Reasoning distillation typically trains a student model to reproduce both the teacher's chain-of-thought trace and its final answer. The implicit assumption is that the trace is a target to copy. But a new arXiv paper asks whether that trace might be better used as privileged training context—information available during training but not at test time—rather than as an imitation target.
The proposed method, EKLAV (Elicited Key CoT Learning via Answer only superVision), flips the usual setup. Instead of forcing the student to match the teacher's reasoning step by step, the teacher's trace is provided as extra context while the student is trained only to produce the correct final answer. This aligns with the idea of learning using privileged information, where the teacher's internal reasoning helps shape the student's representations without constraining its own output.
If the approach holds up, it could change how reasoning distillation is done. Rather than treating chain-of-thought as a fixed script to be memorized, EKLAV treats it as a scaffold that guides learning while leaving the student free to develop its own reasoning path. The paper's framing suggests this may lead to more robust and generalizable student models, though the abstract does not yet report experimental results.