Latent推理靠的不是推理

重点关注

• MobilityBench: A Benchmark for Evaluating Route-Planning Agents in Real-World Mobility Scenarios score 13
  机构: Baidu；入选 HF Daily Papers；HF 热度: 98 upvotes (+4)；有代码实现；关键词(1): tool use
• Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimization score 11
  入选 HF Daily Papers；HF 热度: 30 upvotes (+4)；关键词(1): GRPO；顶会接收: ICLR
• EmbodMocap: In-the-Wild 4D Human-Scene Reconstruction for Embodied Agents score 11
  机构: Zhejiang University；入选 HF Daily Papers；HF 热度: 9 upvotes (+2)；有代码实现；关键词(2): fine-tune, embodied
• The Trinity of Consistency as a Defining Principle for General World Models score 10
  入选 HF Daily Papers；HF 热度: 187 upvotes (+4)；有代码实现；关键词(2): scaling, reasoning
• From Blind Spots to Gains: Diagnostic-Driven Iterative Training for Large Multimodal Models score 10
  入选 HF Daily Papers；HF 热度: 146 upvotes (+4)；有代码实现；关键词(1): reasoning
• OmniGAIA: Towards Native Omni-Modal AI Agents score 10
  入选 HF Daily Papers；HF 热度: 49 upvotes (+4)；有代码实现；关键词(3): reasoning, vision-language, open-source
• Imagination Helps Visual Reasoning, But Not Yet in Latent Space score 10
  入选 HF Daily Papers；HF 热度: 36 upvotes (+4)；有代码实现；关键词(1): reasoning
• AgentDropoutV2: Optimizing Information Flow in Multi-Agent Systems via Test-Time Rectify-or-Reject Pruning score 10
  入选 HF Daily Papers；HF 热度: 24 upvotes (+4)；有代码实现；关键词(4): pruning, fine-tuning, retrieval-augmented, reasoning
• Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization score 9
  入选 HF Daily Papers；HF 热度: 18 upvotes (+3)；有代码实现；关键词(5): scaling, latency, fine-tuning, agentic, reasoning
• MediX-R1: Open Ended Medical Reinforcement Learning score 9
  入选 HF Daily Papers；HF 热度: 16 upvotes (+3)；有代码实现；关键词(4): lightweight, reasoning, vision-language, open-source

也值得关注

• S2O: Early Stopping for Sparse Attention via Online Permutation score 4
  机构: Zhejiang University；关键词(2): lightweight, latency
• Coded-E2LF: Coded Aperture Light Field Imaging from Events score 4
  关键词(1): coding；顶会接收: CVPR
• Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache score 4
  关键词(1): deployment；顶会接收: CVPR
• No Caption, No Problem: Caption-Free Membership Inference via Model-Fitted Embeddings score 4
  关键词(2): vision-language, text-to-image；顶会接收: ICLR
• Replacing Multi-Step Assembly of Data Preparation Pipelines with One-Step LLM Pipeline Generation for Table QA score 4
  机构: Cornell；关键词(2): lightweight, compression
• HulluEdit: Single-Pass Evidence-Consistent Subspace Editing for Mitigating Hallucinations in Large Vision-Language Models score 4
  关键词(2): deployment, vision-language；顶会接收: CVPR
• Towards Better RL Training Data Utilization via Second-Order Rollout score 4
  机构: Peking University；关键词(1): reasoning
• Hierarchy-of-Groups Policy Optimization for Long-Horizon Agentic Tasks score 4
  关键词(2): GRPO, agentic；顶会接收: ICLR
• Towards LLM-Empowered Knowledge Tracing via LLM-Student Hierarchical Behavior Alignment in Hyperbolic Space score 4
  关键词(1): synthetic data；顶会接收: AAAI
• OpenFS: Multi-Hand-Capable Fingerspelling Recognition with Implicit Signing-Hand Detection and Frame-Wise Letter-Conditioned Synthesis score 4
  关键词(1): open-source；顶会接收: CVPR
• SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling score 4
  关键词(1): vision-language；顶会接收: CVPR
• TriLite: Efficient Weakly Supervised Object Localization with Universal Visual Features and Tri-Region Disentanglement score 4
  关键词(2): fine-tuning, pre-training；顶会接收: CVPR
• A Mixture-of-Experts Model for Multimodal Emotion Recognition in Conversations score 4
  机构: Google；关键词(1): MoE
• Understanding Usage and Engagement in AI-Powered Scientific Research Tools: The Asta Interaction Dataset score 4
  机构: Allen Institute；关键词(1): retrieval-augmented
• Scale Can't Overcome Pragmatics: The Impact of Reporting Bias on Vision-Language Reasoning score 4
  机构: AI2；关键词(4): scaling, reasoning, vision-language, data curation