前沿论文周报 · 2026 年第 37 周
本期周报由 AI 根据每日论文数据起草,深度精读基于论文正文,经独立的 AI 事实核查未发现存疑项,未经人工逐篇审核;数字与结论以论文原文为准。
各期周报
本周概览
本周最值得跟踪的变化是具身与机器人学习在顶会口径下明显放量,RoboSPA 与 HuRo 分别从复杂空间—长程诊断和规模化预训练两侧检验现有 VLA[1][2]。基础模型方向仍以 LLM 为主(23 篇),本周工作覆盖工具调用、Agent 强化学习与多语表示对齐,并伴随 VEX-Bench、LexAgentHallu 等 agent 向评测[3][4]。音频方向 10 篇,BuzzASR 与 SEA-SpeechBench 强化低资源与区域多任务语音理解[5][6]。生成侧篇数不多(6 篇),但文生图加速、文化偏见审计与视频推理评测(VWG-Bench)指向质量与 Responsible AI 并重[7][8]。VLM 本周 7 篇,低于前 4 周 P1 周均 11.0,faithfulness 与 RL 自验证等机制分析仍活跃[9][10]。
研究热点
按重要性排序;动向由 AI 依据下方统计和论文内容判断
1VLA 基准与具身数据规模化
升温具身智能
本周具身智能 P1 篇数为 12,前 4 周 P1 周均为 2.3,增量主要来自 CoRL、NeurIPS 等机器人与会论文。RoboSPA 用 10 类 56 基任务与长视界程序规划指标暴露 VLA 在复杂空间关系与记忆型规划上的缺口[1]。HuRo 以约 63 万条 episode 的人类视频机器人化预训练,将四项真实操作的平均完成分从 51.5% 提到 80.3%[2]。ContextFlow 在 LIBERO 未见配置上较 ICRT 平均成功率高 35 个百分点,强调上下文流匹配而非纯缩放[11]。WorldAgen 将世界模型与动作头统一并在测试时轻量 TTT,与干预感知世界模型 WHIRL 共同体现「模型 + 真实交互信号」并进[12][13]。
代表论文
- RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?(在新窗口打开)
- HuRo: Robotizing Human Videos for Scalable VLA Pretraining(在新窗口打开)
- ContextFlow: In-Context Flow Matching for Robot Manipulation(在新窗口打开)
- WorldAgen: Unified State-Action Prediction with Test-Time World Model Training(在新窗口打开)
- How to Learn from What a Human Would Avoid? Intervention-Aware World Models with Real-World RL for Dexterous Manipulation(在新窗口打开)
- Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents(在新窗口打开)
2Agent 评测与工具调用训练闭环
持续LLMGUI Agent
benchmark 关键词本周覆盖 9 篇,Agent 向评测从软件供应链可利用性(VEX-Bench,75 例专家标注)到法律多步轨迹幻觉分层(LexAgentHallu,3414 例)细化失败模式[3][4]。工具侧,KOPA-Bench 与 EDGE live API 合成可执行多步轨迹,GRPO 微调 9B 接近未微调 27B[15];ToolLoop 闭环自反馈合成 11K 样本,4B 在 BFCL 非推理模式达 86.40%[16]。Elastic Horizon 按成功轨迹长度动态定交互 horizon,最多节省约 25% 逐步 token[17]。跨设备 GUI 基准 JarvisGUI 则显示 SOTA 开源 Agent 在长程状态转移上仍弱[18]。
代表论文
- VEX-Bench: Benchmarking LLM Agents for Assessing Exploitability of Software Supply Chain Vulnerabilities(在新窗口打开)
- LexAgentHallu: A Hierarchical Benchmark for Profiling Hallucinations in Legal Agents(在新窗口打开)
- Multi-Step Tool-Calling over Korean Open Public APIs: A Benchmark and a Data-Synthesis Recipe(在新窗口打开)
- ToolLoop: Closed-Loop Tool-Use Data Synthesis via Decomposed Generation and Dynamic Self-Feedback(在新窗口打开)
- Elastic Horizon: Discovering the Effective Interaction Frontier in Agentic Reinforcement Learning(在新窗口打开)
- JarvisGUI: Towards Cross-Device GUI Agents with Dynamic Task Composition(在新窗口打开)
3多语语音理解与 ASR 特化
升温音频
音频 P1 本周 10 篇,高于前 4 周 P1 周均 7.3。BuzzASR 发布 102 语单语 Whisper 特化,77/102 语 CER 优于 Whisper-large-v3[5]。SEA-SpeechBench 以 11 种东南亚语、99 评测集与约 597 小时音频系统暴露低资源与时序理解短板[6]。TRILOGUE 将近 390 小时配对音频的三语口语事实核查与 ASR 退化一并纳入评测[19]。全双工场景下 AV-STE 在冻结 Moshi 的前提下将 GPT-4o 评判连贯性由 1.42 升至 1.91[20]。
代表论文
- BuzzASR: A Swarm of 100+ Monolingual Speech Recognition Models(在新窗口打开)
- SEA-SpeechBench: A Large-Scale Multitask Benchmark for Speech Understanding Across Southeast Asia(在新窗口打开)
- TRILOGUE: A Trilingual Spoken Dialogue Fact-Checking Benchmark with Evidence and Paired Audio(在新窗口打开)
- Noise Adaptive Streaming Audio-Visual Speech Token Enhancement for Robust Full-Duplex Spoken Dialogue Models(在新窗口打开)
- TAD: Token-Adaptive Contrastive Decoding with Confidence-Guided Gating for Hallucination Mitigation in Large Audio-Language Models(在新窗口打开)
4VLM 忠实度与 RL 推理智能体
持续VLM
VLM 本周 P1 为 7 篇,低于前 4 周 P1 周均 11.0,但机制与对齐议题集中。模态缺失下自解释干预协议显示十款 VLM 系统性高估证据充分性[9]。SVRL 用 GRPO 在推理链内自验证检索证据,约 5000 条视觉 QA 微调 Qwen-2.5-VL-7B 提升多跳 VQA 与工具效率[10]。微短剧基准 M-Drama(3.5 万+实例)与 SAGA 图匹配奖励强化叙事结构理解[22]。IVSGround 与免训练 GTA 分别推进 3D grounding 与注视—语义统一[24][23]。
代表论文
- I Don't Miss You, but I Do: Self-Explanation Faithfulness of Modality Missingness in Vision-Language Models(在新窗口打开)
- Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning(在新窗口打开)
- Beyond Sparse Rewards: A New Benchmark and Structure-Aware Graph Alignment for Micro-Drama Understanding(在新窗口打开)
- From Gaze to Meaning: A Training-Free AI Agent for Unified Grounding and Explanation(在新窗口打开)
- Where to Look Matters: Learning Influential Views for VLM-based 3D Visual Grounding(在新窗口打开)
5生成模型加速、推理与 Responsible 审计
持续文生图视频生成
生成分组本周 6 篇。文生图侧 SJD-SV 在子序列级做语义验证以缓解 Speculative Jacobi 歧义[25];GP-Refiner 用高斯过程校正 Diffusion Transformer 特征缓存[26]。WORLDVIEW 以 8960 条 prompt、15 语言审计 DALL-E-3 等系统的隐式 prompt 修订层为文化偏见主因[7]。视频生成 VWG-Bench 覆盖 9 维推理,揭示渲染分高但规则约束任务普遍失败,Vid-PRE 用纯文本 RL 将推理卸载至 VLM[8]。
代表论文
- SJD-SV: Speculative Jacobi Decoding with Semantics Verification for Autoregressive Image Generation(在新窗口打开)
- Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems(在新窗口打开)
- From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models(在新窗口打开)
- Accelerating Diffusion Transformers with Gaussian Process Rectified Feature Cache(在新窗口打开)
- Shedding Light: A Benchmark for Evaluating Lighting Understanding in Generative Image Models(在新窗口打开)
热点统计
由站点根据每日数据计算,与周报起草时提供给 AI 的统计口径一致
本周入选 60 篇:P1 顶会 60
各方向本周与前 4 周均值
按论文主方向 · 对比 W33、W34、W35、W36
查看数据表
| 研究方向 | 本周 | 前 4 周均值 | 本周 P1 | 前 4 周均值(P1) |
|---|---|---|---|---|
| 文生图 | 4 | 2.0 | 4 | 2.0 |
| 图编辑 | 1 | 0.5 | 1 | 0.5 |
| 视频生成 | 1 | 2.5 | 1 | 2.5 |
| 3D 生成 | 0 | 0.3 | 0 | 0.3 |
| Visual2Code | 0 | 0.0 | 0 | 0.0 |
| 3D2Code | 0 | 0.3 | 0 | 0.3 |
| GUI Agent | 1 | 0.8 | 1 | 0.8 |
| 世界模型 | 0 | 1.0 | 0 | 1.0 |
| 具身智能 | 12 | 2.3 | 12 | 2.3 |
| VLM | 7 | 11.0 | 7 | 11.0 |
| OCR | 1 | 0.5 | 1 | 0.5 |
| 音频 | 10 | 7.3 | 10 | 7.3 |
| LLM | 23 | 17.8 | 23 | 17.8 |
| 统一多模态 | 0 | 0.0 | 0 | 0.0 |
口径:按论文主方向计数;均值为前几周中有每日数据的周的周均篇数,保留 1 位小数。2026-10-14 及以前为历史补采,作者单位只来自 OpenAlex,P3 论文明显偏少,前 4 周全部是补采数据;前后对比以 P1 为可比口径(默认显示)。
本周高频关键词
AI 要点关键词 · 前 15 个,按论文数
查看数据表
| 关键词 | 论文数(篇) |
|---|---|
| benchmark | 9 |
| LLM agent | 5 |
| VLM | 5 |
| multilingual | 4 |
| reinforcement learning | 4 |
| VLA | 4 |
| GRPO | 3 |
| imitation learning | 3 |
| tool use | 3 |
| ASR | 2 |
| data synthesis | 2 |
| faithfulness | 2 |
| hallucination | 2 |
| humanoid | 2 |
| LLM | 2 |
口径:同一篇论文中的同一关键词只算一次,大小写与全半角归并。
深度精读
基于论文正文撰写;每条出处的英文原句都经管线核对,确认在正文中逐字出现,结论中的数字出现在原句中
RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?(在新窗口打开)
RoboSPA:VLA 模型能否超越简单场景与短视界任务?
- 要解决的问题
- 现有 VLA 数据集与基准多在固定设定下测任务完成率,难以在空间与程序复杂度递增时诊断模型的具身推理能力,尤其缺少细粒度空间消歧与长视界逐步诊断。
- 方法
- RoboSPA 以 Fine-Grained Spatial Reasoning 与 Long-Horizon Procedural Planning 两维能力 taxonomy 组织 10 类 56 基任务;每任务 5 档难度得 280 变体,仿真采集 527K 轨迹与多本体场景。除 Success Rate 外引入 ONTA 与 Progress Score,并支持 step-level 评估。
- 主要结果
- 四款代表 VLA 随 L1→L5 成功率普遍下滑;最难设定下整体平均成功率低于 25%,部分任务为 0%。π0.5 整体最好,L5 仅 22.3%,X-VLA 为 19.9%;长视界上 π0.5 从 L1 的 71.2% 降至 L5 的 23.1%。Memory-Intensive Planning 在最高难度上各模型均失败。
- 局限
- 任务均在仿真中构建,sim-to-real 可能限制直接迁移;基准侧重桌面操作,未覆盖可形变物体、人机交互与开放式指令等更广设定。
- 意义
- 将细粒度空间推理与长视界程序规划设为可分级诊断维度,并给出大规模轨迹与互补指标,为评估与改进 VLA 在复杂场景下的 grounding 与规划提供较系统的测试床。
正文出处
AbstractRoboSPA 含 10 类 56 基任务、五档难度共 280 变体,并采集约 527K 轨迹。
covering 10 task categories and 56 base tasks. Each task is instantiated across five difficulty levels, yielding 280 variants with increasing spatial ambiguity and procedural complexity. We collect 527K trajectories across multiple embodiments and diverse scenes.
§1 Introduction最难任务上各模型平均成功率低于 25%,部分任务可降至 0%。
On the hardest tasks, all models achieve an average success rate below 25%, with some tasks dropping to 0%.
§4.2 Main Resultsπ0.5 在 L5 整体 Success Rate 为 22.3%,X-VLA 为 19.9%;RDT 与 GO-1 在 L5 分别为 6.9% 与 8.8%。
Among the evaluated models, π0.5 performs best overall, but reaches only 22.3% success at L5, followed by X-VLA with 19.9%. Both drop by over 30 percentage points as difficulty increases. GO-1 and RDT perform worse, achieving only 8.8% and 6.9% success at L5, respectively.
§4.2 Main Results长视界程序规划上,表现最好的 π0.5 从 L1 的 71.2% 降至 L5 的 23.1%。
For Long-Horizon Procedural Planning, the performance drop is even more pronounced, with the best-performing model, π0.5, decreasing from 71.2% at L1 to 23.1% at L5.
§4.2 Main Results细粒度空间推理中 X-VLA 相对最好,但 L1 的 41.9% 仍跌至 L5 的 23.9%。
X-VLA shows the best performance in this dimension, but its success rate still drops from 41.9% at L1 to only 23.9% at L5.
Limitations论文承认任务全在仿真中,且基准聚焦桌面操作。
First, all tasks are constructed in simulation, and the sim-to-real gap may limit direct transfer to physical robots. Second, our benchmark focuses on tabletop manipulation, which may not fully capture the diversity and open-endedness of real-world environments.
HuRo: Robotizing Human Videos for Scalable VLA Pretraining(在新窗口打开)
HuRo:将人类视频机器人化以规模化 VLA 预训练
- 要解决的问题
- VLA 策略依赖大规模预训练,但真实机器人数据成本高。人类视频更易获取且多样,却存在人机 embodiment gap。既有方法多在任务匹配设定下做观测–动作 robotization,尚不清楚异构人类视频能否作为可扩展、联合对齐的机器人监督。
- 方法
- HuRo 流水线分三阶段:标注阶段估计相机内参、手部 pose、egocentric 轨迹并用 VLM 生成 chunk 级语言指令;动作阶段用 PyRoKi 将手部运动 retarget 到目标机器人关节,并由 retarget 后的状态轨迹构造动作标签(a_t = s_{t+1});视觉阶段分割 inpaint 去除人体后在原场景叠加 Isaac Sim 渲染的机器人。流水线可据各源已有标注补全缺失信号,统一输出 robotized observation–action episode,并构建以 ALLEX 为主目标的 HuRo 数据集。
- 主要结果
- HuRo 含超过 630K 条 robotized episode、142M 处理帧,来自五个 egocentric 人类视频源。在四项 ALLEX 真实操作任务上,无 HuRo 预训练至使用全量 HuRo,平均完成分由 51.5% 升至 80.3%;ID 平均完成分由 68.1% 至 88.4%,OOD(空间与视觉偏移)平均完成分由 34.9% 至 72.2%。全量 HuRo 在 ID/OOD 上优于 π0.5 与 GR00T N1.6 参考。无 overlay 的 human 观测在 OOD 上为 55.7%,低于完整 robotization 的 72.2%;Visual + Action 端到端预训练在 Diverse Pick-and-Place 上达 61.1%(ID)与 50.0%(OOD),明显优于 visual-only。
- 局限
- 论文指出:robotized 观测保真度受 inpaint 与 overlay 质量限制,且未显式建模机器人与场景的遮挡;数据仅有视觉与运动学动作,无 force/tactile;运动学 retarget 未建模自碰撞与物理接触,五源审计中仅 55.2% 轨迹在采样帧内无 detected non-grasp self-contact,轨迹主要作预训练监督而非可直接执行的演示。
- 意义
- 在异构、非任务匹配的人类视频上系统验证 joint observation–action robotization 可用于 VLA 预训练,并发布规模显著大于先前 robotized-video 预训练工作的 HuRo。实验表明数据规模、视觉 robotization 与 retarget 动作监督对真实机泛化均有贡献,为人机迁移数据工程提供有真实机器人实验支持的可扩展路径。
正文出处
§1 IntroductionHuRo 超过 630K episode、142M 帧,来自五个人类视频源。
Using this pipeline, we construct the HuRo dataset, a large-scale dataset comprising over 630 K robotized episodes and 142 M processed frames from five human-video sources, over an order of magnitude larger than those used in prior robotized-video pretraining methods [21, 19].
§4.2四项真实任务上,预训练 HuRo 使平均完成分从 51.5% 升至 80.3%。
the average completion score increases from 51.5% without HuRo pretraining to 80.3% with the full HuRo dataset.
§4.2ID 平均完成分从 68.1% 到 88.4%,OOD 平均完成分从 34.9% 到 72.2%。
with ID completion increasing from 68.1% to 88.4% and OOD completion from 34.9% to 72.2%.
§4.3no-overlay 在 OOD 为 55.7%,完整 HuRo 为 72.2%;全量 no-overlay 仍低于 10% PT 的 OOD 59.5%。
substantially lower OOD performance (55.7% vs. 72.2%). Notably, despite using the full HuRo dataset, the no-overlay variant also underperforms the 10% PT model under OOD evaluation (55.7% vs. 59.5%).
§4.4Visual + Action 预训练在 Diverse Pick-and-Place 上 ID/OOD 为 61.1% 与 50.0%。
PT (Visual + Action) substantially improves completion under both ID and OOD evaluation, reaching 61.1% and 50.0%, respectively.
§6 Limitations五源审计中仅 55.2% 轨迹无 detected non-grasp self-contact。
In a five-source audit, only 55.2% of trajectories had no detected non-grasp self-contact, highlighting that the resulting trajectories serve as pretraining supervision rather than executable robot demonstrations.
From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models(在新窗口打开)
从评测到增强:视频生成模型的 Think-with-Video 推理基准与改进
- 要解决的问题
- 视频生成模型画面已较好,但是否能在视频中执行符号规则、遵守物理约束并完成意图目标仍不清楚;现有基准常把视觉质量与认知正确性混为一谈,难以定位失败发生在流畅度、规则遵循还是目标达成。
- 方法
- 提出 VWG-Bench:9 维推理、38 类任务,三级 VLM-as-Judge(视频级流畅、任务级规则、样本级目标),1–5 分解耦评测。提出 Vid-PRE:与生成器无关的提示重写器,由 VLM 从首帧与用户指令推理 CoT 并输出约束明确的 prompt;约 20K 样本 SFT(拒绝采样≥4 分)+ 1.5K 高难度样本 GRPO,奖励纯文本(意图、空间一致、解正确、格式),不依赖视频渲染信号。
- 主要结果
- VWG-Bench 上最强商用 Veo3.1 总分仅 3.55(1–5);各模型视频质量明显高于规则与目标项,如 Wan2.2 视频质量 2.32、隐式规则 1.56。接 Vid-PRE 后,MME-CoF 上 Wan2.2 总分由 1.72 升至 2.58;V-ReasonBench 平均由 26.65 升至 44.48;VWG-Bench 上 Wan2.2 由 2.13 升至 2.51,隐式规则由 1.56 升至 2.12;Wan2.6-Flash 达 3.26,高于原生 PE 的 3.04。自动评测与人类 Spearman ρ 为 0.69–0.80。
- 局限
- 每任务类仅 10 条实例以控制评测成本;Vid-PRE 增强实验仅覆盖开源 Wan2.2 或可关闭内置 Prompt Extension 的 Wan2.5、Wan2.6-Flash,未改架构的闭源商用生成器未系统报告;结论依赖 VLM-as-Judge 与固定采样设置(如 2 fps、最多 16 帧)。
- 意义
- 在较完整的 think-with-video 维度上给出可解释诊断,并表明用纯文本 RL 做推理卸载的提示增强可在多生成器上提升规则与目标推理,而不必改视频模型结构。
正文出处
§4.2Veo3.1 在 VWG-Bench 上 1–5 分制总分仅 3.55,说明基准难度高。
Even the strongest commercial model (Veo3.1) achieves only 3.55 overall on a 1–5 scale, underscoring the challenge of our benchmark.
§4.2Wan2.2 视频质量 2.32,隐式规则遵循仅 1.56,呈现渲染强、推理弱的分化。
For example, Wan2.2 attains 2.32 in video quality but only 1.56 in implicit rule following
§4.2MME-CoF 上 Vid-PRE 将 Wan2.2 总分从 1.72 提到 2.58。
On MME-CoF (Table 3(a)), Vid-PRE boosts Wan2.2's overall score from 1.72 to 2.58, a 50% gain
§4.2V-ReasonBench 上 Wan2.2 平均准确率从 26.65 升至 44.48。
Vid-PRE improves Wan2.2's average accuracy from 26.65 to 44.48, a 67% gain
§4.2VWG-Bench 上 Wan2.2 总分 2.13→2.51,隐式规则 1.56→2.12。
On VWG-Bench (Table 3(c)), Vid-PRE raises Wan2.2 from 2.13 to 2.51, an 18% gain, and improves its weakest baseline dimension, implicit rule following, from 1.56 to 2.12. For Wan2.6-Flash, Vid-PRE achieves 3.26, surpassing the Wan PE baseline of 3.04
§4.3自动 VLM 评测与人类判断的 Spearman ρ 在 0.69 至 0.80 之间。
As shown in Table 4, Spearman's ρ ranges from 0.69 to 0.80 and the mean absolute error remains below 0.55 across all criteria.
论文推荐
按必读、方法亮点、数据与评测、开源项目分组,每篇附推荐理由
必读3
RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?(在新窗口打开)
RoboSPA:VLA 模型能否超越简单场景与短视界任务?
RoboSPA 大规模基准同时诊断细粒度空间推理与长视界程序规划,直接回答 VLA 能否走出简单场景。
HuRo: Robotizing Human Videos for Scalable VLA Pretraining(在新窗口打开)
HuRo:将人类视频机器人化以规模化 VLA 预训练
HuRo 人类视频机器人化与 63 万 episode 预训练,四项真实操作的平均完成分 51.5%→80.3%。
从评测到增强:视频生成模型的 Think-with-Video 推理基准与改进
VWG-Bench 同时评测视频质量、规则遵循与目标达成,并以 Vid-PRE 在多个生成器上验证提示增强。
方法亮点2
ContextFlow: In-Context Flow Matching for Robot Manipulation(在新窗口打开)
ContextFlow:机器人操作的上下文流匹配
ContextFlow 上下文流匹配在 LIBERO 未见配置大幅领先 ICRT,且接近微调 VLA π₀。
Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning(在新窗口打开)
用强化学习激发多模态推理智能体的自验证能力
SVRL 用 GRPO 激发多模态 Agent 链内自验证,小模型 VQA 与检索效率双提升。
数据与评测3
BuzzASR: A Swarm of 100+ Monolingual Speech Recognition Models(在新窗口打开)
BuzzASR:100+ 单语语音识别模型群
102 语 BuzzASR 开源群与 FLEURS 系统对比,低资源 ASR 工程参考价值高。
SEA-SpeechBench:东南亚大规模多任务语音理解基准
SEA-SpeechBench 区域多任务与时序理解,补东南亚语音评测空白。
Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems(在新窗口打开)
文生图系统中提示修订作为文化偏见来源
WORLDVIEW 多语言审计将文生图文化偏见因果归因于部署中的 prompt 修订层。
开源项目2
Accelerating Diffusion Transformers with Gaussian Process Rectified Feature Cache(在新窗口打开)
基于高斯过程 rectified 特征缓存的 Diffusion Transformer 加速
GP-Refiner 即插即用 Diffusion Transformer 缓存加速,开源实现便于复现与组合。
思维链表面之下:大语言模型推理操作的机制解释
CoT 推理操作在隐藏表示中的可分离性与机制解释,开源利于后续干预研究。
领域洞察
按研究方向分组,各写一段本周的研究动向与判断
生成
本周 6 篇图编辑 1 文生图 4 视频生成 1
本周生成向 P1 共 6 篇,文生图占 4。议题分裂为效率(SJD-SV、GP-Refiner)与 Responsible/物理一致性评测(WORLDVIEW、Shedding Light 光照基准)[25][27]。视频单篇但分量足:VWG-Bench 明确生成模型在逻辑约束上的系统性失败[8]。图编辑 MIEdit 强调免训练 SDE 反演与 EditEval++[29]。
相关论文
- From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models(在新窗口打开)
- Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems(在新窗口打开)
- SJD-SV: Speculative Jacobi Decoding with Semantics Verification for Autoregressive Image Generation(在新窗口打开)
- Multi-History-Step SDE Inversion for Image Editing with Superior Regional Awareness(在新窗口打开)
智能体
本周 1 篇GUI Agent 1
智能体分组仅 1 篇 P1,JarvisGUI 以跨 Android/Windows/Ubuntu 动态组合任务暴露开源 GUI Agent 在状态转移与长程依赖上的缺口[18]。与 foundation 中多条工具调用、BFCL/AppWorld 类 Agent RL 论文形成「基准在 agent、训练在 LLM」的互补,但本周 dedicated GUI 产出仍偏少。
世界模型与具身
本周 12 篇具身智能 12
世界模型与具身合计 12 篇且均属 embodied 主方向,为本周最突出分组。人形与四足并进:SkillX、CHIP 与 Spot 手倒立课程、表示有效秩分析 sim-to-real[30][31][32][33]。操作学习侧扩散策略 ObstaDiff 在真实温室 366 次试验验证[34]。主标 world_model 本周 P1 为 0,但 WHIRL、WorldAgen 在次方向承接世界模型叙事[13][12]。
相关论文
- RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?(在新窗口打开)
- HuRo: Robotizing Human Videos for Scalable VLA Pretraining(在新窗口打开)
- How to Learn from What a Human Would Avoid? Intervention-Aware World Models with Real-World RL for Dexterous Manipulation(在新窗口打开)
- SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer(在新窗口打开)
理解
本周 8 篇OCR 1 VLM 7
理解分组 8 篇,VLM 7、OCR 1。除 3D grounding 与注视智能体 GTA 等任务创新外,faithfulness(模态缺失自解释、规范色 probing)与 RL(SAGA、SVRL)占相当比重[9][10][22]。SupGRPO 将 GRPO 与匹配式在线 SFT 用于 text spotting[35]。
相关论文
- I Don't Miss You, but I Do: Self-Explanation Faithfulness of Modality Missingness in Vision-Language Models(在新窗口打开)
- Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning(在新窗口打开)
- Where to Look Matters: Learning Influential Views for VLM-based 3D Visual Grounding(在新窗口打开)
- SupGRPO: Enhancing GRPO with Matching-based Online SFT for Text Spotting(在新窗口打开)
音频
本周 10 篇音频 10
音频 10 篇 P1,ASR 与 benchmark 双线:BuzzASR、希腊语 ALT 与 Whisper 适配规律[5][36];SEA-SpeechBench、流式多说话人架构权衡研究补系统选型[6][37]。LALM 幻觉抑制 TAD 与 AST 训练加速 SpecAugment-Patch Merging 体现推理/训练效率关注[21][38]。
相关论文
- BuzzASR: A Swarm of 100+ Monolingual Speech Recognition Models(在新窗口打开)
- SEA-SpeechBench: A Large-Scale Multitask Benchmark for Speech Understanding Across Southeast Asia(在新窗口打开)
- Noise Adaptive Streaming Audio-Visual Speech Token Enhancement for Robust Full-Duplex Spoken Dialogue Models(在新窗口打开)
- TAD: Token-Adaptive Contrastive Decoding with Confidence-Guided Gating for Hallucination Mitigation in Large Audio-Language Models(在新窗口打开)
基础模型
本周 23 篇LLM 23
基础模型 23 篇 P1,LLM 独占。本周工具调用合成与评测(KOPA、ToolLoop、PTA 教师锚定)、Agent RL(Elastic Horizon、DDO 多样化成功轨迹)均有多篇代表工作[15][16][40][17]。效率线含 layer dropout、Osprey 投机解码与 MoE ACE 合并[41][42][43]。安全与忠实:SAFEGuard 越狱检测、CodeTD 注意力拓扑幻觉、CAROT 跨语对齐最高 +11.2 点[44][45][39]。
相关论文
- ToolLoop: Closed-Loop Tool-Use Data Synthesis via Decomposed Generation and Dynamic Self-Feedback(在新窗口打开)
- Elastic Horizon: Discovering the Effective Interaction Frontier in Agentic Reinforcement Learning(在新窗口打开)
- Cross-Lingual Representation Alignment by Token-Level Optimal Transport in a Language-Agnostic Space(在新窗口打开)
- VEX-Bench: Benchmarking LLM Agents for Assessing Exploitability of Software Supply Chain Vulnerabilities(在新窗口打开)
参考来源
- [1]RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?(在新窗口打开)↩
- [2]HuRo: Robotizing Human Videos for Scalable VLA Pretraining(在新窗口打开)↩
- [3]VEX-Bench: Benchmarking LLM Agents for Assessing Exploitability of Software Supply Chain Vulnerabilities(在新窗口打开)↩
- [4]LexAgentHallu: A Hierarchical Benchmark for Profiling Hallucinations in Legal Agents(在新窗口打开)↩
- [5]BuzzASR: A Swarm of 100+ Monolingual Speech Recognition Models(在新窗口打开)↩
- [6]SEA-SpeechBench: A Large-Scale Multitask Benchmark for Speech Understanding Across Southeast Asia(在新窗口打开)↩
- [7]Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems(在新窗口打开)↩
- [8]From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models(在新窗口打开)↩
- [9]I Don't Miss You, but I Do: Self-Explanation Faithfulness of Modality Missingness in Vision-Language Models(在新窗口打开)↩
- [10]Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning(在新窗口打开)↩
- [11]ContextFlow: In-Context Flow Matching for Robot Manipulation(在新窗口打开)↩
- [12]WorldAgen: Unified State-Action Prediction with Test-Time World Model Training(在新窗口打开)↩
- [13]How to Learn from What a Human Would Avoid? Intervention-Aware World Models with Real-World RL for Dexterous Manipulation(在新窗口打开)↩
- [14]Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents(在新窗口打开)↩
- [15]Multi-Step Tool-Calling over Korean Open Public APIs: A Benchmark and a Data-Synthesis Recipe(在新窗口打开)↩
- [16]ToolLoop: Closed-Loop Tool-Use Data Synthesis via Decomposed Generation and Dynamic Self-Feedback(在新窗口打开)↩
- [17]Elastic Horizon: Discovering the Effective Interaction Frontier in Agentic Reinforcement Learning(在新窗口打开)↩
- [18]JarvisGUI: Towards Cross-Device GUI Agents with Dynamic Task Composition(在新窗口打开)↩
- [19]TRILOGUE: A Trilingual Spoken Dialogue Fact-Checking Benchmark with Evidence and Paired Audio(在新窗口打开)↩
- [20]Noise Adaptive Streaming Audio-Visual Speech Token Enhancement for Robust Full-Duplex Spoken Dialogue Models(在新窗口打开)↩
- [21]TAD: Token-Adaptive Contrastive Decoding with Confidence-Guided Gating for Hallucination Mitigation in Large Audio-Language Models(在新窗口打开)↩
- [22]Beyond Sparse Rewards: A New Benchmark and Structure-Aware Graph Alignment for Micro-Drama Understanding(在新窗口打开)↩
- [23]From Gaze to Meaning: A Training-Free AI Agent for Unified Grounding and Explanation(在新窗口打开)↩
- [24]Where to Look Matters: Learning Influential Views for VLM-based 3D Visual Grounding(在新窗口打开)↩
- [25]SJD-SV: Speculative Jacobi Decoding with Semantics Verification for Autoregressive Image Generation(在新窗口打开)↩
- [26]Accelerating Diffusion Transformers with Gaussian Process Rectified Feature Cache(在新窗口打开)↩
- [27]Shedding Light: A Benchmark for Evaluating Lighting Understanding in Generative Image Models(在新窗口打开)↩
- [28]Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs(在新窗口打开)↩
- [29]Multi-History-Step SDE Inversion for Image Editing with Superior Regional Awareness(在新窗口打开)↩
- [30]SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer(在新窗口打开)↩
- [31]Unifying Physics-Based Humanoid Interaction with a Context-Conditioned Interaction Prior(在新窗口打开)↩
- [32]Actuator Dynamics Curricula for Narrow-Viability Tasks in Legged Robot Learning(在新窗口打开)↩
- [33]Mind the Phase: Effective Rank and Representation Health in Legged Locomotion(在新窗口打开)↩
- [34]ObstaDiff: Generalizable Diffusion Policy Learning via Obstacle-aware Representations(在新窗口打开)↩
- [35]SupGRPO: Enhancing GRPO with Matching-based Online SFT for Text Spotting(在新窗口打开)↩
- [36]Automatic Lyric Transcription for Greek Songs: Scaling and Task Composition Effects in Whisper Adaptation(在新窗口打开)↩
- [37]Pushing the Boundaries of Streaming Multi-Speaker ASR: A Systematic Study of Architectural Trade-offs(在新窗口打开)↩
- [38]From Masking to Merging: Rethinking SpecAugment for Efficient Audio Spectrogram Transformer(在新窗口打开)↩
- [39]Cross-Lingual Representation Alignment by Token-Level Optimal Transport in a Language-Agnostic Space(在新窗口打开)↩
- [40]Persistent Teacher Anchoring for Tool-Using Agents(在新窗口打开)↩
- [41]Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference(在新窗口打开)↩
- [42]Osprey: Target-agnostic Pre-training Makes Stronger Drafters in Speculative Decoding(在新窗口打开)↩
- [43]ACE: Adapter Consolidation across Experts for Parameter-Efficient Fine-Tuning of MoE LLMs(在新窗口打开)↩
- [44]SAFEGuard: Detect Optimization-Based Jailbreak Attacks Through Harmful Semantic Analysis and Fluency Measurement(在新窗口打开)↩
- [45]CodeTD: Topology of Attention Detects Hallucinations in Code LLMs(在新窗口打开)↩