Know2Say: Unveiling the Gap Between What Reasoning Models 'Know' and What They 'Say'

The Know2Say study focuses on the 'Detection-Extraction Gap' phenomenon in the reasoning process of large language models—models internally 'know' the answer early in reasoning, but are prone to errors when forced to extract it immediately. Based on this, the study proposes the Black-box Adaptive Early Exit (BAEE) strategy, which reduces reasoning costs by 70-85% while improving accuracy, and is applicable to closed-source models like GPT-4.

As the complex reasoning capabilities of large language models improve, Chain-of-Thought (CoT) prompting has become a standard technique, but it comes with high computational costs due to numerous intermediate steps. Core question: Does the model need to generate all steps to 'know' the answer? The Know2Say study found that models form answers internally early on, but are prone to errors when forced to answer immediately—this is defined as the 'Detection-Extraction Gap'.

Detection-Extraction Gap refers to: In the early stages of CoT, models can detect the high probability of an answer through free continuation (detection), but have low accuracy when forced to answer immediately (extraction). Researchers formalized the modeling using total variation distance (lower bound of the distance between the free continuation distribution P_free and the forced extraction distribution P_forced), providing a mathematical foundation for improvement strategies.

The core mechanisms of BAEE (Black-box Adaptive Early Exit) are: 1. Pause generation at preset checkpoints in CoT; 2. Sample continuations to check answer consistency (PSC metric); 3. If consistency exceeds a threshold (e.g., 0.75), exit early and return the majority answer. Its black-box nature does not require access to the model's internals, making it applicable to closed-source models like GPT-4 and Claude.

In benchmark tests like MATH-500 and GPQA-Diamond, BAEE achieves: 70-85% reduction in sequence generation, 1-5 percentage points improvement in accuracy, and 52-88% of tokens accounted for after the commitment point. Overly long CoT may cause the model to deviate from the correct path; appropriate early exit instead maintains a clear reasoning main thread.

Theoretical Contributions: Revealing the asymmetry between internal knowledge and external expression, the optimality of CoT length, and the feasibility of black-box optimization. Practical Value: Providing developers with ready-to-use optimization solutions to reduce API costs, improve latency, and enhance quality. Methodologically, it demonstrates a rigorous paradigm from phenomenon observation to practical solutions.

Current Limitations: PSC threshold requires task-specific tuning, additional overhead from sampling costs, reduced benefits for complex long-dependency tasks. Future Directions: Adaptive threshold strategies, efficient PSC estimation, multimodal expansion, white-box version of early exit mechanisms.