Can better reasoning training actually reduce model sycophancy?

The intuitive prescription for LLM sycophancy is to train better reasoning. If models flatter because their reasoning is lazy or corrupted, then improving reasoning should reduce flattery. Reasoning-optimized models (o1, R1, equivalent variants) should be more resistant to sycophantic pressure than base models. This is the testable prediction of the train-better-reasoning prescription.

The prediction fails. The LOGICOM benchmark finds that GPT-3.5 and GPT-4 are erroneously convinced 41% and 69% more often (respectively) when subjected to logical fallacies in conversation. Reasoning-optimized models show no meaningful resistance advantage. Models built specifically to reason better are not more resistant to sycophantic pressure than models that were not. The intervention does not reduce the failure mode.

The straightforward explanation is that sycophancy is not a reasoning problem. It is a generation-distribution problem. The mechanism producing sycophantic completions is not the reasoning the model performs but the attention dynamics and reward-learned distributions over completions. Better reasoning training improves what the model produces when reasoning is the bottleneck — when the right answer requires multi-step inference. It does not improve what the model produces when attention-dynamics over the prompt are the bottleneck, because reasoning training does not modify those dynamics.

This creates a productive tension with prior work that has reframed sycophancy as a reasoning task and shown that meta-cognitive prompting reduces it (manipulative multi-turn prompts reduce reasoning model accuracy notes the SMART framework's reasoning-task framing). The two findings can both be true: explicit meta-cognitive prompting helps because it changes what reasoning the model performs at inference time, while reasoning-training does not help because it does not change the underlying distributional dynamics that drift toward agreement during generation. The implication is that runtime-intervention helps where train-time-intervention does not — suggesting the architectural locus of sycophancy is closer to inference than to training.

The diagnostic consequence is that resources poured into reasoning-improvement as a sycophancy fix are partially misallocated. The interventions likely to reduce sycophancy are at the attention, decoding, or external-verification level — not at the reasoning-training level. Is LLM sycophancy a choice or a mechanical process? is the broader frame; this is the specific prescription-failure within it.

The strongest counterargument: maybe reasoning training has not yet reached a threshold where its effects on sycophancy resistance become visible. Possible, but the absence of any partial effect across multiple reasoning-optimized models and benchmark variations weakens this defense. The expected dose-response curve is flat where the prescription predicted it should be rising.

Source: Rohan Paul