8%的token决定reasoning差距

重点关注

• WOW-Seg: A Word-free Open World Segmentation Model score 6
  机构: Meta；顶会接收: ICLR

也值得关注

• TRACE: Evidence Grounding-Guided Multi-Video Event Understanding and Claim Generation score 4
  关键词(3): reasoning, vision-language, leaderboard；顶会接收: ACL
• The Unlearnability Phenomenon in RLVR for Language Models score 4
  关键词(1): reasoning；顶会接收: ICML
• HAD: Hallucination-Aware Diffusion Priors for 3D Reconstruction score 4
  关键词(1): reasoning；顶会接收: CVPR
• Reasoning Can Be Restored by Correcting a Few Decision Tokens score 4
  关键词(1): reasoning；顶会接收: ICML
• SE-GA: Memory-Augmented Self-Evolution for GUI Agents score 4
  关键词(1): open source；顶会接收: ICML
• D$^2$Evo: Dual Difficulty-Aware Self-Evolution for Data-Efficient Reinforcement Learning score 4
  关键词(1): reasoning；顶会接收: ICML
• New Wide-Net-Casting Jailbreak Attacks Risk Large Models score 4
  关键词(1): jailbreak；顶会接收: ICML
• The Geometry of Projection Heads: Conditioning, Invariance, and Collapse score 4
  关键词(1): pretraining；顶会接收: ICML
• PluRule: A Benchmark for Moderating Pluralistic Communities on Social Media score 4
  关键词(2): reasoning, vision-language；顶会接收: ACL
• Controlling Decision Drift in Multimodal Sentiment Analysis with Missing Modalities score 3
  顶会接收: IJCAI
• Multi-task Linear Regression without Eigenvalue Lower Bounds: Adaptivity, Robustness and Safety score 3
  顶会接收: ICML