MoE安全集中在少数专家、独占批提速42%

重点关注

• RoboStressBench: Benchmarking VLM Robustness to Physical Visual Stress in Embodied Scenes score 7
  入选 HF Daily Papers；HF 热度: 3 upvotes (+1)；有代码实现；关键词(4): agentic, reasoning, vision-language, embodied

也值得关注

• Threshold-Based Exclusive Batching for LLM Inference score 4
  关键词(1): throughput；顶会接收: ICML
• The Assistant as a Privileged Persona: A canonical reference in cross-persona self-recognition score 4
  机构: Anthropic；关键词(1): post-training
• CAFOSat: A Strongly Annotated Dataset for Infrastructure-Aware CAFO Mapping Using High-Resolution Imagery score 4
  关键词(2): production, vision-language；顶会接收: CVPR
• Revisiting Parameter-Based Knowledge Editing in Large Language Models: Theoretical Limits and Empirical Evidence score 4
  关键词(1): reasoning；顶会接收: ICML
• FlowNar: Scalable Streaming Narration for Long-Form Videos score 4
  关键词(3): deployment, throughput, real-time；顶会接收: ICML
• MESA: Improving MoE Safety Alignment via Decentralized Expertise score 4
  关键词(1): MoE；顶会接收: ICML
• Demystifying the Optimal Fair Classifier in Multi-Class Classification score 4
  关键词(1): fine-tuning；顶会接收: ICML
• COPF: An Online Framework for Deployment-Stable Counterfactual Fairness in Evolving Graphs score 4
  关键词(1): deployment；顶会接收: ICML
• Ryze: Evidence-Enriched Data Synthesis from Biomedical Papers score 4
  关键词(4): fine-tuning, post-training, open source, synthetic data；顶会接收: ACL
• On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance score 3
  顶会接收: ICML
• Improving Visual Grounding in Remote Sensing via Cluster-Guided Refinement and Model Ensemble Voting score 3
  顶会接收: CVPR
• Semi-Supervised Noise Adaptation: Transferring Knowledge from Noise Domain score 3
  顶会接收: ICML
• SORA: Free Second-Order Attacks in Fast Adversarial Training score 3
  顶会接收: ICML
• Generative Diffusion Priors for 3D Mapping of the Dark Universe score 3
  顶会接收: CVPR