Mechanism by Which Reasoning Distillation Affects Steering

Determine whether reasoning distillation changes the conditional distributions governing logits steering in a way that explains the reduced or negative incremental benefit of the M-step for DeepSeek-R1-0528-Qwen3-8B, using matched pre- and post-distillation checkpoints, contexts, prefixes, and decoding budgets.

Background

The paper observes that DeepSeek-R1-0528-Qwen3-8B improves substantially over its base model with EMO but exhibits a negative mean incremental contribution from the M-step relative to context-only TextGrad. The authors propose that reasoning distillation may alter the conditional distributions on which the steering procedure operates, including the softmax Jacobian and context-induced logit displacement.

This explanation is unresolved because the evaluated checkpoints do not form a controlled pre- and post-distillation pair. Establishing the mechanism would require matched checkpoints, contexts, prefixes, and decoding budgets, together with measurements of conditional entropy, context-induced logit displacements, and reference-token probability changes.

References

A possible explanation is that distillation changes the conditional distributions on which steering operates. In Eq.~\ref{eq:context_displacement}, both the softmax Jacobian $\mathbf J_{P_{\mathrm{opt}}$ and the context-induced displacement $\Delta\mathbf{z}_t$ depend on the checkpoint; changes to either can alter the steering vector even with fixed $\mu$ and $\alpha$. Our Qwen3-8B comparison is not a controlled pre-/post-distillation pair, so this remains a mechanistic hypothesis. Testing it would require matched checkpoints, contexts, prefixes, and decoding budgets, measuring conditional entropy, context-induced logit displacements, and reference-token probability changes.

— EMO: A Closed-Loop Test-Time Computation Framework via Alternating Context Optimization and Logits Steering  (2610.22160 - Xia et al., 15 Oct 2026) in Appendix, Section "Model Selection," paragraph "Reasoning distillation"