👍 90
09/26 08:00
Reinforcement learning (RL) for code agents often uses executable tests to provide binary rewards. With these rewards, Group Relative Policy Optimization (GRPO) assigns identical advantages to test-passing trajectories within each rollout group, overlooking differences in implementation quality and
中文介绍 针对代码 agent 的 RL 训练中,GRPO 仅依据可执行测试的二元奖励,给同一 rollout 组内通过测试的轨迹分配相同优势,忽略实现质量差异。论文提出组内 agentic 评分与优势重分配,按实现质量对轨迹区分并重新分配优势,缓解二元奖励的不公平,提升代码 agent 的 RL 训练效果。
👍 55
09/28 08:00
Most modern multimodal large language models (MLLMs) build on a pretrained visual encoder that provides a strong visual prior. Encoder-free MLLMs instead learn visual representations directly from raw pixels, offering a simple and unified architecture, but their scaling behavior has not been systema
中文介绍 多数多模态大模型依赖预训练视觉编码器提供视觉先验,无编码器 MLLM 直接从像素学习视觉表示,架构统一但 scaling 行为缺乏系统研究。本文建立无编码器多模态预训练的 scaling law,对比编码器方案在数据与模型规模上的表现差距,明确移除视觉编码器的可行边界与代价。
👍 29
09/27 08:00
A central goal in robotics is to enable manipulation across changing tasks and environments. Vision-language-action (VLA) models provide broad manipulation capabilities but can struggle when execution requires diagnosing failures and adapting behavior. Strong agents can discover effective interventi
中文介绍 VLA 模型具备广泛操作能力,但在需要失败诊断与行为适应的任务上表现不佳,而强 agent 能发现有效干预策略。论文提出跨 agent 的递归 harness 蒸馏,把强 agent 的干预与适应能力迁移给 VLA 策略,在变化任务与环境中改善机器人操作的鲁棒性与自适应能力。
👍 27
09/27 08:00
Large language model (LLM) agents increasingly undertake extreme-long (xlong) horizon tasks, where a single execution can span hours, hundreds of model--environment interactions, and nearly 1M tokens per rollout. Applying online reinforcement learning (RL) to such executions poses two fundamental ch
中文介绍 超长程 agent 任务单次执行可跨越数小时、数百次模型-环境交互、接近 1M token,在线 RL 面临两大挑战。论文提出 QwenGyre 弹性 RL 框架,针对超长程训练的采样与资源调度做弹性设计,支撑大规模长程 agent 的在线强化学习训练。
👍 25
09/28 08:00
Vision-language (VL) pretraining using paired chest X-ray (CXR) images and radiology reports has shown strong potential for medical image understanding. However, existing methods often remain dependent on task-specific finetuning because radiology reports are lengthy, clinically dense, and difficult
中文介绍 胸片图像-报告视觉语言预训练有潜力,但报告冗长、临床信息密集,现有方法多依赖任务微调,难以零样本迁移。论文提出以句子为中心的 SentZero 预训练框架,增强细粒度文本监督,在多项胸片分析任务上实现零样本表现提升。
👍 17
09/26 08:00
Fast matrix multiplication algorithms keep the product fixed and search for a cheaper way to evaluate it. We instead ask whether a Transformer's learned projections can use a different, cheaper product altogether. Building on an associative-algebra construction that replaces ordinary matrix multipli
中文介绍 快速矩阵乘法在固定乘积下寻找更省算力的算法,本文反向提问:Transformer 学到的投影能否直接使用更廉价的产品。基于结合代数构造替换普通矩阵乘法,提出结合代数层,在保持参数不变的前提下改变乘积运算,降低计算开销。
👍 16
09/27 08:00
Coding-agent evaluation has progressed from resolving individual issues to carrying out long-horizon development, yet task completion is still largely assessed within a single codebase. In software ecosystems, many features and bug fixes require coordinated changes across multiple repositories. We i
中文介绍 编码 agent 评测已从解决单个 issue 走向长程开发,但任务完成仍多在单一代码库内评估;而软件生态中许多功能与修复需要跨仓库协同改动。论文提出 WideSWE 基准,考察编码 agent 跨多个仓库协调变更的能力,揭示现有方法在跨库任务上的不足。
👍 15
09/28 08:00
Frontier general-purpose systems are rapidly expanding beyond visual understanding into capabilities traditionally handled by dedicated computer-vision models. As these capabilities expand, a central question for the computer-vision community is how far this reach extends, and what remains hard. We
中文介绍 通用前沿系统正从视觉理解扩展到传统专用视觉模型的能力范围。论文系统评测 GPT-6 Astra 在多种计算机视觉任务上的表现,区分「难视觉」与「易视觉」任务,分析通用模型能替代哪些专用模型、哪些任务仍然困难,为视觉社区提供能力边界参考。
👍 15
09/28 08:00
Tool-based image editing (image retouching) is commonly formulated with autoregressive multimodal large language models (MLLMs) that sequentially generate reasoning, tool selections, and parameter values. In this work, we present a novel approach to tool-based image editing by framing the task as a
中文介绍 基于工具的图像修图通常用自回归 MLLM 依次生成推理、工具选择与参数值,参数控制精度受限。论文把该任务重构为 flow matching 问题,提出 FlowTool 直接生成工具参数,实现更精细的修图参数控制与更优的编辑效果。
👍 11
09/28 08:00
As language models take a growing role in AI development, a natural aspiration is for them to reflect on their own learning process, as humans do, and use that reflection to improve themselves. At the same time, these models have an advantage that human learners lack, since training leaves parameter
中文介绍 语言模型训练会在参数中留下更新痕迹,为模型反思自身学习过程提供可能。论文提出 Imprint Reader,从权重更新中读出训练信息,并进一步用其进行行为干预,将权重层面的学习信号转化为可影响模型行为的机制,连接内部表征读取与外部行为控制。
👍 10
09/28 08:00
We study on-policy distillation (OPD) through the lens of reinforcement learning, establishing a connection between the reverse-KL objective in OPD and KL-regularized policy optimization. Building on this connection, we introduce Least-Square Policy Distillation (LSPD), an RL-inspired framework that
中文介绍 论文从强化学习视角重新审视 on-policy distillation(OPD),建立 OPD 中 reverse-KL 目标与 KL 正则化策略优化的联系。据此提出 Least-Square Policy Distillation(LSPD)框架,将蒸馏转化为最小二乘形式的策略学习,提升 LLM 推理训练的样本效率。
👍 9
09/28 08:00
Reinforcement learning (RL) post-training for large language models (LLMs) coordinates multiple models across generation, inference, and training on GPU clusters. Several factors may change during a run, including resource availability, sequence length, memory pressure, and stage bottlenecks. As a c
中文介绍 LLM 的 RL 后训练需在 GPU 集群上协调生成、推理与训练多个模型,运行中资源可用性、序列长度、显存压力与阶段瓶颈都会变化。论文提出 Nereus,面向后训练的自适应并行方案,根据运行时状况动态调整并行策略,缓解瓶颈并提升集群利用率。
👍 8
09/26 08:00
Learning from experience in LLM agents has become a key paradigm for developing self-evolving agents that continuously learn and expand their capabilities. Within this paradigm, synthesizing the agent skill has emerged as a promising solution for transforming accumulated experience into reusable pro
中文介绍 从经验中学习是构建自演化 LLM agent 的关键范式,核心难点在于把累积经验转化为可复用技能。论文提出 ExpVoyager,通过直接经验导航实现动态技能合成,让 agent 在持续学习与能力扩展中更有效地组织并复用历史经验,提升自演化能力。
👍 5
09/26 08:00
Multiple agents may often conflict in an organization: for example, one coding agent changes an interface in a repository, but another continues to develop on the old version where existing tests become stale. A conversation can resolve the episode, but when the participants change, what makes the l
中文介绍 多 agent 组织常出现冲突,例如一个编码 agent 改了仓库接口,另一个仍在旧版本上开发导致测试失效;对话可解决当下问题,但成员更换后能力难以留存。论文提出 Relic,把多 agent 协作沉淀为持久组织能力,使经验与约定不随参与者更替而丢失。
👍 4
09/28 08:00
Multi-vector retrievers built on vision-language models lead visual document retrieval (VDR), but they run a multi-billion-parameter query encoder on every search. Distilling this encoder into a small student that queries the teacher's existing index would remove the bottleneck. The standard recipe,
中文介绍 基于视觉语言模型的多向量检索器在视觉文档检索中领先,但每次搜索都需运行数十亿参数的查询编码器,成为瓶颈。论文提出 ColNanoVDR,利用最优传输做无文档查询蒸馏,将查询编码器蒸馏为小模型并复用教师索引,显著降低检索计算开销。
👍 4
09/28 08:00
Reliable AI safeguards require both control mechanisms that reduce unsafe behavior and monitoring mechanisms that detect safety risks during model interactions. Established behavioral safeguards include alignment methods that optimize model outputs and text monitors that assess interaction text. Rep
中文介绍 可靠 AI 防护既需要降低不安全行为的控制机制,也需要检测风险的监控机制,现有方案多基于行为层面对齐或文本监控。论文系统探究 representation engineering 在 LLM 安全中的作用,分析何种条件下模型内部表征能带来增益,明确内部表示方法相对行为防护的适用边界。
👍 3
09/28 08:00
RLVR provides reliable trajectory-level credit, while OPSD offers dense supervision for token-level credit. This exposes a fundamental coupling when updating step-level credit direction and magnitude with teacher supervision, preventing steps from receiving reliable credit directions and contributio
中文介绍 RLVR 提供可靠的轨迹级奖励信号,OPSD 提供 token 级稠密监督,但用教师监督同时更新 step 级信用的方向与幅度会造成耦合,使步骤难以获得可靠方向与合理权重。论文提出解耦信用的方向与幅度用于自蒸馏,缓解该耦合问题。
👍 3
09/27 08:00
Connectionist temporal classification (CTC) naturally supports offline and streaming speech recognition with utterance-level supervision, but conventional implementations materialize frame-by-vocabulary activations in memory, making CTC training with native LLM vocabularies prohibitively memory-inte
中文介绍 CTC 天然支持离线与流式语音识别,仅需话语级监督,但传统实现需在显存中物化帧×词表的激活,使原生 LLM 大词表下的 CTC 训练显存开销过高。论文提出剪枝 CTC,通过缩减有效词表激活降低显存占用,使大词表 ASR 的 CTC 训练更可行。
👍 3
09/28 08:00
Distillation attacks copy the reasoning capabilities of closed-source large language models, allowing bad actors to replicate state-of-the-art performance at low cost. Attackers systematically collect a large volume of frontier model reasoning traces and then train (i.e., "distill") their own models
中文介绍 蒸馏攻击可大量收集前沿模型推理轨迹并训练自有模型,低成本复制其能力,防御手段随之出现。论文发现针对蒸馏的防御在强化学习后容易被突破:攻击者只需对蒸馏模型做 RL,即可重新逼近被防御的目标能力,说明此类防御的鲁棒性不足。
👍 3
09/28 08:00
Lookup-based memory has been a promising way to scale the parameters of large language models (LLMs). It retrieves learned representations of local token patterns, such as n-grams, instead of reconstructing them through successive layers of computation. However, existing designs such as Engram treat