论文

本文提出 Token-Level Off-Policy Labeling (TOPL)，一种将后训练重构为 token 级正确性预测任务的离线策略训练范式。其核心思路是：通过训练模型区分响应中的好 token 与坏 token，自然引导模型生成好 token，同时避免直接训练模型生成离线策略 token 所带来的缺陷。在文档摘要任务上的实验表明，TOPL 在 11 个数据集上针对多种序列级与 token 级方法实现了强大的分布外泛化能力。We propose Token-Level Off-Policy Labeling (TOPL), an off-policy training paradigm that reframes post-training as a token-level correctness prediction task. Our key intuition is that by training the model to distinguish good and bad tokens in a response, we naturally guide the model towards generating good tokens, while avoiding the pitfalls that come with directly training the model to generate off-policy tokens. Experiments on document summarization tasks show that TOPL achieves strong out-of-distribution generalization across 11 datasets against a diverse set of sequence-level and token-lev

本文提出 FlashRT，一种 Agent Harness，引导编码 Agent 将开发者编写的简易参考实现提升为优化的多 GPU 部署，并可灵活权衡时延与吞吐量等目标指标，证明在专家优化尚不成熟的平台上，由 Agent 驱动的优化具有更高的可扩展性。FlashRT is presented, an agent harness that guides coding agents to lift simple developer-written reference implementations into optimized multi-GPU deployments that flexibly weigh target metrics like latency and throughput, demonstrating that agent-driven optimization can be more scalable on platforms with less mature expert optimization.

在七个基准上，TimeLens2-2B 在所有基准上均优于规模相当的所有基线，4B 和 8B 变体则取得了 SOTA 性能，超越了参数量高达 397B 的开源模型。Across seven benchmarks, TimeLens2-2B outperforms all size-matched baselines on every benchmark, while the 4B and 8B variants achieve state-of-the-art performance, surpassing open-source models with up to 397B parameters.

实验表明，EvolvingWorld 能够通过有效维持持久且一致的角色与世界发展，提升长程模拟能力。Experiments show that EvolvingWorld can improve long-horizon simulation by effectively maintaining persistent, coherent character and world development.

大量实验表明，Distilled RL 在 pass@1 和 pass@k 上均显著优于标准 RL 和 OPD，并能将教师模型中此前无法获得的知识有效迁移至学生模型。Extensive experiments show that Distilled RL substantially outperforms standard RL and OPD in terms of both pass@1 and pass@k, and can effectively transfer previously unavailable knowledge from a teacher model to a student model.

提出了ReflectWorld-MM，一个面向开放视频流的以实体为中心的多模态记忆系统，在六个长视频和终身记忆基准测试上均达到最优准确率，超越了强记忆Agent和前沿模型。ReflectWorld-MM is proposed, an entity-oriented multimodal memory system for open-ended video streams that achieves the best accuracy on all six long-video and lifelong-memory benchmarks, outperforming strong memory agents and a frontier model.

该工作形式化了一个攻击空间，并使用四种 agent 工作负载与 Linux telemetry pipeline 评估了代表性 OS 防御，表明主要限制并非 OS observability。This work formalizes an attack space and evaluates representative OS defenses using four agent workloads and a Linux telemetry pipeline and shows that the main limitation is not OS observability.

该工作将两步方法的遗憾分解为候选生成遗憾和候选内选择遗憾，并利用最近邻估计器和Transformer的预测误差保证对后者进行了界。This work decomposes the regret of the two-step method into candidate-generation regret and within-candidate choice regret, and bound the latter using prediction-error guarantees for nearest-neighbor estimators and transformers.

提出了ShotPlan，一个基于视频扩散基础模型构建的、用于显式多镜头电影级视频生成的框架，显著优于现有的电影级视频生成方法，提供更灵活的镜头管理和更强的跨镜头一致性。ShotPlan is proposed, a framework for explicit multi-shot cinematic video generation built upon a video diffusion foundation model that significantly outperforms existing cinematic video generation methods, offering more flexible shot management and stronger inter-shot consistency.

OCT-Bench能够对MLLM进行全面且细粒度的评估，为识别能力瓶颈和推进临床可信的OCT理解奠定基础。OCT-Bench enables comprehensive and fine-grained evaluation of MLLMs, providing a foundation for identifying capability bottlenecks and advancing clinically grounded OCT understanding.

结果因果性地表明，在主要的 Qwen 设置中，关键支撑信息的遗漏是原因之一，并解释了为何必须同时审计支撑保真度与部署成本。The results causally implicate decision-critical support omission as one contributor in the primary Qwen setting and show why both support fidelity and deployment costs must be audited.

本文提出一个时间感知的音频 LLM，能够基于大规模合成监督（来自级联 pipeline）在长达 120 分钟的输入上回答带有显式时间戳的问题，并在短时长和长时长 benchmark 上取得强劲的时间定位准确率。This work presents a time-aware audio LLM that answers questions with explicit timestamps over up to 120 minutes of input using large-scale synthetic supervision from a cascaded pipeline and achieves strong temporal-grounding accuracy on short and long benchmarks.

本文针对中亚低资源语言构建鲁棒基础模型的挑战，提出 GigaAM Multilingual：一个基于 HuBERT 风格目标在 2M 小时音频上预训练的 Conformer encoder，并在预训练阶段引入 cluster 级的数据均衡策略，在微调阶段引入域感知采样方法，以缓解头部语言主导问题。This work addresses the challenge of building robust foundation models for underrepresented Central Asian languages by presenting GigaAM Multilingual, a Conformer encoder pre-trained on 2M hours of audio using a HuBERT-style objective, and introduces a cluster-level data balancing strategy during pre-training and a domain-aware sampling method during fine-tuning to mitigate head-language dominance.

本文提出 DeepSearch-Evolve，一个面向 web agent 的自蒸馏框架，基于 DeepSearch-World——一个具备可复现搜索与页面读取工具的确定性、可验证环境——从而实现长程 web agent 的可扩展自演化。DeepSearch-Evolve is presented, a self-distillation framework for web agents built on DeepSearch-World, a deterministic and verifiable environment with reproducible search and page-reading tools that enables scalable self-evolution for long-horizon web agents.

本文研究了在采用极小卷积滤波器的架构下，卷积网络深度对大规模图像识别精度的影响，并表明将深度推进至 16-19 个权重层，可在先前 SOTA 配置基础上取得显著提升。This work investigates the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting using an architecture with very small convolution filters, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16-19 weight layers.

在 YouCook2、CrossTask 等教学视频数据集上，基于该数据训练的文本-视频 embedding 在文本到视频检索与动作定位任务上达到了 SOTA 结果。It is demonstrated that a text-video embedding trained on this data leads to state-of-the-art results for text-to-video retrieval and action localization on instructional video datasets such as YouCook2 or CrossTask.

提出在通过跨模态注意力融合之前对齐图像与文本表示的对比损失（ALBEF），可实现更扎实的视觉-语言表征学习；并提出动量蒸馏，一种利用动量模型生成伪目标进行自训练的方法。A contrastive loss to ALign the image and text representations BEfore Fusing (ALBEF) them through cross-modal attention, which enables more grounded vision and language representation learning and proposes momentum distillation, a self-training method which learns from pseudo-targets produced by a momentum model.

本文提出 LLaVA：Large Language and Vision Assistant，一个端到端训练的大型多模态模型，将视觉编码器与 LLM 相结合用于通用视觉和语言理解；并引入 GPT-4 生成的视觉指令微调数据，模型与代码库已开源。This paper presents LLaVA: Large Language and Vision Assistant, an end-to-end trained large multimodal model that connects a vision encoder and LLM for general-purpose visual and language understanding and introduces GPT-4 generated visual instruction tuning data, the model and code base publicly available.

m-RNN 模型直接对给定先前词语和图像条件下生成下一个词的概率分布建模，相较于直接优化排序目标函数进行检索的 SOTA 方法，取得了显著的性能提升。The m-RNN model directly models the probability distribution of generating a word given previous words and an image, and achieves significant performance improvement over the state-of-the-art methods which directly optimize the ranking objective function for retrieval.

分析表明，借助 LLM，美国约 15% 的工作任务可在保持同等质量的前提下显著提速完成，意味着 LLM 驱动的软件将对底层模型经济影响的规模化产生实质性作用。The analysis suggests that, with access to an LLM, about 15% of all worker tasks in the US could be completed significantly faster at the same level of quality, implying that LLM-powered software will have a substantial effect on scaling the economic impacts of the underlying models.

提出极简的 VLP 模型 Vision-and-Language Transformer (ViLT)，其一体化设计将视觉输入处理大幅简化为与文本输入相同的无卷积方式；ViLT 比此前的 VLP 模型快达数十倍，同时下游任务性能具有竞争力甚至更优。A minimal VLP model, Vision-and-Language Transformer (ViLT), monolithic in the sense that the processing of visual inputs is drastically simplified to just the same convolution-free manner that the authors process textual inputs, showing that ViLT is up to tens of times faster than previous VLP models, yet with competitive or better downstream task performance.

Contrastive Captioner (CoCa) 采用极简设计，对图文编码器-解码器基础模型联合使用对比损失与字幕损失进行预训练，从而兼具 CLIP 等对比方法与 SimVLM 等生成方法的能力。Contrastive Captioner (CoCa), a minimalist design to pretrain an image-text encoder-decoder foundation model jointly with contrastive loss and captioning loss, thereby subsuming model capabilities from contrastive approaches like CLIP and generative methods like SimVLM.

本文提出 MiniGPT-4，通过一个投影层将冻结的视觉编码器与冻结的先进 LLM Vicuna 对齐，发现将视觉特征与先进大语言模型恰当对齐可获得类似 GPT-4 所展现的多种先进多模态能力。MiniGPT-4 is presented, which aligns a frozen visual encoder with a frozen advanced LLM, Vicuna, using one projection layer to uncovers that properly aligning the visual features with an advanced large language model can possess numerous advanced multi-modal abilities demonstrated by G PT-4.

本文表明视频掩码自编码器（VideoMAE）是自监督视频预训练（SSVP）的数据高效学习器，并受近期 ImageMAE 启发，提出采用极高掩码比例的定制化视频管状掩码策略。This paper shows that video masked autoencoders (VideoMAE) are data-efficient learners for self-supervised video pre-training (SSVP), and proposes customized video tube masking with an extremely high ratio, inspired by the recent ImageMAE.

本文基于预训练 BLIP-2 模型，对视觉-语言指令微调展开系统全面研究，并提出指令感知的 Query Transformer，用于提取针对给定指令的信息丰富特征。This paper conducts a systematic and comprehensive study on vision-language instruction tuning based on the pretrained BLIP-2 models, and introduces an instruction-aware Query Transformer, which extracts informative features tailored to the given instruction.

本文提出 Chunk Coverage (CC)，一种独立于 oracle 的 RAG 系统检索组件测试充分性准则，结果表明 CC 在无需测试 oracle 的情况下捕获了与有效测试相关的检索多样性。Chunk Coverage (CC), an oracle-independent test adequacy criterion for testing the retrieval component of RAG systems, is introduced and results show that CC captures retrieval diversity relevant to effective testing without requiring test oracles.

研究揭示了 LLM 空间能力中的清晰规律：尽管其仍落后于 SOTA 方法，但具有潜力并能同时处理多种空间约束，从而可扩展到异构场景。A clear pattern is revealed in LLM spatial capabilities: while they still lag behind state-of-the-art approaches, they are promising and can handle multiple spatial constraints simultaneously, enabling scaling to heterogeneous setups.

提出 Manager Coercion Benchmark：被测管理者需要完成一项良性任务并有完成的动机，但唯一能礼貌且坚定地拒绝的智能体本身就是被测管理者自己。The Manager Coercion Benchmark is introduced: the manager under test needs a benign task done and has an incentive to deliver, but the only agent that can do it politely and immovably declines, but the only agent that can do it politely and immovably declines is the manager under test itself.

ReViV 是首个用于整体第一人称 4D 重建的统一框架，能够从单个单目 RGB 视频中同时提取观察者与视角动态，在整体自身体、手部与注视重建以及相机跟踪方面达到 SOTA 精度与效率，同时保持极具竞争力的第一人称深度估计能力。ReViV is the first unified framework for holistic egocentric 4D reconstruction that extracts both viewer and view dynamics from a single monocular RGB video and achieves state-of-the-art accuracy and efficiency across holistic ego-body, hand, and gaze reconstruction, camera tracking, while maintaining highly competitive egocentric depth estimation.

本工作通过 teacher forcing 与 ODE 蒸馏，将一个双向动作条件教师模型逐步蒸馏为因果学生模型，并提出 LongForcing，将学生模型的长时间自展开与扩展时域教师模型对齐，从而缓解累积的分布漂移与自回归漂移。This work progressively distill a bidirectional action-conditioned teacher into a causal student through teacher forcing and ODE distillation, and introduces LongForcing to align long student self-rollouts with an extended-horizon teacher, mitigating accumulated distribution shift and autoregressive drift.

本文是面向业务流程中长时间运行、有状态、多步生成式 AI 系统的基于图的工作流路径实践指南，并通过三个可执行示例展示类型化状态、条件路由、确定性工具、重试、中断、检查点与 trace 如何协同工作。This paper is a practitioner guide to graph-based workflow pathways for long-running, stateful, multi-step generative AI systems in business processes and presents three executable recipes to show how typed state, conditional routing, deterministic tools, retries, interrupts, checkpoints, and traces fit together.

在该研究的评估协议下，部分 AI 生成回答获得了高于可用参考答案的评分，尤其当其包含准确且相关的支撑细节时；但这不应被解读为模型总体上优于合格法律专业人士的证据。Under the study’s evaluation protocol, some AI-generated responses received higher ratings than the available reference answers, particularly when they contained accurate and relevant supporting detail, should not be interpreted as evidence that the models generally outperform qualified legal professionals.

本文提出 RF-Agent，通过多 Agent 的 Question-Thinking-Solution-Answer 流水线，基于教材驱动的知识蒸馏来弥补 RF 领域专用推理的空白，为面向 LLM 辅助 RF 电路设计的未来工作提供了可复用的基础。RF-Agent is presented, which addresses the gap in domain-specific RF reasoning through textbook-driven knowledge distillation through a multi-agent Question-Thinking-Solution-Answer pipeline and provides a reusable foundation for future work on LLM-aided RF circuit design.

这些发现表明，即便长上下文评估正从简单检索转向复杂推理，对相关证据的准确 grounding 仍是一项不可或缺且仍有大幅提升空间的能力。These findings suggest that, even as long-context evaluation shifts from simple retrieval toward complex reasoning, accurate grounding in relevant evidence remains an indispensable capability with substantial room for improvement.

本文提出 HACO，一种运行时控制方案，将每次角色请求视为在候选 agent 实例上的可靠性约束选择问题，每个候选实例耦合了角色类型、LLM 与具体执行环境。HACO is proposed, a runtime control scheme that treats each role request as a reliability-constrained selection problem over candidate agent instances, each coupling a role type, an LLM, and a concrete execution environment.