数据来源：HuggingFace Papers

Latest Papers

1. Compile by Training: Turning Natural-Language Specifications into Local Neural Functions

Abstract:Many recurring text functions are easy to describe but difficult to implement with rules, while calling a large remote model for every input introduces repeated cost, latency, and dependency on a provider. We present compile by training, which turns a natural-language specification into a reusable neural function. At compile time, teacher models generate task-specific examples that are used to train a small adapter for a compact interpreter. The resulting function runs without the teachers and can be stored, versioned, and composed like ordinary software. On FuzzyBench-Hard, a subset on which the Program-as-Weights fast compiler produced no exact matches, compile by training reaches 83.6% semantic accuracy. This higher accuracy comes with a higher compile-time cost: roughly a minute rather than seconds for the fast compiler. We deploy the compiler in a public interactive service and demonstrate compiled functions in a multi-site website helper, a language-controlled 3D avatar, and a bidirectional English-Claudish translator.

中文摘要

摘要：许多重复出现的文本功能容易描述，但难以通过规则实现，而每次输入都调用大型远程模型会带来重复的成本、延迟以及对服务提供商的依赖。我们提出了“通过训练编译”（compile by training）的方法，它将自然语言规范转化为可重用的神经函数。在编译阶段，教师模型生成特定任务的示例，这些示例用于训练一个用于紧凑解释器的小型适配器。生成的函数可以在没有教师模型的情况下运行，并且可以像普通软件一样存储、版本控制和组合。在 FuzzyBench-Hard 数据集（一个快速编译器 Program-as-Weights 没有生成精确匹配的子集）上，“通过训练编译”的语义准确率达到 83.6%。这种更高的准确率伴随着更高的编译时间成本：大约一分钟，而快速编译器只需几秒钟。我们将该编译器部署在公共交互服务中，并展示了在多站点网站助手、语言控制的 3D 化身以及双向英-克劳迪语翻译器中的编译函数应用。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04199 (Request failed with status code 502)

Authors: Yuntian Deng,Pengyu Nie,Stuart Shieber

Categories:

PDF URL: https://arxiv.org/pdf/2609.04199.pdf

Arxiv URL: https://arxiv.org/abs/2609.04199

Arxiv ID: 2609.04199

CoolPaper URL: https://papers.cool/arxiv/2609.04199

Published: 2026-09-07T01:18:08.176Z

Updated: 2026-09-07T01:18:08.176Z

2. Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments

Abstract:As terminal-based code agents become prevalent, agent trajectories have accumulated at scale, while realistic, executable environments remain scarce. However, environments are what agent post-training actually requires: each can be re-queried into many verifiable tasks and provides execution feedback, whereas a trajectory is a single frozen demonstration. Rather than generating environments from scratch, we observe that the tool-execution history in existing trajectories exposes the structure and contents of the environments in which they ran, making it possible to reconstruct those environments from the trajectories themselves. Thus, we introduce Terminal-Universe, a framework which turns each trajectory into a reusable environment and explores it for synthesizing new tasks and continued interactions. Specifically, Terminal-Universe replays the file operations recorded in a trajectory to restore each file before the agent modified it, yielding a partial workspace; a completion agent then supplies the missing files and dependencies. On this recovered workspace, we both reconstruct the original intent task and synthesize entirely new ones. Besides, we also scale the tasks along two complementary axes: breadth and depth. For breadth, we mine directional dependency relations between related environments and synthesize cross-workspace queries spanning multiple codebases, as developers routinely do in real-world development. For depth, we extend the initial single-turn query into a multi-round session that captures iterative user feedback and requirement refinement via a user agent. Applied to public terminal agent trajectories, Terminal-Universe produces 37.3k task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on this corpus improves single-round performance on Terminal-Bench 2.1 by 11.9 points and multi-round performance on EvoCode-Bench v2 MT@4 by 13.8 points.

中文摘要

摘要：随着基于终端的代码代理变得普遍，代理轨迹已大规模累积，而现实的、可执行的环境仍然稀缺。然而，环境才是代理后训练真正需要的：每个环境可以被多次查询成许多可验证的任务并提供执行反馈，而轨迹只是一个固定的演示。我们观察到，与其从零生成环境，不如利用现有轨迹中的工具执行历史，这些历史揭示了代理运行环境的结构和内容，从而可以直接从轨迹本身重建这些环境。因此，我们提出了Terminal-Universe，一个将每条轨迹转化为可重复使用环境并在其中探索以合成新任务和持续交互的框架。具体而言，Terminal-Universe会重放轨迹中记录的文件操作，以在代理修改前恢复每个文件，从而生成一个部分工作区；然后，完成代理会补全缺失的文件和依赖。在这个恢复的工作区上，我们既重建原始意图任务，也合成全新的任务。此外，我们还沿两个互补维度扩展任务规模：广度和深度。在广度方面，我们挖掘相关环境之间的方向性依赖关系，并合成跨工作区的查询，涉及多个代码库，就像开发者在真实开发中常做的那样。在深度方面，我们将初始的单轮查询扩展为多轮会话，通过用户代理捕获迭代的用户反馈和需求细化。在应用于公开的终端代理轨迹后，Terminal-Universe产生了37.3k个任务充分环境。在此语料上对Qwen3.5-27B进行监督微调，使其在Terminal-Bench 2.1上的单轮表现提升了11.9分，在EvoCode-Bench v2 MT@4上的多轮表现提升了13.8分。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04148 (Request failed with status code 502)

Authors: Jie Wu,Zhenru Zhang,Beichen Zhang,Xuwu Wang,Yuhui Su,Mouxiang Chen,Peng Wang,Zhihai Wang,Que Shen,Hao Zhou,An Yang,Fei Huang,Yujiu Yang,Dayiheng Liu

Categories:

PDF URL: https://arxiv.org/pdf/2609.04148.pdf

Arxiv URL: https://arxiv.org/abs/2609.04148

Arxiv ID: 2609.04148

CoolPaper URL: https://papers.cool/arxiv/2609.04148

Published: 2026-09-07T01:18:08.122Z

Updated: 2026-09-07T01:18:08.122Z

3. LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

Abstract:We introduce LLaDA-Image, a unified framework that pairs a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. Instead of relying heavily on paired image-text data from the beginning, we first build a strong visual generative prior through image-only pre-training and mid-training. The generation pipeline comprises 220M samples, 98 of which are real images. For efficient and scalable optimization, we use parameter-free RMSNorm throughout the DiT together with the Muon optimizer. The resulting unified model produces highly photorealistic images while accurately following fine-grained editing instructions. We further distill LLaDA-Image into LLaDA-Image-Turbo, enabling fast inference in 2-4 sampling steps. On Qwen-Image-Bench, LLaDA-Image achieves overall scores of 53.53 and 53.38 on the English and Chinese tracks, respectively, setting a new state-of-the-art among open-source models on both tracks. To support further research on capable and efficient generative models, we release our model weights, training code, and detailed recipes.

中文摘要

摘要：我们介绍了 LLaDA-Image，这是一个统一框架，它将从零训练的 6B 扩散变换器（DiT）与基于 LLaDA2.0-Mini 扩散语言模型骨干的冻结视觉语言理解模块配对。我们并非从一开始就严重依赖配对的图像-文本数据，而是首先通过仅图像的预训练和中期训练建立强大的视觉生成先验。生成流程包含 2.2 亿个样本，其中 98% 是真实图像。为了实现高效且可扩展的优化，我们在整个 DiT 中使用无参数的 RMSNorm 并配合 Muon 优化器。由此得到的统一模型能够生成高度逼真的图像，同时准确遵循细粒度编辑指令。我们进一步将 LLaDA-Image 蒸馏为 LLaDA-Image-Turbo，实现 2-4 步采样的快速推理。在 Qwen-Image-Bench 上，LLaDA-Image 在英文和中文轨道上分别取得了 53.53 和 53.38 的总体评分，在两个轨道上的开源模型中建立了新的最先进水平。为了支持在高效生成模型方面的进一步研究，我们发布了模型权重、训练代码和详细的训练方案。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.03796 (Request failed with status code 502)

Authors: Chuyan Chen,Haoxing Chen,Kun Chen,Zhenglin Cheng,Long Cui,Ruishan Fang,Zhangxuan Gu,Zhicheng Huang,Zhenzhong Lan,Yuanting Lei,Haoquan Li,Jianguo Li,Rongchuan Li,Sidu Li,Tao Lin,Deyuan Liu,Jiacheng Liu,Lin Liu,Yuxuan Lou,Zhisheng Lu,Yuxin Ma,Shuheng Shen,Peng Sun,Chaoyang Wang,Hongjun Wang,Xiaomei Wang,Yongxin Wang,Chengzhang Wu,Hongru Wu,Jun Xie

Categories:

PDF URL: https://arxiv.org/pdf/2609.03796.pdf

Arxiv URL: https://arxiv.org/abs/2609.03796

Arxiv ID: 2609.03796

CoolPaper URL: https://papers.cool/arxiv/2609.03796

Published: 2026-09-07T01:18:08.150Z

Updated: 2026-09-07T01:18:08.150Z

4. Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning

Abstract:Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, and keep the top-scoring ones. We show that the selection signal contributes almost nothing. Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all; across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput than it in vLLM deployment. Controlled experiments explain this by showing that 1) the prompt is the fragile part of the cache, and most of the gap between selectors is just whether their selection signal happened to keep it; 2) the reasoning trace protects itself against eviction with redundancy at two levels, in the text (the model restates what it still needs as it works) and across attention heads (each keeps its own copy of the trace), so once the prompt is safe, a random draw retains enough copies of what the model still needs, and no score is required to pick them. Our code is publicly available at this https URL.

中文摘要

摘要：大型语言模型在需要长时间推理的任务上表现出色，但长链的思维过程使 KV 缓存成为严重的内存瓶颈。现有的 KV 缓存压缩方法共有一个范式：通过某种对未来重要性进行估计的方式对每个缓存的 token 进行评分，然后保留得分最高的。我们表明，选择信号几乎没有贡献。随机注意力（Random Attention）保留提示（prompt），在每个注意力头内以均匀随机的方式逐出，并且完全不计算任何分数；在四个模型和六个推理任务中，它与最强的先前逐出器匹配，同时在 vLLM 部署中提供比其高 32-43% 的吞吐量。控制实验解释了这一点：1）提示是缓存中最脆弱的部分，大多数选择器之间的差距仅仅是因为它们的选择信号碰巧保留了提示；2）推理轨迹在两个层面上通过冗余保护自己免受逐出：在文本层面（模型在工作时会重述仍然需要的内容）和在注意力头层面（每个头保留自己的轨迹副本），所以一旦提示安全，随机抽取就能保留模型仍然需要的足够副本，而且无需分数来选择它们。我们的代码已在此 https URL 公共提供。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.03430 (Request failed with status code 502)

Authors: Heng Wang,Jielin Qiu,Wenting Zhao,Cheng Qian,Liangwei Yang,Jiawei Han,Heng Ji,Silvio Savarese,Shelby Heinecke,Huan Wang

Categories:

PDF URL: https://arxiv.org/pdf/2609.03430.pdf

Arxiv URL: https://arxiv.org/abs/2609.03430

Arxiv ID: 2609.03430

CoolPaper URL: https://papers.cool/arxiv/2609.03430

Published: 2026-09-07T01:18:08.160Z

Updated: 2026-09-07T01:18:08.160Z

5. Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training

Abstract:Large language models offer broad capabilities, but adapting them to evolving domains, tools, and requirements often entails repeated post-training. Autonomous systems automate parts of this process by proposing updates, training candidates, and using evaluation feedback to select subsequent proposals. As evidence accumulates, a central problem emerges: which past update evidence remains actionable after subsequent training has changed the parent model? An update’s effect depends on its parent, data, and training stage. Treating past success as context-free permission can waste compute. If the resulting child is promoted, it can also degrade the subsequent training trajectory. We formulate this problem as conditional experience transfer and introduce Boundary-Calibrated Intervention Transfer (BCIT), a method that authorizes experience reuse before weight-changing training. BCIT binds an observed effect to its source context, checks applicability conditions, vetoes candidates with named hard conflicts, and obtains current-state evidence through a bounded training trial when needed. Fully trained candidates still face a shared adoption rule, and only observed events extend memory. On one 4B model adapted across finance reasoning, text-to-SQL, and function calling, candidate updates exhibit heterogeneous target and retention effects across the evaluated contexts. Under matched candidates, evidence, and compute, BCIT authorizes fewer harmful updates and attains higher equal-budget final-model quality than the evaluated alternatives. These results support treating experience authorization as a distinct problem in autonomous post-training.

中文摘要

摘要：大语言模型提供了广泛的能力，但将它们适应不断变化的领域、工具和需求通常需要反复进行训练后调整。自主系统通过提出更新、训练候选模型并利用评估反馈选择后续提案来自动化这一过程的部分环节。随着证据的积累，一个核心问题出现了：在随后的训练改变了父模型之后，哪些过去的更新证据仍然可操作？一次更新的效果取决于其父模型、数据和训练阶段。将过去的成功视为无上下文的许可可能会浪费计算资源。如果生成的子模型被推广，也可能会削弱随后的训练轨迹。我们将这一问题形式化为条件性经验转移，并提出界限校准干预转移（BCIT）方法，该方法在改变权重的训练之前授权复用经验。BCIT将观察到的效果绑定到其源上下文，检查适用条件，否决具有明确冲突的候选，并在需要时通过有限训练试验获取当前状态的证据。即便是完全训练的候选模型，仍需遵循共享的采纳规则，且只有观察到的事件会延伸到记忆中。在一款跨金融推理、文本转SQL和函数调用适应的4B模型上，候选更新在不同评估上下文中表现出异质的目标和保留效果。在匹配候选模型、证据和计算条件下，BCIT授权的有害更新更少，并且在相同预算下最终模型质量高于评估的其他方法。这些结果支持将经验授权视为自主训练后处理中的独立问题。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2608.26730 (Request failed with status code 502)

Authors: Tingyun Li,Wenfeng Feng,Weiqing Li,Abudukelimu Wuerkaixi,Guohua Liu,Yuewei Zhang

Categories:

PDF URL: https://arxiv.org/pdf/2608.26730.pdf

Arxiv URL: https://arxiv.org/abs/2608.26730

Arxiv ID: 2608.26730

CoolPaper URL: https://papers.cool/arxiv/2608.26730

Published: 2026-09-07T01:18:08.199Z

Updated: 2026-09-07T01:18:08.199Z

6. RoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning

Abstract:Robot learning increasingly depends on broad and diverse demonstrations, yet collecting robot data remains expensive and poorly suited to covering the long tail of real-world tasks. To address this bottleneck, we introduce RoboTok, an internet-scale data engine that, given a query human manipulation video, retrieves manipulation-relevant human demonstrations from web videos for training dexterous robot policies. Specifically, we learn a latent motion space from 3D hand trajectories expressed in estimated actor-centered reference frames. This representation enables manipulation behaviors to be compared across variations in camera viewpoint, scene appearance, and actor occlusions, while remaining compact enough for efficient search and continual indexing over internet-scale video collections. We evaluate RoboTok against existing robot-data retrieval approaches on retrieval benchmarks and downstream robot policy performance. Our results show that RoboTok retrieves more relevant manipulation demonstrations and improves downstream task success, establishing hand-pose trajectory-aware retrieval as a way to make web video a scalable and continuously growing source of supervision for robot learning.

中文摘要

摘要：机器人学习越来越依赖广泛而多样的示范，但收集机器人数据仍然成本高昂，且难以覆盖现实世界任务的长尾部分。为了解决这一瓶颈，我们引入了RoboTok，一个互联网规模的数据引擎。给定一个查询的人类操作视频，它可以从网络视频中检索与操作相关的人类示范，用于训练灵巧的机器人策略。具体来说，我们从以估计的以演员为中心的参考系表示的3D手部轨迹中学习潜在运动空间。这种表示能够在摄像机视角、场景外观和演员遮挡变化下比较操作行为，同时足够紧凑以便在互联网规模的视频集合上进行高效搜索和持续索引。我们在检索基准和下游机器人策略性能上评估了RoboTok相对于现有机器人数据检索方法的表现。结果表明，RoboTok能够检索更多相关的操作示范并提升下游任务成功率，确立了手部姿态轨迹感知的检索方式，使网络视频成为机器人学习可扩展且持续增长的监督来源。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.03199 (Request failed with status code 502)

Authors: Howard Qian,Yiting Chen,Yunfei Xie,Kejia Ren,Podshara Chanrungmaneekul,Gaotian Wang,Bowen Wen,Chen Wei,Kaiyu Hang

Categories:

PDF URL: https://arxiv.org/pdf/2609.03199.pdf

Arxiv URL: https://arxiv.org/abs/2609.03199

Arxiv ID: 2609.03199

CoolPaper URL: https://papers.cool/arxiv/2609.03199

Published: 2026-09-07T01:18:26.745Z

Updated: 2026-09-07T01:18:26.745Z

7. LatentPress: Context Compression Beyond Text and Vision

Abstract:Compressed context is usually carried as human-readable text or as rendered images that must be decoded, even when its consumer is a language model. We introduce LatentPress, which writes conversational histories and long documents into a third representation: continuous memory tokens that a frozen decoder reads directly through its input-embedding interface, with no text reconstruction at inference. A small reader-matched writer compresses $4$-$16\times$ while training only an adapter (4.2M-26.2M parameters, $\sim!0.1\%$ of the decoder). On LongMemEval, LatentPress reaches $0.504$ accuracy at $7.70\times$ compression versus $0.490$ for uncompressed evidence, outperforming text summaries (0.184) and OCR-based compression (0.426 to 0.312). On LongBench-QA, in-domain writers match or exceed raw-context reading at $4$-$8\times$ compression, while $16\times$ trails raw. Writing takes 43ms per conversation, roughly an order of magnitude faster than text summarization or OCR reconstruction, and reading is $5$-$9\times$ faster than raw context or cached OCR. We validate the interface under two transfer settings, zero-shot from UltraChat to LongMemEval memory QA and from LongMemEval-derived QA to unseen LongBench document domains, establishing direct soft tokens as a practical machine-facing context interface beyond text and vision. The implementation of the experiments could be found at: this https URL .

中文摘要

摘要：压缩上下文通常以人类可读的文本或必须解码的渲染图像的形式存在，即使其使用者是语言模型。我们引入了 LatentPress，它将对话历史和长文档写入第三种表示形式：连续记忆标记，冻结的解码器通过其输入嵌入接口直接读取这些标记，在推理过程中无需文本重建。一个小型、与读取器匹配的写入器在仅训练适配器（4.2M-26.2M 参数，约占解码器的 0.1%）的情况下能压缩 4-16 倍。在 LongMemEval 上，LatentPress 在 7.70 倍压缩下达到 0.504 的准确率，而未压缩证据为 0.490，优于文本摘要（0.184）和基于 OCR 的压缩（0.426 到 0.312）。在 LongBench-QA 上，域内写入器在 4-8 倍压缩下匹配或超过原始上下文读取，而 16 倍压缩低于原始读取。写入每次对话需要 43ms，大约比文本摘要或 OCR 重建快一个数量级，而读取比原始上下文或缓存的 OCR 快 5-9 倍。我们在两种迁移设置下验证了该接口：从 UltraChat 到 LongMemEval 记忆 QA 的零样本迁移，以及从 LongMemEval 衍生 QA 到未见的 LongBench 文档领域，从而确定直接软标记作为一种实用的面向机器的上下文接口，超越了文本和视觉。实验实现可在此 https URL 找到。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.01507 (Request failed with status code 502)

Authors: Zhengze Zhou,Hejian Sang

Categories:

PDF URL: https://arxiv.org/pdf/2609.01507.pdf

Arxiv URL: https://arxiv.org/abs/2609.01507

Arxiv ID: 2609.01507

CoolPaper URL: https://papers.cool/arxiv/2609.01507

Published: 2026-09-07T01:18:26.863Z

Updated: 2026-09-07T01:18:26.863Z

8. Rethinking On-Policy Distillation of Large Language Models II: One Training Example

Abstract:On-policy distillation (OPD) combines student-generated rollouts with dense token-level supervision from a teacher. Existing work has mainly studied its algorithmic behavior, leaving the role of training data unclear. We examine this role at the data-minimal limit by training on a single query. One-shot OPD keeps improving for hundreds of steps and recovers most of full-data OPD’s gain across task domains and model families. We explain this result through the states visited during training and the rate at which the student aligns with the teacher. We measure \emph{state coverage}, the fraction of the states full-data OPD visits that a query set’s rollouts reach. A single query already reaches (71.5\%), most of it within the first 100 steps. Adding semantically distinct queries raises coverage and validation accuracy together, until 16 queries reach (98.9\%) and match full-data training. Yet alignment slows at a similar pace whether OPD trains on one query or the whole dataset, and even a fixed set of states takes hundreds of steps to absorb. OPD is therefore data-overfed but algorithm-starved. Its rollouts quickly expose broad supervision, while the student absorbs that supervision increasingly slowly. The state-coverage result extends to multi-teacher OPD, where 16 semantically diverse queries per domain match full-data MOPD. As a further stress test, content-light templates and off-domain WildChat queries also approach the real-query baseline. Task content and induced state coverage can therefore come apart. We hope these findings direct future work toward the step efficiency of OPD, and prompt a re-examination of the data and the mechanisms behind its recent successes in frontier post-training.

中文摘要

摘要：策略提纯（OPD）结合了学生生成的推广与教师密集的代币级监督。现有研究主要研究其算法行为，训练数据的作用尚不明确。我们通过对单一查询进行训练，在数据极小限制下考察这一角色。一次性OPD在数百步内不断改进，并恢复了全数据OPD在任务域和模型族间的大部分收益。我们通过培训期间访问的状态以及学生与教师对齐的速度来解释这一结果。我们测量\emph{state coverage}，即查询集的部署覆盖到的州全数据OPD访问比例。单个查询已达到\（71.5\%），大部分在前100步内完成。添加语义上不同的查询可以同时提升覆盖率和验证准确性，直到16个查询达到\（98.9\%\）并匹配全数据训练。然而，无论是针对一个查询还是整个数据集训练，对齐速度都差不多，即使是固定状态集也需要数百步才能吸收。因此，OPD数据过剩但算法不足。其推广迅速暴露出广泛的监督，而学生吸收监督的速度却越来越慢。状态覆盖结果延伸到多教师OPD，每个领域16个语义多样查询匹配全数据MOPD。作为进一步的压力测试，内容轻模板和域外WildChat查询也接近实查询基线。任务内容与诱导状态覆盖可能分离。我们希望这些发现能引导未来提升门诊的阶梯效率，并促使对数据及其近期前沿培训成功背后的机制进行重新审视。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04172 (Request failed with status code 502)

Authors: Zixuan Fu,Bingxiang He,Yuxin Zuo,Haohuan Huang,Jinqian Zhang,Ruhang Xiao,Cheng Qian,Qinyu Luo,Huan-ang Gao,Yudong Wang,Zhiyuan Liu,Ning Ding,Chaojun Xiao

Categories:

PDF URL: https://arxiv.org/pdf/2609.04172.pdf

Arxiv URL: https://arxiv.org/abs/2609.04172

Arxiv ID: 2609.04172

CoolPaper URL: https://papers.cool/arxiv/2609.04172

Published: 2026-09-07T01:18:26.996Z

Updated: 2026-09-07T01:18:26.996Z

9. Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM

Abstract:Hybrid LLMs pair softmax attention with linear-attention layers such as Gated DeltaNet (GDN), whose recurrent state summarizes the context in fixed size. Early community 4-bit quantizations of Qwen3.8-27B (48 GDN layers, 16 attention layers) left the GDN block in 8- or 16-bit precision — especially its decay and write-strength gates — on the intuition that errors in a recurrence accumulate over long contexts. We test that intuition by building Minima: NVFP4 W4A4 on all 496 linear layers, GDN included. Across perplexity at 4K/32K, MMLU-Pro, GSM8K, AIME’25, GPQA-Diamond, LiveCodeBench, and RULER retrieval to 64K, Minima matches BF16 within seed noise (5-task average -0.52) while being the smallest (17.5 GiB) and fastest-prefill (+14-19%) recipe we compare, and its 32K perplexity gap shrinks with position. A four-part mechanism study explains why: (i) NVFP4’s 16-element block scaling localizes the residual stream’s extreme outliers, equalizing activation error across layer roles; (ii) the supposedly fragile gate projections are the least sensitive — softplus/exponential and sigmoid parameterizations compress ~11% GEMM error to ~2% output error; (iii) the delta-rule recurrence holds injected noise at a flat plateau over 32K tokens and forgets a state impulse within hundreds of steps, because each write overwrites the state along the current key direction; (iv) the per-token quantization cost washes out with context instead of compounding. We also repair a global-scale mismatch that arises when per-module-calibrated NVFP4 checkpoints are served by kernels that fuse those modules into one GEMM, and show calibrated FP8 KV-cache scales are performance-free. The result: a practical recipe — quantize everything, ship KV scales — and a mechanistic account of why the recurrent half of a hybrid LLM is the easy half to quantize. Checkpoint: this https URL

中文摘要

摘要：混合大语言模型（Hybrid LLM）将 softmax 注意力与线性注意力层（如 Gated DeltaNet（GDN））结合使用，其循环状态以固定大小总结上下文。早期社区对 Qwen3.8-27B（48 层 GDN，16 层注意力）的 4 位量化尝试中，将 GDN 块保持在 8 位或 16 位精度——尤其是其衰减和写入强度门——基于这样一种直觉：循环中的错误会在长上下文中积累。我们通过构建 Minima 来验证这一直觉：对所有 496 个线性层，包括 GDN，使用 NVFP4 W4A4。在 4K/32K 对困惑度（perplexity）、MMLU-Pro、GSM8K、AIME’25、GPQA-Diamond、LiveCodeBench 和 RULER 检索至 64K 的测试中，Minima 在种子噪声范围内与 BF16 匹配（5 任务平均 -0.52），同时成为我们比较的最小（17.5 GiB）且预填充速度最快（+14-19%）的方案，并且其 32K 困惑度差距随位置减小。一项四部分的机制研究解释了原因：（i）NVFP4 的 16 元块缩放将残差流的极端异常值局部化，使激活误差在各层角色之间均衡；（ii）被认为脆弱的门投影实际上最不敏感——softplus/指数和 sigmoid 参数化将约 11% 的 GEMM 误差压缩到约 2% 的输出误差；（iii）delta-rule 循环在 32K 个标记上保持注入噪声于平坦平台，并在数百步内遗忘状态冲击，因为每次写入都沿当前键方向覆盖状态；（iv）每个标记的量化成本随着上下文被稀释，而非累积。我们还修复了一种全局尺度不匹配问题，即当针对每模块校准的 NVFP4 检查点由将这些模块融合为一个 GEMM 的内核提供服务时出现的错误，并显示校准的 FP8 KV 缓存尺度对性能无影响。结果：一个实用方案——量化所有内容，提供 KV 缩放——以及一个机制解释，说明混合 LLM 的循环部分为何是容易量化的一半。检查点：此 https URL

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04098 (Request failed with status code 502)

Authors: Sergii Kozyrev,Davyd Maiboroda

Categories:

PDF URL: https://arxiv.org/pdf/2609.04098.pdf

Arxiv URL: https://arxiv.org/abs/2609.04098

Arxiv ID: 2609.04098

CoolPaper URL: https://papers.cool/arxiv/2609.04098

Published: 2026-09-07T01:18:27.016Z

Updated: 2026-09-07T01:18:27.016Z

10. Puffin-World: Scaling a Unified Multimodal Model with Native 3D World States

Abstract:We propose Puffin-World, a unified multimodal architecture that integrates physical understanding, spatial simulation, and 3D world generation and reconstruction without relying on external offline modules. To reliably construct and interact with 3D worlds, our framework jointly models three native world states: physics (gravity field and latitude), geometry (depth), and appearance (image), together with a unified Omni-Camera representation that supports diverse tasks and flexible motions. Beyond modeling these states, we introduce a strategy for propagating physical dynamics across future frames. By grounding absolute camera properties in the real world, Puffin-World enables physically consistent and visually stable world generation. We further couple appearance and geometry within a single generative process, jointly synthesizing each future view and reconstructing its underlying geometry. This unified paradigm enables interleaved closed-loop applications requiring synergy across multiple tasks, including mimic and self-calibrated world exploration. To scale Puffin-World to complex scenarios, we construct Puffin-16M, comprising 15 million vision-language-camera triplets and 1 million trajectories featuring various and challenging motions. To foster further research in this area, we released the code, models, and datasets.

中文摘要

摘要：我们提出了Puffin-World，一种统一的多模态架构，它集成了物理理解、空间模拟以及3D世界的生成与重建，无需依赖外部离线模块。为了可靠地构建和交互3D世界，我们的框架联合建模了三种原生世界状态：物理（重力场和纬度）、几何（深度）和外观（图像），并结合了支持多样任务和灵活动作的统一全景相机（Omni-Camera）表示。除了建模这些状态外，我们还提出了一种跨未来帧传播物理动态的策略。通过将绝对相机属性与现实世界对齐，Puffin-World能够实现物理一致且视觉稳定的世界生成。我们进一步在单一生成过程中将外观和几何结合起来，同时合成每个未来视图并重建其底层几何结构。这一统一范式支持需要多任务协同的交错闭环应用，包括模仿和自校准的世界探索。为了将Puffin-World扩展到复杂场景，我们构建了Puffin-16M，该数据集包含1500万个视觉-语言-相机三元组和100万条包含各种复杂运动的轨迹。为了促进该领域的进一步研究，我们公开了代码、模型和数据集。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04196 (Request failed with status code 502)

Authors: Kang Liao,Yihang Luo,Xiao-Ming Wu,Linyi Jin,Size Wu,Chunyu Lin,Yao Zhao,Fei Wang,Wei Li,Chen Change Loy

Categories:

PDF URL: https://arxiv.org/pdf/2609.04196.pdf

Arxiv URL: https://arxiv.org/abs/2609.04196

Arxiv ID: 2609.04196

CoolPaper URL: https://papers.cool/arxiv/2609.04196

Published: 2026-09-07T01:18:31.082Z

Updated: 2026-09-07T01:18:31.082Z

11. Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction

Abstract:Online 3D reconstruction models perform poorly on long videos. This happens because regressing poses relative to a fixed first-frame anchor forces extrapolation far beyond the training distribution. Small drifts accumulate and amplify into significant geometric collapse. However, we observe that per-frame depth remains stable throughout this failure. The backbone’s local geometry remains intact; only the global pose head breaks down. Motivated by this decoupling, we introduce Scal3R. This approach reformulates online reconstruction as multi-reference relative pose querying. We use lightweight learnable tokens, which make up about ~1% of the parameters, and inject them into a completely frozen backbone via asymmetric attention. This setup queries poses relative to multiple past keyframes. An online pose-graph optimization system with loop closure suppresses long-range drift. Scal3R reaches convergence in 8 hours on a single GPU. It reduces the average ATE by over 60% on KITTI compared to the online baseline. It also achieves state-of-the-art performance across Virtual KITTI, Sintel, TUM-Dynamic, ScanNet, and 7-Scenes. Project page: this https URL

中文摘要

摘要：在线 3D 重建模型在长视频上表现不佳。这是因为相对于固定的第一帧锚点回归位姿，会导致远超训练分布的外推。小的漂移会累积并放大，最终导致显著的几何崩塌。然而，我们观察到在此失败过程中，每帧的深度仍然保持稳定。主干网络的局部几何保持完整，只有全局位姿头部出现问题。受到这种解耦的启发，我们提出了 Scal3R。该方法将在线重建重新表述为多参考相对位姿查询。我们使用轻量级可学习的 token，占参数的约 1%，并通过非对称注意机制将其注入完全冻结的主干网络。这种设置能够查询相对于多个过去关键帧的位姿。带有回路闭合的在线位姿图优化系统抑制了长距离漂移。Scal3R 在单 GPU 上 8 小时内即可收敛。与在线基线相比，它在 KITTI 上将平均绝对位姿误差 (ATE) 降低了超过 60%。它还在 Virtual KITTI、Sintel、TUM-Dynamic、ScanNet 和 7-Scenes 上实现了最先进的性能。项目页面：此 https URL

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04201 (Request failed with status code 502)

Authors: Chin-Yang Lin,Yang-Che Sun,Cheng Sun,Fu-En Yang,Min-Hung Chen,Yen-Yu Lin,Wei-Chen Chiu,Yu-Lun Liu

Categories:

PDF URL: https://arxiv.org/pdf/2609.04201.pdf

Arxiv URL: https://arxiv.org/abs/2609.04201

Arxiv ID: 2609.04201

CoolPaper URL: https://papers.cool/arxiv/2609.04201

Published: 2026-09-07T01:18:45.295Z

Updated: 2026-09-07T01:18:45.295Z

12. Editable Visual Design

Abstract:While diffusion base models such as GPT-Image-2 and Nano-Banana exhibit remarkable visual expressiveness, their end-to-end generation inherently yields flattened bitmaps with error-prone text, precluding layer-wise post-editing. Conversely, code-based visual generation via Coding Agents provides precise layout control and decoupled layers, yet remains constrained by a lack of global aesthetic intuition and the difficulty of coding complex visual assets. To address this, we propose Editable Visual Design, a new paradigm driven by a Coding Agent. We designate the VLM as the creative brain'' for requirement comprehension, task planning, and aesthetic judgment, while utilizing the image generation model as an on-demandvisual world simulator’’ to synthesize standalone visual assets. Operating under an ``imagine first, then act’’ closed-loop workflow, the agent generates isolated assets, writes native HTML/CSS, and iteratively refines the design against visual rendering feedback. Furthermore, Agent Design Replay faithfully reproduces the creative and reasoning trajectory akin to that of professional human designers. Ultimately, the system delivers editable artifacts with decoupled layers and real text, enabling users to perform intuitive mouse dragging and layout adjustments on a graphical user interface. Validations on posters, infographics, and other scenarios show that this paradigm successfully achieves both refined aesthetics and production-grade editability.

中文摘要

摘要：尽管扩散基础模型如 GPT-Image-2 和 Nano-Banana 展现了出色的视觉表现力，其端到端生成本质上会产生扁平化的位图，并伴随易出错的文本，从而无法进行按层后期编辑。相反，通过编码代理进行的基于代码的视觉生成可以实现精确的布局控制和解耦层，但仍受限于缺乏整体美学直觉以及编码复杂视觉资产的难度。为了解决这一问题，我们提出了可编辑视觉设计（Editable Visual Design），这是由编码代理驱动的新范式。我们将视觉语言模型（VLM）指定为“创意大脑”，用于需求理解、任务规划和美学判断，同时将图像生成模型作为按需的“视觉世界模拟器”来合成独立的视觉资产。在“先想象，后行动”的闭环工作流程下，代理生成独立资产、编写原生 HTML/CSS，并根据视觉渲染反馈迭代优化设计。此外，代理设计回放（Agent Design Replay）能够忠实再现类似专业人类设计师的创意和推理轨迹。最终，系统提供解耦层和真实文本的可编辑成果，使用户能够在图形用户界面上直观地进行鼠标拖拽和布局调整。在海报、信息图表及其他场景中的验证表明，该范式能够成功实现精细的美学效果和生产级的可编辑性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04034 (Request failed with status code 502)

Authors: Junyan Ye,Wei Liu,Dongzhi Jiang,Zichen Wen,HaoDong Li,Zhutao Lv,Jiaxin Lin,Jinhua Yu,Jun He,Zilong Huang,Rui Chen,Weijia Li

Categories:

PDF URL: https://arxiv.org/pdf/2609.04034.pdf

Arxiv URL: https://arxiv.org/abs/2609.04034

Arxiv ID: 2609.04034

CoolPaper URL: https://papers.cool/arxiv/2609.04034

Published: 2026-09-07T01:18:46.313Z

Updated: 2026-09-07T01:18:46.313Z

Abstract:Joint audio-video generation models have made substantial progress in visual quality and audio-visual synchronization. However, they still provide limited control over when shot transitions occur and dialogue is spoken. This limitation constrains their application in script-driven content creation, where timing errors can undermine narrative coherence and the viewing experience. Current joint generators align video and audio representations on a shared temporal axis, yet the precise timing of shots and dialogue specified in a structured prompt is encoded only in the prompt’s text representation and remains unaligned with the temporal coordinates of either modality. Consequently, video and audio may remain synchronized with each other while both fail to follow the script timeline. This mismatch motivates us to extend temporal alignment beyond video and audio to include the structured script. We therefore introduce Temporal Context Routing (TCR), which maps the script timing onto the shared temporal axis of video and audio generation and routes each prompt’s guidance to the corresponding positions in both modalities. Compared with the baseline on 200 test scripts, TCR reduces Shot Boundary MAE by 96%, from 1.11 s to 0.042 s, and raises Dialogue Acc@0.5 s from 28.3% to 84.1%. TCR achieves these improvements while maintaining visual quality and audio-visual synchronization comparable to those of the baselines. A user study further shows that participants prefer TCR on all five evaluated dimensions.

中文摘要

摘要：联合音视频生成模型在视觉质量和视听同步方面取得了显著进展。然而，它们在控制镜头切换发生时间和对白播放时间方面仍然有限。这种限制限制了它们在剧本驱动内容创作中的应用，因为时间上的错误可能破坏叙事的连贯性和观看体验。当前的联合生成器在共享时间轴上对齐视频和音频表示，但结构化提示中指定的精确镜头和对白时间仅编码在提示的文本表示中，并未与任何一种模态的时间坐标对齐。因此，视频和音频可能互相同步，但都未能遵循剧本时间线。这种不匹配促使我们将时间对齐扩展到视频和音频之外，包括结构化剧本。因此，我们引入了时间上下文路由（Temporal Context Routing, TCR），它将剧本时间映射到视频和音频生成的共享时间轴，并将每个提示的指导路由到两种模态中的对应位置。与基线在200个测试剧本上的表现相比，TCR将镜头边界平均绝对误差（MAE）减少了96%，从1.11秒降至0.042秒，并将对白准确率@0.5秒从28.3%提高到84.1%。TCR在取得这些改进的同时，仍保持与基线相当的视觉质量和视听同步性。用户研究进一步显示，参与者在所有五个评估维度上更偏好TCR。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.02367 (Request failed with status code 502)

Authors: Yichen Liu,Quanwei Zhang,Haozhe Wang,Donghao Zhou,Xiaojie Li,Yang Shi,Jiaming Liu,Ruihua Huang,Yingtian Zou,Daquan Zhou

Categories:

PDF URL: https://arxiv.org/pdf/2609.02367.pdf

Arxiv URL: https://arxiv.org/abs/2609.02367

Arxiv ID: 2609.02367

CoolPaper URL: https://papers.cool/arxiv/2609.02367

Published: 2026-09-07T01:18:47.475Z

Updated: 2026-09-07T01:18:47.475Z

14. Beyond Retrieval: Progressive Latent Memory Evolution for Streaming Video Understanding

Abstract:Streaming video understanding requires multimodal large language models (MLLMs) to process continuous visual inputs and respond to user queries under strict causality and bounded memory. Existing approaches typically compress historical observations into an external memory bank and retrieve query-relevant evidence as additional visual context. Though effective, this store-and-retrieve paradigm keeps historical evidence as external visual context, preventing it from being internalized into a compact, evolving latent memory that can continuously guide streaming reasoning. To bridge this gap, we introduce LatentStream, a progressive latent working memory framework that shifts streaming memory from store-and-retrieve to retrieve-and-internalize. Specifically, LatentStream comprises three coordinated components. First, Query-agnostic Hierarchical Streaming Memory organizes visual history into short-, mid-, and long-term levels under a fixed memory budget through Jenks-guided adaptive consolidation. Once a query arrives, Hierarchical Latent Memory Evolution equips groups of latent memory tokens with progressively expanding memory receptive fields, enabling them to iteratively retrieve historical evidence from their corresponding scopes and internalize it into a compact, fixed-length latent memory. Finally, Progressive Confidence-guided Latent Memory Optimization constructs a hierarchical progression reward from group-wise predictive entropy and jointly refines the latent memory tokens and retrieved evidence, encouraging increasingly confident streaming reasoning. Extensive experiments demonstrate that LatentStream achieves new state-of-the-art results on existing online and offline video benchmarks.

中文摘要

摘要：流视频理解要求多模态大语言模型（MLLMs）处理连续的视觉输入，并在严格的因果性和有限的记忆条件下响应用户查询。现有方法通常将历史观察压缩到外部记忆库中，并检索与查询相关的证据作为额外的视觉上下文。虽然有效，但这种存储-检索范式将历史证据保持为外部视觉上下文，阻止其被内化为紧凑、可演化的潜在记忆，从而持续指导流式推理。为弥合这一差距，我们提出了LatentStream，一种渐进式潜在工作记忆框架，将流式记忆从存储-检索转变为检索-内化。具体而言，LatentStream包含三个协调组件。首先，无查询层次流式记忆通过Jenks引导的自适应整合，将视觉历史按短期、中期和长期层次组织在固定内存预算下。一旦查询到来，层次潜在记忆演化为潜在记忆令牌组配备逐步扩展的记忆感受野，使它们能够从对应范围迭代地检索历史证据并将其内化为紧凑的固定长度潜在记忆。最后，渐进式置信引导的潜在记忆优化通过组级预测熵构建层次进展奖励，并联合优化潜在记忆令牌和检索到的证据，促进日益自信的流式推理。大量实验表明，LatentStream在现有在线和离线视频基准上实现了新的最先进结果。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04131 (Request failed with status code 502)

Authors: Hongyu Qu,Guangming Yao,Ling Xing,Xiaobin Hu,Rongxing Ding,Guibin Zhang,Fan Zhang,Yi Yuan,Xiangbo Shu,Shuicheng Yan

Categories:

PDF URL: https://arxiv.org/pdf/2609.04131.pdf

Arxiv URL: https://arxiv.org/abs/2609.04131

Arxiv ID: 2609.04131

CoolPaper URL: https://papers.cool/arxiv/2609.04131

Published: 2026-09-07T01:18:49.329Z

Updated: 2026-09-07T01:18:49.329Z

15. Last Translation Benchmark

Abstract:For scientific progress, we need benchmarks that test the limits of state-of-the-art models, and evaluation methods that inform us about failure cases. As models get stronger, standard benchmarks for machine translation are approaching saturation. Further, automatic translation metrics are unreliable, vulnerable to reward-hacking, and provide unactionable assessments. Even gold human evaluation is not problem-free, because it often lacks reproducibility, objectivity, and scalability. Overall, this prevents us from tracking objective progress in the field and identifying pathways for improvement. We introduce the Last Translation Benchmark, a collection of human-authored and peer-reviewed examples (texts, images, audio, videos) that break leading machine translation models. We also present a new evaluation approach: each example comes with handcrafted verification rules describing concrete failure cases on that example, therefore allowing reliable and actionable future evaluation. The Last Translation Benchmark is a live dataset that accepts ongoing contributions. The latest version is LTBv1, containing accepted contributions prior to September 1st 2026, with future releases planned as new data is continuously collected.

中文摘要

摘要：为了科学进步，我们需要能够测试最先进模型极限的基准，并能够告知我们失败案例的评估方法。随着模型变得更强大，机器翻译的标准基准正在接近饱和。此外，自动翻译指标不可靠，容易被奖励操纵，并且提供的评估无法付诸行动。即使是黄金级人工评估也并非没有问题，因为它往往缺乏可重复性、客观性和可扩展性。总体而言，这阻碍了我们追踪该领域的客观进展并识别改进途径。我们引入了“最后翻译基准”（Last Translation Benchmark），这是一个由人工创作并经过同行评审的示例集合（文本、图像、音频、视频），旨在挑战领先的机器翻译模型。我们还提出了一种新的评估方法：每个示例都附有手工制作的验证规则，描述该示例的具体失败案例，从而实现可靠且可操作的未来评估。“最后翻译基准”是一个实时数据集，接受持续的贡献。最新版本为 LTBv1，包含截至2026年9月1日已接受的贡献，未来计划随着新数据的不断收集而发布新版本。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04173 (Request failed with status code 502)

Authors: Vilém Zouhar,Niyati Bafna,Mukund Choudhary,Maike Züfle,Sara Rajaee,Pinzhen Chen,Jannis Vamvas,Sara Papi,Ona de Gibert,Bhavitvya Malik,Eliya Habba,Orfeas Menis Mastromichalakis,Patrícia Schmidtová,Michelle Wastl,Sheriff Issaka,Leshem Choshen,Stella Biderman,Antonis Anastasopoulos,Jan Niehues,Rico Sennrich,Mrinmaya Sachan,Ondřej Bojar,Kenton Murray,Jörg Tiedemann,Alham Fikri Aji,Philipp Koehn,Christof Monz,Alexandra Birch,Sowmya Vajjala,Chalamalasetti Kranti,Cristina España-Bonet,Nobin Sarwar,David Kaczér,Shunta Asano,Malik Marmonier,Daban Q. Jaff,Vaisakhi Mishra,Hend Al- Khalifa,Gabriele Sarti,Sourajit Saha,Nils Rehlinger,Juan Daniel Cuervo Villa,Jonathan Tonglet,Saugata Purkayastha,Dominik Macháček,Jagannathan Ramanujam,Heejin Do,Zuzana Nadova,Fred Philippy,Fabian Retkowski,Maria Lymperaiou,Silvia Casola,Hanna Yukhymenko,Shubhashis Roy Dipta,Sangwon Ryu,Andrés Jerez,Ron Keinan,Shuaib Shuaib Yusuf,Avantica Vempati,Maria Carmen Staiano,Sukannya Purkayastha,Adrian Cosma,Vitalii Babenko,Erivan Inan,Aviral Nigam,Wafa Aissa,Fatima Haouari,Venkata Prasanth Kumar Gummadi,Mehdi Jafarzadeh,Valentin Scourneau,Lukas Edman,Kaiser Sun,Shaomu Tan,Mohammad Sadegh Gholizadeh,Johannes-Rudolf David,Dipankar Srirag,Javier García Gilabert,Ruta Binkyte,Manar Ali,Ana-Maria Bucur,Sabry E. Farrag,Youssef Saber,Yihong Liu,Jean Maillard,Cojocaru Nicoleta,Xiaochuang Yuan,Sina Ahmadi,Philipp Mondorf,Kaustubh Dhole,Roman Wixinger,Shenbin Qian,Manuel Tuor,Sergey Troshin,Jonathan Yahav,Fida Mohammad Thoker,Amir Arsalan Rezapour,Lance Calvin Lim Gamboa,Manon Reusens,Kätriin Kukk,Koel Dutta Chowdhury,Giuseppe Gallipoli,Christian Hoang,Shaswati Saha,Seth Aycock,Jan Kocoń,Bo Chen,Linh Vu,Vatsal Venkatkrishna,Arafat Ahsan,Luan Thanh Nguyen,Hassan Soliman,Daryna Dementieva,Theresia Veronika Rampisela,Ngoc Quynh Tram Do,Marius Huber,Kazuki Egashira,Azmine Toushik Wasi,Vladislav Poritski,Mike Zhang,Deep Shah,Paul Gavrikov,Luis Frentzen Salim,David Africa,R. Damanhuri,Bello Umar Bello,Anumit Garg,Gengyu Rao,Pawan Sasanka Ammanamanchi,Kamile Dementaviciute,Andrianos Michail,L D M S Sai Teja,Dawei Zhu,Yi Fan,Wei Liu,Farhan Farsi,Elias Herranen,Sankalan Pal Chowdhury,Karen Sanchez,Farzad Shami,Ashok Urlana,Zimu Wang,Tomasz Limisiewicz,Priyaranjan Pattnayak,Marii Ojastu,Hongbin Na,Emilian Radoi,Chenyi Zhao,Carlos Hinojosa,Andrea Gregor de Varda,Zaid Alyafeai,Reem Alzahrani,Nehal Kathrotia,Alex Flückiger,Ulysses Sekai Tully Carr,Jimson Paulo Layacan,Guy Kaplan,Ritwik Tiwari,Rishit Dagli,Oksana Volchek,Isaac R Caswell,Bowen Yi,Blanka Kövér,Amir Hossein Yari,Aicha Chorana,Zhengxiang Wang,Selja Keränen,Samuel Simko,Joy Olusanya,Jenny Chim,Enzo Doyen,Vivek Harsha Lakkamaneni,Sophia Conrad,Pouya Sadeghi,Panayiotis Panayiotou,Luis Lara,Jannatul Nayem,Eran Yahav,Debanshu Das,Antonia Karamolegkou,Anmol Goel,Aishik Mandal,Tommaso Cerruti,Raoyuan Zhao,Mykola Haltiuk,Thura Aung,Naser Almousa,Amir Hossein Kargaran,Rachel Bawden,Qiaoyuan Zheng,Mateusz Lango,Beni Egressy,Fidel Rodríguez Velásquez,Natchapon Jongwiriyanurak,Minh Ngoc Do,Marco Gaido,Lena Libon,Dzmitry Kuzmin,Badal Nyalang,Antoine Taroni,Andrei Niculae,Abdulaziz Nura Kani,Rushikesh Zawar,Marek Šuppa,Beatrice Savoldi,Andreas Simons,Rayyan Merchant,Ilai Yaron Levy,Francesco Pinto,Ziyi Yang,Yolanda Xavier,Samuel Frontull,Muhammad Ravi Shulthan Habibi,Kenneth Enevoldsen,Harris Abdul Majid,Francesca Padovani,Tim Graf,Tatiana Bielakova,Sharifa Djurabaeva,Shaoxiong Ji,Raia Abu Ahmad,Pavel Stepachev,Jirui Qi,Ayush Sunil Munot,Alireza Pakniat,Ayla Rigouts Terryn,Yuxing Lu,Yurii Paniv,Xiyan Fu,Tosin Adewumi,Sunisth Kumar,Stéphane J.P.S. Thunus,Shree Harsha Bokkahalli Satish,Shayan Bali,Prakhar Gupta,Papa Abdou Karim Karou Diallo,Matija Akrap,Marko Culjak,Kristýna Onderková,Joseph Attieh,Esrael Teferi Tensay,Elisabeth Fittschen,Benoît Sagot,Jingwei Ni,Yu Fan,et al. (144 additional authors not shown)

Categories:

PDF URL: https://arxiv.org/pdf/2609.04173.pdf

Arxiv URL: https://arxiv.org/abs/2609.04173

Arxiv ID: 2609.04173

CoolPaper URL: https://papers.cool/arxiv/2609.04173

Published: 2026-09-07T01:18:53.206Z

Updated: 2026-09-07T01:18:53.206Z

16. RealSWE: A Compositional Evaluation of Coding Agents under Realistic User Requests

Abstract:Coding agents are now commonly evaluated on the SWE-bench family of benchmarks, whose tasks are built from curated GitHub issues: long, structured, and information-rich. Real user requests, however, are typically far shorter and less structured. To characterize this gap, we define a six-category information taxonomy and four dimensions of linguistic style, and apply them to real user prompts from SWE-chat and problem statements from SWE-bench Verified and Pro. We find that requests carrying only a problem statement, alone or with limited additional context, account for 88% of real prompts but just 7% of benchmark problems. Furthermore, 87% of real prompts are casually written whereas 94% of benchmark problems are formal. Guided by these observations, we introduce RealSWE, 381 multi-variant task families derived from SWE-bench Verified and Pro. Variants within each family share the same underlying task and gold patch while differing only in information composition and linguistic style. Evaluating seven contemporary LLMs with RealSWE, we find that i) realistic inputs reduce resolution rates by 6.4 pp on average and can change model rankings. Controlled analysis further shows that ii) including Desired Behavior and Motivation significantly affects performance, whereas Environment Information and Reproduction Steps merely add tokens without measurable benefit; iii) linguistic style has only small, model-dependent effects. These findings provide actionable guidance for users and agents: explicitly stating the desired behavior and motivation, which most real prompts omit, substantially improves the LLM’s software engineering performance.

中文摘要

摘要：现在，编码代理通常通过 SWE-bench 系列基准进行评估，这些任务由精心挑选的 GitHub 问题构建：任务冗长、结构化且信息丰富。然而，真实用户的请求通常要短得多且不那么结构化。为了描述这种差距，我们定义了一个六类信息分类法和四个语言风格维度，并将其应用于来自 SWE-chat 的真实用户提示以及 SWE-bench Verified 和 Pro 的问题陈述。我们发现，仅携带问题陈述的请求，无论是否附带有限的补充上下文，构成真实提示的 88%，但仅占基准问题的 7%。此外，87% 的真实提示是随意书写，而 94% 的基准问题是正式的。根据这些观察结果，我们引入了 RealSWE，这是由 SWE-bench Verified 和 Pro 衍生出的 381 个多变体任务家族。每个家族内的变体共享相同的基础任务和标准修补方案，仅在信息组成和语言风格上有所不同。使用 RealSWE 对七个当代大语言模型进行评估，我们发现：i) 真实输入平均减少 6.4 个百分点的解决率，并可能改变模型排名。受控分析进一步显示：ii) 包含期望行为和动机显著影响性能，而环境信息和复现步骤仅增加了标记而没有可测量的益处；iii) 语言风格仅对性能有小幅、依赖模型的影响。这些发现为用户和代理提供了可操作的指导：明确陈述大多数真实提示中缺失的期望行为和动机，能够显著提升大语言模型的软件工程表现。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2608.27831 (Request failed with status code 502)

Authors: Gyuhyeong Kim,Hyojung Gwon,Jeonghyeon Kim,Kyuhong Shim,Sunjae Lee

Categories:

PDF URL: https://arxiv.org/pdf/2608.27831.pdf

Arxiv URL: https://arxiv.org/abs/2608.27831

Arxiv ID: 2608.27831

CoolPaper URL: https://papers.cool/arxiv/2608.27831

Published: 2026-09-07T01:19:02.759Z

Updated: 2026-09-07T01:19:02.759Z

17. DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training

Abstract:Reinforcement Learning from Verifiable Rewards works well when a task has a programmatic checker, but most long-horizon agent domains have none. We work in the outcome-blind setting, where ground-truth success signals are not available. Multi-criteria rubrics are a popular way to supply such a reward; they are scored once per trajectory, but a single scalar is a poor signal across tens of steps. We propose DRACO: Distributing Rubric-based Advantage for Credit Optimization. It generates rubrics dynamically during training to track the policy’s evolving capability, scores those rubrics once per completed trajectory, and redistributes that judgment over the steps responsible for annotated rubrics to produce differentiated per-step advantages in GRPO. The redistribution is closed-form and does not introduce any trained attribution module. On AppWorld, DRACO gains 15.9 points over the base model and 5.3 points over GRPO trained with a sparse ground-truth reward, despite not using any verifiers itself. On out-of-domain Tau-Bench, it gains 5.3 points over the base model even without a frontier judge, beating both ground-truth-reward training and other rubric-based training settings. The code for DRACO is available at this https URL.

中文摘要

摘要：当任务有可编程的检查器时，从可验证奖励中进行强化学习效果良好，但大多数长时间跨度的智能体任务领域并没有这样的方法。我们在“结果不可见”的环境下工作，即无法获得真实的成功信号。多标准评分标准是一种提供此类奖励的流行方法；它们每条轨迹只评分一次，但单一的标量信号在数十步中是很差的指导。我们提出了 DRACO：基于评分标准优势的信用优化分配。它在训练过程中动态生成评分标准以跟踪策略不断变化的能力，每次完成轨迹后对这些评分标准进行评分，并将该评分重新分配到负责评分标准注释的步骤上，以在 GRPO 中产生区分化的每步优势。该重新分配是封闭形式的，并且不引入任何训练的归因模块。在 AppWorld 上，DRACO 比基础模型提升了 15.9 分，比使用稀疏真实奖励训练的 GRPO 提升了 5.3 分，尽管它自身不使用任何验证器。在异域 Tau-Bench 上，即使没有前沿评判器，DRACO 仍比基础模型提升了 5.3 分，超越了真实奖励训练和其他基于评分标准的训练设置。DRACO 的代码可在此 https URL 获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04094 (Request failed with status code 502)

Authors: Shubham Gandhi,Saurabh Goyal,Kiran Kate,Yara Rizk

Categories:

PDF URL: https://arxiv.org/pdf/2609.04094.pdf

Arxiv URL: https://arxiv.org/abs/2609.04094

Arxiv ID: 2609.04094

CoolPaper URL: https://papers.cool/arxiv/2609.04094

Published: 2026-09-07T01:19:04.824Z

Updated: 2026-09-07T01:19:04.824Z

18. CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation

Abstract:MLLM-based embedding models remain limited in compositional retrieval, often failing to distinguish scenes containing the same concepts but different attribute-object bindings. Yet the same backbone can resolve such distinctions when used as a cross-attentive reranker, motivating us to distill its compositional judgments into the embedding model. We propose CORE, which synthesizes candidate lists spanning five compositional matching levels and introduces a Rank-KL objective that trains the embedding model to reproduce the reranker’s fine-grained ranking. We further introduce a graded evaluation protocol and compare contrastive learning, pairwise CoSENT, and listwise Rank-KL under the same data and tuning budget. Our comparison shows that both CoSENT and Rank-KL use the multi-level supervision more effectively than contrastive learning, with Rank-KL achieving the strongest overall performance. Across three compositional reasoning benchmarks (COLA, SUGARCREPE++, NEGBENCH), CORE-RERANKER-8B achieves an 82.7% total average, outperforming Jina-Reranker by 10.7 points, while CORE-EMBED-8B achieves the best total average (0.666) among all evaluated embedding models. The improvements transfer to the MCMR benchmark without sacrificing retrieval performance on COCO and Flickr30K.

中文摘要

摘要：基于 MLLM 的嵌入模型在组合检索方面仍然有限，通常无法区分包含相同概念但属性-对象绑定不同的场景。然而，当相同的骨干网络用作交叉注意重排序器时，它能够解决这些区分问题，这促使我们将其组合判断蒸馏到嵌入模型中。我们提出了 CORE，它合成了跨五个组合匹配水平的候选列表，并引入了 Rank-KL 目标，该目标训练嵌入模型以再现重排序器的细粒度排序。我们进一步引入了分级评估协议，并在相同的数据和调参预算下比较了对比学习、成对 CoSENT 以及列表式 Rank-KL。我们的比较显示，CoSENT 和 Rank-KL 都比对比学习更有效地利用了多级监督，其中 Rank-KL 实现了最强的整体性能。在三个组合推理基准（COLA、SUGARCREPE++、NEGBENCH）中，CORE-RERANKER-8B 实现了 82.7% 的总平均分，比 Jina-Reranker 高出 10.7 个百分点，而 CORE-EMBED-8B 在所有评估的嵌入模型中实现了最佳总平均分（0.666）。这些改进也转移到 MCMR 基准，同时在 COCO 和 Flickr30K 上的检索性能没有下降。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.04083 (Request failed with status code 502)

Authors: Tingyu Song,Mingxin Li,Yanzhao Zhang,Dingkun Long,Chu Liu,Pengjun Xie,Yilun Zhao,Shu Wu

Categories:

PDF URL: https://arxiv.org/pdf/2609.04083.pdf

Arxiv URL: https://arxiv.org/abs/2609.04083

Arxiv ID: 2609.04083

CoolPaper URL: https://papers.cool/arxiv/2609.04083

Published: 2026-09-07T01:19:05.571Z

Updated: 2026-09-07T01:19:05.571Z

19. WorldReward: Reward Modeling for Camera-Conditioned World Models

Abstract:Camera-conditioned world models generate interactive videos in which commanded actions should induce the expected scene changes while appearance, geometry, and temporal dynamics remain coherent. Existing rewards assess these requirements separately: geometry-based rewards estimate trajectory execution but cannot judge the visual quality of the executed motion, whereas image-based rewards measure frame quality without capturing action execution or temporal dynamics. We posit that a vision-language model (VLM) offers a shared reasoning space for relating actions to their visual outcomes. However, judging a complete long video against its full action sequence creates a lengthy, noisy context in which short-lived local action evidence can be missed or diluted. We present WorldReward, a VLM-based pairwise preference reward model that unifies action-consistency and visual-quality evaluation for camera-conditioned world models. WorldReward decomposes paired videos into action-aligned chunks, organizes each chunk into structured visual evidence, and aggregates chunk-level decisions by voting into separate video-level action and visual-quality preferences. To train it, we construct a large-scale reasoning-augmented preference dataset using structured judgments generated by a frontier VLM and refined through tool-based agent auditing and targeted human review. We further introduce WorldReward-Bench, a human-annotated benchmark measuring reward-model agreement with human preferences across action consistency, appearance quality, and motion quality. WorldReward achieves the highest agreement on all three dimensions, exceeding GPT-5.5 by 3.42, 1.45, and 3.56 percentage points, respectively. When used for RL post-training of HY-WorldPlay 1.5, it consistently improves both action execution and visual quality across short- to long-term horizons.

中文摘要

摘要：基于摄像机条件的世界模型生成交互式视频，其中指令动作应引起预期的场景变化，同时外观、几何和时间动态保持一致。现有奖励分别评估这些要求：基于几何的奖励估计轨迹执行情况，但无法判断执行动作的视觉质量，而基于图像的奖励衡量帧质量，却无法捕捉动作执行或时间动态。我们假设视觉-语言模型（VLM）提供了一个共享的推理空间，可将动作与其视觉结果关联。然而，将完整的长视频与其完整动作序列进行评判，会产生冗长、噪声较大的上下文，其中短暂的局部动作证据可能被忽略或淡化。我们提出了 WorldReward，一种基于 VLM 的成对偏好奖励模型，统一了基于摄像机条件的世界模型的动作一致性与视觉质量评估。WorldReward 将成对视频分解为动作对齐的片段，将每个片段组织为结构化视觉证据，并通过投票聚合片段级决策为单独的视频级动作和视觉质量偏好。为了训练它，我们构建了一个大规模的推理增强偏好数据集，该数据集使用前沿 VLM 生成的结构化判断，并通过基于工具的智能体审核和针对性人工审查进行优化。我们进一步引入 WorldReward-Bench，这是一种人工标注的基准，用于衡量奖励模型在动作一致性、外观质量和运动质量方面与人类偏好的契合度。WorldReward 在所有三个维度上达到了最高一致性，分别比 GPT-5.5 高出 3.42、1.45 和 3.56 个百分点。在用于 HY-WorldPlay 1.5 的 RL 后训练时，它在短期到长期的时间范围内持续提升动作执行和视觉质量。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.03952 (Request failed with status code 502)

Authors: Yibin Wang,Zehan Wang,Junshu Tang,Zhimin Li,Yujie Zhou,Jiazi Bu,Pengyang Ling,Feng Han,Zhixiong Zhang,Long Xing,Shengyuan Ding,Ziang Li,Cheng Jin,Yuhang Zang,Jiaqi Wang,Tianyu Pang

Categories:

PDF URL: https://arxiv.org/pdf/2609.03952.pdf

Arxiv URL: https://arxiv.org/abs/2609.03952

Arxiv ID: 2609.03952

CoolPaper URL: https://papers.cool/arxiv/2609.03952

Published: 2026-09-07T01:19:09.058Z

Updated: 2026-09-07T01:19:09.058Z

20. PACE: Towards Surfacing Hidden Conflicts in User Requests

Abstract:Personalized assistants should not only comply with user requests but also assess whether those requests are appropriate given the user’s current circumstances. However, prior work has primarily focused on accurately executing requests, overlooking the need for assistants to account for context and engage in conflict-based refusal. Furthermore, while existing work on conflict or safety detection relies on explicitly provided factors, real-world scenarios often involve implicit factors that must be retrieved from a knowledge base (KB). To this end, we introduce Personalized Assistants for Conflict Evaluation (PACE), a dataset for evaluating whether models can identify latent constraints, expressed as egocentric knowledge or events, that render seemingly reasonable user requests inappropriate. PACE pairs user requests grounded in well-defined personas with egocentric KB facts, requiring models to integrate contextual evidence to determine whether a request is conflicting. This implicit retrieval setting hinders the direct association between user requests and conflict-inducing knowledge, making it difficult for existing models to identify relevant user-specific facts. To address this challenge, we further propose PaceMaker, a multi-agent framework in which specialized agents coordinate across query reformulation, multi-hop graph traversal, and conflict-aware filtering to retrieve contextually decisive evidence. Experiments on PACE evaluate both evidence retrieval quality and conflict decision accuracy, showing that PaceMaker consistently outperforms existing approaches.

中文摘要

摘要：个性化助手不仅应遵循用户的请求，还应评估这些请求在用户当前情况下是否合适。然而，以往的工作主要集中在准确执行请求，忽略了助手需要考虑上下文并进行基于冲突的拒绝的需求。此外，尽管现有关于冲突或安全检测的工作依赖于明确提供的因素，但现实场景中往往涉及必须从知识库（KB）中获取的隐性因素。为此，我们提出了用于冲突评估的个性化助手（PACE），这是一个用于评估模型是否能够识别潜在约束的数据集，这些约束以自我中心的知识或事件形式表达，使看似合理的用户请求变得不适当。PACE 将基于明确定义的角色的人类用户请求与自我中心的 KB 事实配对，要求模型整合上下文证据以判断请求是否存在冲突。这种隐性检索环境阻碍了用户请求与引发冲突的知识之间的直接关联，使现有模型难以识别相关的用户特定事实。为解决这一挑战，我们进一步提出了 PaceMaker，这是一个多智能体框架，其中专门的智能体在查询重构、多跳图遍历和冲突感知过滤方面进行协调，以检索上下文上决定性的证据。在 PACE 上的实验评估了证据检索质量和冲突决策准确性，结果显示 PaceMaker 始终优于现有方法。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.03293 (Request failed with status code 502)

Authors: Yoojin Kim,Jihyoung Jang,Hyounghun Kim

Categories:

PDF URL: https://arxiv.org/pdf/2609.03293.pdf

Arxiv URL: https://arxiv.org/abs/2609.03293

Arxiv ID: 2609.03293

CoolPaper URL: https://papers.cool/arxiv/2609.03293

Published: 2026-09-07T01:19:11.053Z

Updated: 2026-09-07T01:19:11.053Z