一条证据压成1个token，生成省3-10倍

重点关注

• How Does Reasoning Flow? Tracing Attention-Induced Information Flow for Targeted RL in LLMs score 9
  入选 HF Daily Papers；HF 热度: 6 upvotes (+2)；关键词(3): throughput, edge, reasoning；顶会接收: ICML
• SCAIL-2: Unifying Controlled Character Animation with End-to-end In-Context Conditioning score 10
  入选 HF Daily Papers；HF 热度: 37 upvotes (+4)；有代码实现；关键词(2): DPO, synthetic data
• One Token per Multimodal Evidence: Latent Memory for Resource-Constrained QA score 9
  入选 HF Daily Papers；HF 热度: 14 upvotes (+3)；有代码实现；关键词(3): distillation, RAG, vision-language
• Dynamic Linear Attention score 7
  入选 HF Daily Papers；HF 热度: 3 upvotes (+1)；顶会接收: ICML
• Test-Time Gradient Guidance of Flow Policies in Reinforcement Learning score 7
  入选 HF Daily Papers；HF 热度: 3 upvotes (+1)；有代码实现；关键词(2): scaling, pre-training
• ARM: An AutoRegressive Large Multimodal Model with Unified Discrete Representations score 10
  入选 HF Daily Papers；HF 热度: 23 upvotes (+4)；有代码实现；关键词(2): vision-language, text-to-image

也值得关注

• When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models score 4
  入选 HF Daily Papers；关键词(2): reasoning, jailbreak
• Dissect and Prune: Enhancing Robustness in AI-Generated Image Detection score 4
  关键词(1): compression；顶会接收: ICML
• Parallel Causal Associative Fields: Gated Sparse Memory for Long-Context Language Modeling score 4
  机构: NVIDIA；关键词(1): pretraining
• UniSVQ: 2-bit Unified Scalar-Vector Quantization score 4
  关键词(5): quantization, deployment, throughput, fine-tuning, post-training；顶会接收: ICML
• Machine Learning Methods for Studying Latent Neural Activity Dynamics score 4
  关键词(1): pre-training；顶会接收: IJCAI
• LC-QAT: Data-Efficient 2-Bit QAT for LLMs via Linear-Constrained Vector Quantization score 4
  关键词(3): quantization, deployment, post-training；顶会接收: ICML
• ActiveMem: Distributed Active Memory for Long-Horizon LLM Reasoning score 4
  机构: Huawei；关键词(4): scaling, lightweight, pruning, reasoning
• Embedding Hybrid Systems into Continuous Latent Vector Fields score 4
  关键词(1): state space；顶会接收: ICML
• Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields score 7
  入选 HF Daily Papers；HF 热度: 19 upvotes (+3)；关键词(1): agentic
• WorldOlympiad: Can Your World Model Survive a Triathlon? score 13
  机构: Alibaba；入选 HF Daily Papers；HF 热度: 29 upvotes (+4)；有代码实现；关键词(3): reasoning, robotics, embodied
• AnyMod-LLVE: Low-Light Video Enhancement with Modality-Agnostic Inference score 4
  关键词(1): pretraining；顶会接收: ICML
• ERAlign: Energy-based Representation Alignment of GNNs and LLMs on Text-attributed Graphs score 3
  顶会接收: ICML
• SSR-Merge: Subspace Signal Routing for Training-Free LoRA Merging in Diffusion Models score 3
  顶会接收: ICML
• IPSM-Bench: A New Intermediate Phase Segmentation Benchmark in Microstructure Images of Zinc-Based Absorbable Biomaterials score 3
  顶会接收: IJCAI
• MOFA-VTON: More Fashion Possibilities with Fine-Grained Adaptations in Virtual Try-On score 3
  顶会接收: CVPR