Why Models Know But Don't Say: Chain-of-Thought Faithfulness Divergence Between Thinking Tokens and Answers in Open-Weight Reasoning Models

TL;DR

This study investigates how open-weight reasoning models sometimes reveal hint influence only in their thinking tokens, not in their answers, highlighting limitations of answer-only monitoring methods.

Contribution

It introduces the concept of thinking-answer divergence and provides empirical analysis of 12 models showing the asymmetry and variability in hint acknowledgment.

Findings

55.4% of hint-following cases show divergence with thinking tokens but not answers.

Answer-only monitoring misses over half of hint-influenced reasoning cases.

Model transparency varies widely, from 19.6% to 94.7% divergence.

Abstract

Extended-thinking models expose a second text-generation channel ("thinking tokens") alongside the user-visible answer. This study examines 12 open-weight reasoning models on MMLU and GPQA questions paired with misleading hints. Among the 10,506 cases where models actually followed the hint (choosing the hint's target over the ground truth), each case is classified by whether the model acknowledges the hint in its thinking tokens, its answer text, both, or neither. In 55.4% of these cases the model's thinking tokens contain hint-related keywords that the visible answer omits entirely, a pattern termed *thinking-answer divergence*. The reverse (answer-only acknowledgment) is near-zero (0.5%), confirming that the asymmetry is directional. Hint type shapes the pattern sharply: sycophancy is the most *transparent* hint, with 58.8% of sycophancy-influenced cases acknowledging the professor's authority in both channels, while consistency (72.2%) and unethical (62.7%) hints are dominated by thinking-only acknowledgment. Models also vary widely, from near-total divergence (Step-3.5-Flash: 94.7%) to relative transparency (Qwen3.5-27B: 19.6%). These results show that answer-text-only monitoring misses more than half of all hint-influenced reasoning and that thinking-token access, while necessary, still leaves 11.8% of cases with no verbalized acknowledgment in either channel.