微调即加速1.7倍，token崩的不是channel

重点关注

• MARS: Enabling Autoregressive Models Multi-Token Generation score 10
  入选 HF Daily Papers；HF 热度: 26 upvotes (+4)；有代码实现；关键词(7): lightweight, deployment, serving, latency, throughput
• INSPATIO-WORLD: A Real-Time 4D World Simulator via Spatiotemporal Autoregressive Modeling score 10
  入选 HF Daily Papers；HF 热度: 24 upvotes (+4)；有代码实现；关键词(3): distillation, real-time, synthetic data
• TC-AE: Unlocking Token Capacity for Deep Compression Autoencoders score 9
  入选 HF Daily Papers；HF 热度: 11 upvotes (+3)；有代码实现；关键词(2): scaling, compression
• FlowInOne:Unifying Multimodal Generation as Image-in, Image-out Flow Matching score 8
  入选 HF Daily Papers；HF 热度: 6 upvotes (+2)；有代码实现；关键词(2): text-to-image, open-source
• MoRight: Motion Control Done Right score 8
  机构: NVIDIA；入选 HF Daily Papers；HF 热度: 3 upvotes (+1)；关键词(1): reasoning
• FP4 Explore, BF16 Train: Diffusion Reinforcement Learning via Efficient Rollout Scaling score 7
  入选 HF Daily Papers；HF 热度: 12 upvotes (+3)；关键词(5): scaling, quantization, throughput, post-training, text-to-image
• Personalized RewardBench: Evaluating Reward Models with Human Aligned Personalization score 8
  入选 HF Daily Papers；HF 热度: 7 upvotes (+2)；有代码实现；关键词(2): serving, PPO
• Fast Spatial Memory with Elastic Test-Time Training score 6
  入选 HF Daily Papers；HF 热度: 4 upvotes (+1)；有代码实现
• Q-Zoom: Query-Aware Adaptive Perception for Efficient Multimodal Large Language Models score 7
  入选 HF Daily Papers；HF 热度: 2 upvotes (+1)；有代码实现；关键词(5): scaling, lightweight, distillation, throughput, fine-tuning

也值得关注

• To Lie or Not to Lie? Investigating The Biased Spread of Global Lies by LLMs score 4
  关键词(1): retrieval-augmented；顶会接收: ACL
• SubFLOT: Submodel Extraction for Efficient and Personalized Federated Learning via Optimal Transport score 4
  关键词(4): scaling, pruning, deployment, edge；顶会接收: CVPR
• Feedback Adaptation for Retrieval-Augmented Generation score 4
  关键词(3): deployment, retrieval-augmented, RAG；顶会接收: ACL
• VDPP: Video Depth Post-Processing for Speed and Scalability score 4
  关键词(3): deployment, real-time, edge；顶会接收: CVPR
• ChemVLR: Prioritizing Reasoning in Perception for Chemical Vision-Language Understanding score 4
  关键词(4): production, reasoning, vision-language, open-source；顶会接收: ACL
• Reasoning Fails Where Step Flow Breaks score 4
  关键词(2): coding, reasoning；顶会接收: ACL
• StructKV: Preserving the Structural Skeleton for Scalable Long-Context Inference score 4
  关键词(2): compression, pruning；顶会接收: ACL
• MoBiE: Efficient Inference of Mixture of Binary Experts under Post-Training Quantization score 4
  关键词(3): quantization, post-training, MoE；顶会接收: ACL
• Generate, Analyze, and Refine: Training-Free Sound Source Localization via MLLM Meta-Reasoning score 4
  关键词(1): reasoning；顶会接收: CVPR
• WRAP++: Web discoveRy Amplified Pretraining score 4
  机构: Tencent；关键词(4): scaling, pretraining, reasoning, synthetic data
• Do We Need Distinct Representations for Every Speech Token? Unveiling and Exploiting Redundancy in Large Speech Language Models score 4
  关键词(2): compression, deployment；顶会接收: ACL
• iTAG: Inverse Design for Natural Text Generation with Accurate Causal Graph Annotations score 4
  关键词(1): reasoning；顶会接收: ACL
• Making MLLMs Blind: Adversarial Smuggling Attacks in MLLM Content Moderation score 4
  关键词(3): scaling, reasoning, open-source；顶会接收: ACL
• Learning to Query History: Nonstationary Classification via Learned Retrieval score 4
  关键词(2): scaling, deployment；顶会接收: ICLR