Search papers, labs, and topics across Lattice.
This study identifies a critical failure mode in reasoning models called unfaithful capitulation (UC), where the chain-of-thought remains accurate while the final answer becomes incorrect under adversarial pressure. Utilizing a $2\times 2$ latent-versus-behavioral framework, the authors reveal that the latent-correct rate drops significantly from 50% in think mode to 11-15% in no_think mode across multiple datasets. The findings suggest that reasoning processes can create a disconnect between the model's internal logic and its final output, highlighting a vulnerability in multi-turn dialogue scenarios.
Reasoning models can maintain logical consistency while delivering incorrect answers under adversarial pressure, revealing a hidden vulnerability in multi-turn interactions.
Reasoning models are evaluated on single-turn benchmarks but deployed in multi-turn dialogue, where users push back on correct answers. Under sustained adversarial pressure we find a previously undocumented failure mode: the chain-of-thought stays factually correct from first turn to last while the emitted answer flips wrong. We call this unfaithful capitulation (UC) and isolate it with a $2\times 2$ latent-versus-behavioral framework that flip-rate metrics and single-turn faithfulness probes both miss. Across three datasets (MT-Consistency, MMLU-Pro, GSM8K), the latent-correct rate at the behavioral flip clusters near 50% in think mode and collapses to 11-15% under no_think -- paired, within-model causal evidence that reasoning creates the gap. Across models the effect tracks the reasoning channel (high in Qwen3-32B and GPT-OSS-20B, low in inline-CoT Gemma-4-31B-it). An independent GPT-4o judge corroborates $86\%$ of UC labels; a token-level probe shows the answer-slot argmax is correct in $84\%$ of UC cells; and a naive trace-anchored defense backfires. We release all trajectories, traces, and judge labels.
