PRM标注成本降两个数量级，dLLM幻觉错法不同

重点关注

• TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training score 8
  入选 HF Daily Papers；HF 热度: 5 upvotes (+2)；有代码实现；关键词(2): post-training, reasoning

也值得关注

• From Query to Counsel: Structured Reasoning with a Multi-Agent Framework and Dataset for Legal Consultation score 4
  关键词(1): reasoning；顶会接收: ACL
• CHAIRO: Contextual Hierarchical Analogical Induction and Reasoning Optimization for LLMs score 4
  关键词(3): fine-tuning, RAG, reasoning；顶会接收: ACL
• CARO: Chain-of-Analogy Reasoning Optimization for Robust Content Moderation score 4
  关键词(5): fine-tuning, DPO, retrieval-augmented, RAG, reasoning；顶会接收: ACL
• BareBones: Benchmarking Zero-Shot Geometric Comprehension in VLMs score 4
  关键词(2): reasoning, vision-language；顶会接收: CVPR
• Lost in Diffusion: Uncovering Hallucination Patterns and Failure Modes in Diffusion Large Language Models score 4
  关键词(1): pre-training；顶会接收: ACL
• GeoMeld: Toward Semantically Grounded Foundation Models for Remote Sensing score 4
  关键词(2): pretraining, agentic；顶会接收: CVPR
• Efficient Process Reward Modeling via Contrastive Mutual Information score 4
  关键词(1): reasoning；顶会接收: ACL
• Detecting RAG Extraction Attack via Dual-Path Runtime Integrity Game score 4
  关键词(3): latency, retrieval-augmented, RAG；顶会接收: ACL
• UDAPose: Unsupervised Domain Adaptation for Low-Light Human Pose Estimation score 3
  顶会接收: CVPR
• ReFEree: Reference-Free and Fine-Grained Method for Evaluating Factual Consistency in Real-World Code Summarization score 3
  顶会接收: ACL
• VLN-NF: Feasibility-Aware Vision-and-Language Navigation with False-Premise Instructions score 3
  顶会接收: ACL
• NTIRE 2026 Challenge on Short-form UGC Video Restoration in the Wild with Generative Models: Datasets, Methods and Results score 3
  顶会接收: CVPR
• Expect the Unexpected? Testing the Surprisal of Salient Entities score 3
  顶会接收: ACL