conditioned Generation) 验证压缩前缀在下游生成任务中的有效性: - **设置**:在 ShapeNet 的五个类别(airplane、car、chair、table、rifle)上训练类条件 EDM。ZipTok3D 使用精确的 **2-token** stage-2 潜在代码( 2 × 32 ),配合 5 次共享细化步进行解码。 - **评估指标**:MMD-CD、COV-CD、1-NNA-CD(分布质量);采样吞吐量(Samp.)、完整吞吐量(Full)、峰值显存(Mem.)。 - **结论**(见 Table 2):ZipTok3D-2 的 stage-2 序列比 COD-VAE-32 短 **16 倍**,分布指标与之接近;采样吞吐量略高于 COD-VAE-32,完整吞吐量与显存占用相当,且显著优于 VecSet 系列。 —- ### 3. 消融实验(Ablation Study) 在 ShapeNet K=2 设置下,通过控制变量验证各组件贡献(见 Table 3): | 变体 | 说明 | 关键结论 | |———|———|—————| | **COD-VAE-2** | 原始基线 | IoU 77.7%,重建质量差 | | **Prefix only** | 仅引入嵌套 dropout,保留 COD-VAE 的 12 层单遍解码器 | IoU 提升至 92.3%,证明**前缀训练**本身有效 | | **w/o iter. ref.** | 使用 6 层解码器,但仅单遍前向传播 | IoU 91.9%,低于 Prefix only,说明浅层单次解码不足以展开短前缀 | | **w/o inter. sup.** | 6 层解码器 + 5 次共享细化,但**仅监督最终输出** | IoU 96.6%,显著提升,说明**迭代细化**是主要增益来源 | | **Full ZipTok3D** | 完整模型(加中间监督与自蒸馏) | IoU 96.9%,为最佳结果,中间监督提供额外补充收益 | —- ### 4. 前缀长度 K 与细化深度 L 的交互分析 利用同一 checkpoint 扫描所有支持的 (K, L) 组合: - **定量扫描**(见 Figure 4 与补充材料 Figure A/B): - 固定 L=1 时,增加 K 带来显著收益;当 L ≥ 3 后,性能对 K 的敏感度降低。 - 固定 K=1 或 4 时,增加 L 的收益最大;当 K 较大时,单遍解码已足够强。 - 所有指标(query IoU、mesh F1、mesh CD)在 L=5 附近趋于饱和。 - **逐对象细化行为**(补充材料 Table A):从 L=1 to 3 ,绝大多数对象在 IoU、CD、F1 上均受益;从 L=3 to 5 ,仍有超过半数对象继续改善,但比例下降且更依赖指标。 - **定性细化过程**(见 Figure 5): L=1 时捕捉部分全局结构但可能缺失组件; L=3 时恢复主要结构与连贯表面; L=5 时主要锐化薄结构、边界与局部细节。 —- ### 5. 效率与资源测量(Efficiency Analysis) 在 NVIDIA H20 GPU、batch size 16、FP32 推理、关闭 TF32 与 Flash SDP 的统一协议下测量: - **重建效率**(补充材料 Table B): - ZipTok3D 在 K ∈ 1,2,4 时,完整重建吞吐量约为 72 shapes/s,峰值显存约 2.03 GiB。 - 由于共享细化的循环计算与固定的密集占用场查询,其完整吞吐量略低于单遍 COD-VAE(约 80 shapes/s),但显著优于 VecSet-512(4.20 shapes/s)。 - **生成效率**(见 Table 2):ZipTok3D-2 的潜在采样吞吐量为 50.62 samples/s,完整吞吐量为 36.64 shapes/s,与 COD-VAE-32 相当。 —- ### 6. 自适应预算后验诊断(Post-Hoc Adaptive-Budget Diagnostic) 假设存在真实重建指标 oracle,按“先尝试更短 K ,再尝试更大 L ”的顺序搜索,严格接受“所有三项指标均不劣于 COD-VAE-32”的候选: - **结果**(补充材料 Table C 与 Figure C): - **ShapeNet**:76.31% 的对象可在平均 K=2.43 、 L=3.41 下达到或超过 COD-VAE-32。 - **TRELLIS**:67.62% 的对象可在平均 K=3.02 、 L=3.97 下达标。 - 成功子集中, L=3 与 L=4 分别占 ShapeNet 的 68.3% 与 TRELLIS 的 71.6%。 - **说明**:该诊断仅揭示**自适应预算潜力**,而非可直接部署的预算预测器。 —- ### 7. 额外定性比较(Additional Qualitative Comparisons) - 在 ShapeNet 与 TRELLIS 上补充展示了包含薄壁、大开孔、曲面和重复建筑元素的复杂对象(补充材料 Figure D)。 - ZipTok3D( K=4, L=5 )与 COD-VAE-32 的对比显示:两者均能恢复主要拓扑与空腔结构,差异主要集中在细小凸起、薄支撑、屋檐等极限几何处。 Q5: 有什么可以进一步探索的点? 基于论文结论与实验分析,以下方向值得进一步探索: —- ### 1. 自适应前缀长度与细化深度分配 论文后验诊断(Supplementary Sec. S5)表明,不同几何复杂度对象所需的最小 token 预算与解码轮次差异显著:简单形状仅需 K ≈ 2 、 L ≈ 3 ,而复杂结构需要更深细化。未来可训练一个**轻量级预测器**,根据输入特征或中间重建误差动态决定停止条件,而非对所有对象使用固定的 K 与 L 。这能在保持质量的同时进一步压缩平均计算成本。 —- ### 2. 面向极短序列的专用生成架构 当前类别条件生成仍基于传统 EDM 作用于 latent array,但 1–4 个 token 的前缀已逼近自回归或扩散模型在极短序列上的表达极限。未来可探索: - **非对称编解码生成器**:利用前缀的顺序信息性质设计因果或层次化生成先验; - **离散化/量化压缩**:将连续前缀进一步映射到紧凑码本,结合_lookup-based\
    生成模型降低下游扩散模型的计算开销。 —- ### 3. 解码端的效率与实时性优化 Table B 显示,ZipTok3D 的完整重建吞吐量(约 72 shapes/s)略低于单遍 COD-VAE(约 80 shapes/s),因为共享循环增加了 GPU 计算时间,而密集 128^3 占用查询成为新的瓶颈。未来可研究: - **自适应提前退出(adaptive early exit)**:在循环细化中监测状态收敛,动态减少 L ; - **稀疏/渐进式占用查询**:避免在每次细化后都查询完整密集网格,仅在空间不确定区域进行局部细化; - **与 3D Gaussian Splatting 或神经稀疏网格结合**,替代稠密 triplane 占用查询以降低显存与延迟。 —- ### 4. 从单物体扩展到复杂多物体场景 现有实验集中在 ShapeNet 单物体与 TRELLIS 结构化资产。如 Related Work 中 SceneTok 所示,全局 token 表示天然支持场景级压缩。将 ZipTok3D 的嵌套前缀机制扩展到**多物体、室内/室外场景**时,需要解决: - 场景内对象间的遮挡与空间关系如何编码进有限前缀; - 前缀的语义-几何解耦,使得场景编辑(如增删对象)无需重新编码整个序列。 —- ### 5. 多模态条件生成与编辑 当前仅验证了类别标签条件(class-conditioned)生成。前缀的紧凑性使其适合作为**文本、图像、草图或多视图**等异构模态的瓶颈接口。未来可探索: - 跨模态对齐:将 CLIP 或 3D 视觉语言特征与几何前缀联合训练; - 可控编辑:利用前缀的顺序信息特性,通过修改序列尾部 token 实现局部几何编辑(如风格迁移或部件替换),而不破坏前端编码的全局结构。 —- ### 6. 信息内容的几何可解释性分析 论文通过实验表明前端 token 保留了对象级几何,但缺乏对**前缀中具体编码了哪些频率/尺度信息**的定量分析。未来可从信息论角度: - 分析不同 K 下前缀的互信息下界; - 通过傅里叶域或谱分析,明确第 L 次细化主要恢复高频细节还是填补拓扑空缺,从而指导更优的中间监督策略。 —- ### 7. 结合动态与非刚性几何 现有框架针对静态刚性物体。将嵌套前缀与迭代细化推广到**时序动态、非刚性形变或拓扑变化**(如流体、关节动物)时,需要重新定义前缀的时序一致性,以及细化循环中运动/形变场的传播机制。 Q6: 总结一下论文的主要内容 这篇论文提出了 **ZipTok3D**,一种面向**极短 token 序列高保真 3D 重建**的 tokenizer。以下是对其主要内容的系统总结。 —- ### 1. 研究背景与核心挑战 现有 3D tokenizer 主要采用两类策略,但在极低 token 预算下均面临严重退化: - **空间结构化方法**(如 sparse voxel、octree)通过局部锚点保留空间结构,压缩需依赖更粗或更稀疏的支撑; - **全局 token 方法**(如 VecSet、COD-VAE)将对象级几何压缩为固定 token 集合,但在预算缩减至极少量时,全局结构与细粒度细节竞争有限的潜在容量,重建质量急剧下降。 核心挑战在于:**如何从一个仅含寥寥数个向量的序列中,同时保证编码端保留对象级几何,又能在解码端充分展开为精细的 3D 空间场?** —- ### 2. 方法:ZipTok3D 论文提出 ZipTok3D,通过**编码端的前缀信息重组**与**解码端的共享迭代展开**联合解决上述问题。 #### 2.1 嵌套前缀训练(Nested Prefix Training) - 编码器 Eφ 将输入点云映射为最多 M=128 个潜在向量 Z = (z_1, …, z_M) ∈ R^(M × d) 。 - 训练时对每个样本,从指数间隔集合 K = 1, 2, 4, 8, 16, 32, 64, 128 中均匀采样保留长度 K ,仅保留前缀 Z(:K) = (z1, …, z_K) ,掩码后缀。 - 重建目标强制要求**每个前缀独立重建完整对象**。这种嵌套约束使前端 token 优先保留对象级几何,后续 token 编码残差细节。 #### 2.2 参数共享的迭代细化解码(Iterative Global-to-Spatial Refinement) - 解码器以可学习的 triplane token T^0 初始化,通过选择模块 Sω 与保留前缀交互,得到选中状态 H^0_K 与冗余 token R_K :

(H^0K, R_K) = Sω(T^0, Z_(:K))

  • 一个六层的 Transformer 块 gθ 被**循环复用** L 次,逐步更新选中状态:
    H^ell_K = g
    θ(H^(ell-1)K, Z(:K)), quad ell = 1, …, L
  • 前缀 Z(:K) 与网络参数 θ 在迭代中固定不变,仅 triplane 状态被更新。每一步均可恢复为完整 triplane 并查询占用概率:
    y^ell
    (K,j) = σ(h_(occ)(T^ell_K, q_j))
  • 该设计将有效解码深度与参数量解耦,无需逐层新增参数,也**不依赖额外的生成采样阶段**即可从紧凑前缀直接重建。 #### 2.3 训练目标 - **最终输出监督**:对第 L 步的占用预测施加加权二元交叉熵损失 L_(rec) 。 - **中间监督与自蒸馏**:随机采样中间步 ell ∈ 1, …, L-1 ,对其施加真实标签监督与来自最终输出的自蒸馏损失:

L_(dist)(Y^ell_K, sg(Y^L_K))

  • 完整目标为:
    L(K,ell) = L(rec)(Y^LK, Y) + L(aux) + β[αt L(rec)(Y^ellK, Y) + (1-α_t)L(dist)(Y^ellK, sg(Y^L_K))]
    其中 L
    (aux) 包含 triplane 初始化与不确定性辅助损失。 #### 2.4 推理时的解耦控制 - **前缀长度 K **:决定潜在序列的信息量,通过截断直接调节。 - **细化深度 L **:决定解码计算量,通过循环次数直接调节。 两者互补:短前缀需更深细化,长前缀可配合浅层解码。 —- ### 3. 关键实验结果 #### 3.1 单物体重建 在 ShapeNet-v2 与 TRELLIS-500K 上的定量结果表明: - **ShapeNet**:使用 K=1 、 L=5 时,ZipTok3D 的 CD 与 F1 与 **32-token COD-VAE** 持平,IoU 仅低 0.3 个百分点,序列长度缩短 **32 倍**。 - **TRELLIS**:使用 K=4 、 L=5 时,query IoU 与 COD-VAE-32 相当,mesh CD 与 F1 更优,序列长度缩短 **8 倍**。 #### 3.2 类别条件生成 在 ShapeNet 五类对象上,ZipTok3D 采用 **2-token** 的 stage-2 潜在代码( 2 × 32 ),其分布指标(MMD-CD、COV-CD、1-NNA-CD)接近 COD-VAE-32,但 stage-2 序列短 **16 倍**,且采样与完整吞吐量具有竞争力。 #### 3.3 消融研究 - **Prefix only**:在 COD-VAE 原解码器上仅引入嵌套 dropout,IoU 从 77.7% 提升至 92.3%,验证前缀训练的有效性。 - **w/o iter. ref.**:移除迭代细化后性能显著下降,证明共享循环是主要增益来源。 - **w/o inter. sup.**:保留细化但仅监督最终输出,性能略低于完整模型,说明中间监督有补充收益。 #### 3.4 前缀长度与细化深度的交互 - K 与 L 呈**互补关系**:在 L=1 时增加 K 收益最大;在 K 极小时增加 L 收益最大。 - 性能在 L=5 附近趋于饱和; L=1 to 3 时绝大多数对象受益, L=3 to 5 主要锐化薄结构与边界。 —- ### 4. 结论与意义 ZipTok3D 通过**嵌套前缀训练**与**参数共享迭代细化**的协同设计,显式将对象级几何集中于潜在序列的前端,并通过循环解码逐步展开紧凑信息,首次实现了从 1–4 个全局 token 的高保真 3D 重建。该工作不仅显著降低了 3D 表示的序列长度,也为下游生成模型(如扩散模型、自回归模型)提供了更紧凑、更高效的潜在空间接口。 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: Mingda Lin, Weijie Wang, Zeyu Zhang, Bowen Cui, Yefei He, Haoyu Zhao, Yuanyu He, Donny Y. Chen, Feng Chen, Bohan Zhuang

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01740.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01740

Published: 2026-09-04T01:22:30.627Z


7. Evidential Deep Learning for Multi-Modal Anti-UAV Detection

Abstract:Anti-UAV systems increasingly fuse multiple sensors, yet their detection heads provide no per-modality reliability signal. This study evaluates whether evidential deep learning (EDL) heads, Dempster-Shafer (DS) evidence fusion, and uncertainty-driven temporal sensor gating improve anti-UAV detection through a controlled ablation on three benchmarks: thermal tracking (AntiUAV600), RGB-audio-RF classification (TRIDENT), and RGB-IR tracking (MM-UAV). The EDL training objective improves accuracy over retrained sigmoid baselines (+5.9 percentage points in accuracy and a tripled tracker-on-absent rate in E1; +4.8 percentage points in classification accuracy in E2, surviving a clip-clustered bootstrap, p = 0.011) and ranks classification errors substantially better (entropy UAUC approximately 0.94 vs. 0.51). The remaining components do not support their respective hypotheses. DS fusion does not outperform simple probability averaging. Dirichlet vacuity adds no ranking power beyond predictive entropy and inverts at the detection level, where extreme background imbalance causes it to encode class membership rather than error likelihood, a failure also observed for entropy and sigmoid confidence. Temporal gating preserves accuracy only when nearly inactive and yields no realised latency saving on shared-backbone hardware. The benefit of evidential learning therefore arises primarily from its training objective rather than its uncertainty estimate; a crop-level control further localises the detection-level breakdown to anchor-level evaluation rather than the learned representation.

中文摘要

摘要:反无人机系统越来越多地融合多种传感器,但其检测头未提供每种模式的可靠性信号。本研究通过对三个基准进行受控消融实验,评估证据深度学习(EDL)检测头、Dempster-Shafer(DS)证据融合和不确定性驱动的时序传感器门控是否能够提升反无人机检测性能:热成像跟踪(AntiUAV600)、RGB-音频-RF分类(TRIDENT)以及RGB-红外跟踪(MM-UAV)。EDL训练目标在重新训练的sigmoid基线模型上提高了准确率(E1中准确率提高5.9个百分点,缺失目标跟踪器触发率增加三倍;E2中分类准确率提高4.8个百分点,在基于剪辑集群的自助法中仍然显著,p=0.011),并在分类错误排名方面表现更佳(熵的UAUC约为0.94,对比0.51)。其余组件未支持各自的假设。DS融合未能优于简单概率平均。Dirichlet空性在预测熵之外不增加排名能力,并在检测级别出现反转,在极度背景不平衡情况下,它编码的是类别归属而非错误可能性,这种失败在熵和sigmoid置信度中亦有观察到。时序门控仅在几乎不活跃时能保持准确率,在共享骨干硬件上并未实现任何延迟节约。因此,证据学习的优势主要来自其训练目标,而非其不确定性估计;基于裁剪级的控制进一步将检测级的性能下降定位于锚框级评估,而非学习到的表征。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01742 (HTTP 429)

Authors: Dmitry Golovchits, Seyed Sahand Mohammadi Ziabari, Ali Mohammed Mansoor Alsahag

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01742.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01742

Published: 2026-09-04T01:22:30.627Z


8. SCULPT: Training Edge Vision Models for Post-Training Quantization Readiness

Abstract:Edge vision models are difficult to deploy on resource-constrained hardware, making low-bit post-training quantization (PTQ) attractive. In practice, standard FP32 training often produces heavy-tailed activation distributions whose outliers destabilize activation quantization: preserving the full range wastes quantization bins on rare extremes, while aggressive clipping causes information loss. Existing solutions typically rely on quantization-aware training (QAT), which adds training complexity and bit-width coupling, or advanced PTQ procedures that repair the model after training. We present SCULPT (Statistical Clipping and Uniform Loss for Post-Training), a training-time method that improves PTQ readiness during ordinary FP32 fine-tuning. SCULPT combines a topology-aware activation regularizer that suppresses quantization-hostile skewness and kurtosis with a stable percentile-based clipping mechanism that learns deployment-ready activation bounds. Unlike QAT, SCULPT does not simulate quantization during optimization; unlike post hoc outlier-repair PTQ methods, it does not require runtime activation transformations. The learned clipping bounds can be exported directly into a standard PTQ workflow for low-bit deployment, including INT8 and lower-bit settings such as W4A8.

中文摘要

摘要:边缘视觉模型在资源受限的硬件上难以部署,使得低比特后训练量化(PTQ)具有吸引力。在实践中,标准的 FP32 训练通常会产生重尾激活分布,其异常值会使激活量化不稳定:保留全范围会浪费量化区间在罕见极端值上,而过度裁剪则会造成信息丢失。现有解决方案通常依赖量化感知训练(QAT),这增加了训练复杂性和比特宽度耦合,或依赖先进的 PTQ 流程在训练后修复模型。我们提出了 SCULPT(后训练的统计裁剪与均匀损失),这是一种训练阶段方法,可在普通 FP32 微调期间提高 PTQ 准备性。SCULPT 将一种考虑拓扑的激活正则化器(抑制不利于量化的偏度和峰度)与稳定的基于百分位数的裁剪机制结合起来,从而学习适合部署的激活边界。与 QAT 不同,SCULPT 不在优化期间模拟量化;与事后异常值修复 PTQ 方法不同,它不需要在运行时进行激活变换。所学习的裁剪边界可以直接导出到标准 PTQ 工作流程中以进行低比特部署,包括 INT8 和更低比特设置,如 W4A8。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01743 (HTTP 429)

Authors: Bharadwaj Kavuri, Sourav Babu-PK, Varadhraj Ellapan, Pullarao Maddu, Prasad Deshpande

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01743.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01743

Published: 2026-09-04T01:22:30.627Z


9. Swin Meets EfficientNet: Lightweight Architectures for GAN-Based Face Forensics

Abstract:Modern generative models, such as GANs, diffusion architectures, and autoregressive systems, now produce facial images that are nearly indistinguishable from authentic photographs. This capability makes detecting forged images increasingly difficult, raising serious concerns about identity theft, fraud, and misinformation campaigns. Our research focuses specifically on GAN-generated synthetic faces, which underpin many face-centric deepfakes, and investigates efficient detection approaches using image analysis alone. Existing detection systems rely heavily on either convolutional neural networks (CNNs) or global vision transformers. While CNNs excel at identifying texture-based local features, they struggle with broader contextual understanding. Traditional Vision Transformer (ViT) models can capture long-range structures effectively, but demand substantial computational resources. Our work explores Swin-Transformer-based architectures across three implementations: a compact Swin Transformer trained from the ground up, ImageNet-1K pre-trained Swin-Tiny and Swin-Small models adapted for binary classification, and a novel hybrid combining EfficientNet-B0’s convolutional processing with a Swin Transformer backend. We evaluated all models using the 140K Real and Fake Faces dataset, which includes StyleGAN-generated fake faces alongside authentic images from Flickr and DFDC, with balanced splits for training, validation, and testing. The EfficientNetB0+Swin hybrid achieved 99% accuracy and a 99.44% recall on 5,000 test images, outperforming both pure Swin variants and a previous CNN-only baseline on this dataset. Our results suggest that combining hierarchical CNN features with shifted-window self-attention provides an efficient and computationally lightweight method for detecting GAN-generated synthetic faces.

中文摘要

摘要:现代生成模型,如GAN、扩散架构和自回归系统,现在能够生成几乎与真实照片无法区分的面部图像。这种能力使得检测伪造图像变得越来越困难,引发了关于身份盗用、欺诈和虚假信息传播的严重担忧。我们的研究特别关注基于GAN的合成面部图像,它们是许多以人脸为中心的深度伪造的基础,并研究仅使用图像分析的高效检测方法。现有的检测系统主要依赖卷积神经网络(CNN)或全局视觉Transformer。CNN在识别基于纹理的局部特征方面表现出色,但在更广泛的上下文理解上存在困难。传统的视觉Transformer(ViT)模型能够有效捕捉远程结构,但需要大量计算资源。我们的工作探索了基于Swin-Transformer的架构的三种实现方式:一种从零开始训练的紧凑型Swin Transformer,经过ImageNet-1K预训练并适配于二分类任务的Swin-Tiny和Swin-Small模型,以及一种新的混合模型,将EfficientNet-B0的卷积处理与Swin Transformer的后台结合。我们使用包含140K真实与伪造面部的Dataset进行了模型评估,该数据集包括StyleGAN生成的伪造面部图像,以及来自Flickr和DFDC的真实图像,并在训练、验证和测试中进行了均衡分割。EfficientNetB0+Swin混合模型在5,000张测试图像上达到了99%的准确率和99.44%的召回率,优于纯Swin变体及该数据集上之前的仅CNN基线。我们的结果表明,结合分层CNN特征与移位窗口自注意力提供了一种高效且计算轻量的GAN生成合成面部检测方法。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01749 (HTTP 429)

Authors: Sejuti Basu, Ashima Sood, Vijay Kumar, Sahil Sharma

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01749.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01749

Published: 2026-09-04T01:22:30.627Z


10. AlphaRAD: Grounded Zero-Shot Classification in Chest Radiology via $α$-Corrected Binary Cross Entropy and Factorized Latent Supervision

Abstract:Vision-Language Pretrained Models (VLPMs) offer a scalable path to open-vocabulary chest radiology understanding, yet two aspects remain underexplored: how structured clinical semantics extracted from medical reports can reduce in-batch noise during contrastive learning, and how cross-modal fusion can be designed to produce more faithful spatial grounding without added complexity. We introduce AlphaRAD, addressing these opportunities through two contributions. First, we construct a large-scale structured medical concept space from medical reports parsed by a Large Language Model for training, thereby mitigating in-batch learning noise and removing heuristic pair matching in contrastive learning, and thus naturally positioning AlphaRAD as a medical concept discriminator trained via $\alpha$-Corrected Binary Cross-Entropy. Second, we propose FLaS (Factorized Latent Supervision), an extremely simple yet effective cross-modal feature fusion module that factorizes VLPM representations into independent subspaces, using dedicated alignment supervision to enhance the expressiveness of spatial grounding without introducing additional model parameters. Through extensive empirical validation, AlphaRAD shows strong zero-shot generalization across diverse chest radiology tasks. Notably, it establishes state-of-the-art average performance across 16 classification benchmarks, while achieving individual state-of-the-art results via distinct gains on 7 grounding/phrase grounding and 3 segmentation datasets.

中文摘要

摘要:视觉语言预训练模型(VLPM)为开放词汇胸腔放射学理解提供了可扩展的路径,但有两个方面尚未被充分探讨:如何从医疗报告中提取结构化临床语义,减少对比学习中的批内噪声;以及如何设计跨模态融合以实现更忠实的空间基础,同时不增加复杂度。我们介绍了AlphaRAD,通过两项贡献应对这些机遇。首先,我们构建了一个由大型语言模型解析的医疗报告进行训练的大规模结构化医学概念空间,从而减少批次学习噪声并消除对比学习中的启发式配对匹配,从而自然地将AlphaRAD定位为通过$\alpha$修正二元交叉熵训练的医学概念判别器。其次,我们提出了FLaS(分解潜在监督),这是一个极其简单但高效的跨模态特征融合模块,通过专用比对监督将VLPM表示分解为独立子空间,提升空间基准表达性,而无需引入额外模型参数。通过广泛的实证验证,AlphaRAD在多种胸腔放射学任务中展现出强有力的零样本推广能力。值得注意的是,它在16个分类基准中建立了最先进的平均性能,同时通过7个基础/短语基础数据集和3个分割数据集的显著提升,实现了各自最先进的结果。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01757 (HTTP 429)

Authors: Jianzhong You, Yuan Gao, Chris McIntosh

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.01757.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.01757

Published: 2026-09-04T01:22:30.627Z