论文雷达日报|2026-05-01
一句话结论:今日 132 条候选锁定 RLHF / 视觉偏好优化与 agent 可靠性两条主线,S2 相似图与 HF affiliation 同步缺失导致延伸阅读与新作者发现两段降级。
摘要
今日 132 条候选,三源全部正常召回(arXiv 93 篇 / HF Daily 命中 39 篇 / Semantic Scholar 元数据补全 21 篇)。Top picks 集中在 RLHF 与偏好优化(image editing verifier、Poly-DPO、Semi-DPO)与 LLM agent 可靠性(链上真金交易、长上下文红队、科学基础模型协同),并夹一篇视觉生成 taxonomy 与一篇 LLM 推理可控性。S2 similar_papers 字段未返回,延伸阅读段降级;HF Daily 元数据无 affiliation,新作者发现段无法机构验证。seen-pool 14 天窗口已记录 252 条,今日候选无 seen_before 命中。
📌 Top picks (交叉命中)
-
Leveraging Verifier-Based Reinforcement Learning in Image Editing
- 作者:Hanzhong Guo, Jie Wu, Jie Liu, Yu Gao, 等 5 人
- 信号:HF upvotes=15 · HF rank=10 · score=9.0 · reasons=hf_trending_rank:10; watchlist_keyword:reasoning,rlhf,preference optimization; nice_to_have:sft,fine-tuning
- 中文速读:用推理型 verifier 把 RLHF 引入图像编辑奖励建模
- 入选理由:ranking_score 最高(HF trending #10 + watchlist 命中 reasoning/rlhf/preference optimization);把 verifier-based RL 思路延伸到图像编辑奖励,是 RLHF 跨模态扩展的代表性工作。
- 补充链接:https://huggingface.co/papers/2604.27505
-
Visual Generation in the New Era: An Evolution from Atomic Mapping to Agentic World Modeling
- 作者:Keming Wu, Zuhao Yang, Kaichen Zhang, Shizun Wang, 等 23 人
- 信号:HF upvotes=65 · HF rank=20 · score=8.0 · reasons=hf_trending_rank:20; watchlist_keyword:reasoning,agent,world model; nice_to_have:benchmark,evaluation
- 中文速读:视觉生成五级分类:从原子映射演进到 agentic 世界模型
- 入选理由:HF #20、65 upvotes,watchlist 命中 reasoning + agent + world model;提供视觉生成五级 taxonomy,对从生成模型转向 agentic world model 的研究有 framing 价值。
- 补充链接:https://huggingface.co/papers/2604.28185
-
ViPO: Visual Preference Optimization at Scale
- 作者:Ming Li, Jie Wu, Justin Cui, Xiaojie Li, 等 2 人
- 信号:HF upvotes=1 · HF rank=1 · score=7.9 · reasons=hf_trending_rank:1; watchlist_keyword:dpo,preference optimization; citation_velocity:1.0
- 中文速读:Poly-DPO 解噪扩展,规模化视觉偏好对齐
- 入选理由:HF trending #1,watchlist 命中 dpo/preference optimization 且 S2 已返回 tldr;把 DPO 扩展到视觉偏好规模化,是工业界做视觉对齐时直接可参考的 recipe。
- S2 TLDR:This work proposes Poly-DPO, which extends the DPO objective with an additional polynomial term that dynamically adjusts model confidence based on dataset characteristics, enabling effective learning across diverse data distributions and confirms that addressing both algorithmic adaptability and data quality is essential for scaling visual preference optimization.
- 补充链接:https://huggingface.co/papers/2604.24953 · https://www.semanticscholar.org/paper/8119ea592079b53ce4900d740f62616e5f708bdb
-
- 作者:Xinxin Liu, Ming Li, Zonglin Lyu, Yuzhang Shang, 等 1 人
- 信号:HF upvotes=1 · HF rank=2 · score=7.8 · reasons=hf_trending_rank:2; watchlist_keyword:dpo,preference optimization; citation_velocity:1.0
- 中文速读:Semi-DPO 用半监督处理多维偏好冲突标签
- 入选理由:HF trending #2,与 ViPO 同源团队的姊妹篇;把噪声偏好转成半监督学习问题,避免对扩散 DPO 的盲目优化。
- S2 TLDR:Semi-DPO is proposed, a semi-supervised approach that treats consistent pairs as clean labeled data and conflicting ones as noisy unlabeled data and significantly improves alignment with complex human preferences, without requiring additional human annotation or explicit reward models during training.
- 补充链接:https://huggingface.co/papers/2604.24952 · https://www.semanticscholar.org/paper/1c6df1c167dd2442d1bb86c1dce8e65e0936ce77
-
- 作者:Yanting Wang, Chenlong Yin, Ying Chen, Jinyuan Jia
- 信号:HF upvotes=0 · HF rank=3 · score=7.2 · reasons=hf_trending_rank:3; watchlist_keyword:agent,long context; nice_to_have:evaluation
- 中文速读:FlashRT 高效红队评测长上下文提示注入
- 入选理由:HF trending #3,watchlist 命中 agent + long context;针对长上下文 LLM 提示注入与知识投毒,提出计算/内存高效的红队方法。
- 补充链接:https://huggingface.co/papers/2604.28157
-
Compliance versus Sensibility: On the Reasoning Controllability in Large Language Models
- 作者:Xingwei Tan, Marco Valentino, Mahmud Elahi Akhter, Yuxiang Zhou, 等 2 人
- 信号:HF upvotes=5 · HF rank=6 · score=6.9 · reasons=hf_trending_rank:6; watchlist_keyword:reasoning,inference; nice_to_have:evaluation
- 中文速读:首篇系统拆解 LLM 推理模式可控性的研究
- 入选理由:HF #6,watchlist 命中 reasoning/inference;首次系统刻画归纳/演绎/溯因等推理模式与具体题面的解耦,对推理可控性研究有基线意义。
- 补充链接:https://huggingface.co/papers/2604.27251
-
Heterogeneous Scientific Foundation Model Collaboration
- 作者:Zihao Li, Jiaru Zou, Feihao Fang, Xuying Ning, 等 5 人
- 信号:HF upvotes=146 · HF rank=21 · score=6.9 · reasons=hf_trending_rank:21; watchlist_keyword:reasoning,agent,inference
- 中文速读:Eywa:让科学基础模型异构协作的 agent 框架
- 入选理由:HF #21(146 upvotes,全榜最高),watchlist 命中 reasoning + agent + inference;提出 Eywa 让 LLM agent 调度领域科学基础模型,跨模态/跨学科 agent 工程参考。
- 补充链接:https://huggingface.co/papers/2604.27351
-
Operating-Layer Controls for Onchain Language-Model Agents Under Real Capital
- 作者:T. J. Barton, Chris Constantakis, Patti Hauseman, Annie Mous, 等 3 人
- 信号:HF upvotes=5 · HF rank=27 · score=6.8 · reasons=hf_trending_rank:27; watchlist_keyword:reasoning,agent,inference; nice_to_have:benchmark
- 中文速读:21 天链上 3505 agent 真金交易的操作层评测
- 入选理由:HF #27 + watchlist agent + benchmark + S2 tldr;3505 个 agent、$20M 真金链上交易的实测结果,是 agent 在真实资本环境下可靠性的稀缺工程数据。
- S2 TLDR:It is shown that capital-managing agents should be evaluated across the full path from user mandate to prompt, validated action, and settlement, and it is shown that capital-managing agents should be evaluated across the full path from user mandate to prompt, validated action, and settlement.
- 补充链接:https://huggingface.co/papers/2604.26091 · https://www.semanticscholar.org/paper/8b15d5034f02df575acd77d0a9031c7ad3e485d0
🏷 Watchlist 分类命中
cs.AI / cs.LG (reasoning & agents)
- LaST-R1: Reinforcing Action via Adaptive Physical Latent Reasoning for VLA Models(2604.28192)— LaST-R1:通过自适应物理潜在推理增强 VLA 动作。
- 信号:score=6.5 · reasons=watchlist_keyword:reasoning,vla,world model; nice_to_have:benchmark
- TopBench: A Benchmark for Implicit Prediction and Reasoning over Tabular Question Answering(2604.28076)— TopBench:表格隐式预测与推理基准。
- 信号:score=6.5 · reasons=watchlist_keyword:reasoning,agent,inference; nice_to_have:benchmark
- Echo-α: Large Agentic Multimodal Reasoning Model for Ultrasound Interpretation(2604.28011)— Echo-α:超声判读的 agentic 多模态推理模型。
- 信号:score=6.5 · reasons=watchlist_keyword:reasoning,agent,inference; nice_to_have:benchmark
- A Pattern Language for Resilient Visual Agents(2604.28001)— 为视觉 agent 失败模式提出的模式语言。
- 信号:score=4.0 · reasons=watchlist_keyword:agent,vla
cs.CV / cs.GR (generation & alignment)
- HERMES++: Toward a Unified Driving World Model for 3D Scene Understanding and Generation(2604.28196)— HERMES++:统一驾驶世界模型,含 3D 场景一致性评估。
- 信号:score=5.0 · reasons=watchlist_keyword:reasoning,world model; nice_to_have:benchmark,evaluation
cs.CL (RLHF / preference / fine-tuning)
- PRISM: Pre-alignment via Black-box On-policy Distillation for Multimodal Reinforcement Learning(2604.28123)— PRISM:黑盒 on-policy 蒸馏,做 SFT 前对齐。
- 信号:score=5.5 · reasons=watchlist_keyword:reasoning,moe; nice_to_have:benchmark,sft,fine-tuning
cs.DC / cs.LG (systems & efficiency)
- Enhanced Privacy and Communication Efficiency in Non-IID Federated Learning with Adaptive Quantization and Differential Privacy(2604.23426)— Non-IID 联邦学习的隐私与通信效率改进。
- 信号:score=6.5 · reasons=hf_trending_rank:25; watchlist_keyword:quantization,scheduler; citation_velocity:2.0
- Auto-FlexSwitch: Efficient Dynamic Model Merging via Learnable Task Vector Compression(2604.28109)— Auto-FlexSwitch:基于可学开关的动态模型合并。
- 信号:score=4.0 · reasons=watchlist_keyword:quantization,inference
🔗 延伸阅读 (Semantic Scholar 相似论文)
本段今日无高置信度增量信号(S2 相似论文未返回)。
🧑🔬 新出现的作者 / 团队
本日发现扫描未发现达标候选:HF Daily 候选未携带 affiliation 字段,arXiv 抓取在本期亦未填充机构,无法满足 discovery_rules 的机构交叉验证门槛。规则上宁愿空也不凑数,等下次 abstract scrape 拿到机构再补。
📉 覆盖缺口与不确定性
s2_similar_unavailable:本期 Semantic Scholar 候选项均未返回similar_papers,延伸阅读段降级为占位说明。hf_affiliation_missing:HF Daily JSON 不附作者机构,且 arXiv 抓取本期 affiliation 字段为空,新作者发现段无法机构层验证。- seen-pool 14 天窗口含 252 条,本期 132 条候选
seen_before=False,未触发任何降级;说明窗口内候选无重叠,但同时也意味着 trending 列表本日整体新鲜。
来源与交叉验证说明
arXiv 为 primary(论文 PDF / 摘要 ground truth);HF Daily 提供 trending 信号与 upvote 强度(curated);Semantic Scholar 补 tldr 与 paper id(metadata)。本期未引用任何 other 来源。
结论锚定 arXiv 摘要原文;HF trending rank 与 upvotes 仅作策划信号,未当作论文结果证据。S2 的 tldr 在引用时直接放进 tldr_en 字段(未自行翻译),ranking_reasons 透明展示打分依据。