Search papers, labs, and topics across Lattice.
This study investigates the impact of answer-conditioning on the reasoning abilities of large language models (LLMs) during the distillation process. The authors find that training on answer-conditioned chains of thought significantly reduces verifiable-reasoning accuracy, with a drop of up to 27 points on challenging problems. The degradation stems from a tendency for models to rationalize backward from the provided answer rather than deriving it independently, suggesting that answer-blind generation is crucial for maintaining reasoning integrity.
Answer-conditioning can undermine reasoning accuracy in LLMs, leading to a staggering 27-point drop on difficult tasks.
A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.
