PPL
Albert No
Papers (5)
The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion ModelsWhere Rollouts Begin: Low-Load, High-Leverage First-Token Diversification for RLVRVLMs Trace Without Tracking: Diagnosing Failures in Visual Path FollowingSlower Generalization, Faster Memorization: A Sweet Spot in Algorithmic LearningFew-Shot Truly Benign DPO Attack for Jailbreaking LLMs