Xu He

The paper introduces the Prompt Overflow Attack, demonstrating that guardrail models inspecting truncated or segmented inputs fail to detect malicious instructions that are only actionable when the full, overlong context is provided to the downstream LLM.

The paper analyzes the entropy dynamics of Chain-of-Thought (CoT) reasoning, identifying a transition from an exploratory Uncertainty Region to a stable Confidence Region, which enables superior early exit and test-time scaling strategies.