Preprint proposes steering-vector self-distillation for reasoning models

A preprint posted to arXiv on 29 September 2026 describes Activation-Conditioned Self-Distillation (ACSD), a way to train reasoning models using their own verified outputs. The authors are Zhexi Lu, Subhajit Chaudhury, Tejaswini Pedapati, Keerthiram Murugesan and Lei Yu.

On-policy self-distillation uses a model as its own teacher to give dense, token-level supervision, often by showing the teacher a reference solution. The authors argue that supplying such privileged information does not by itself guarantee effective supervision across long responses.

ACSD takes a different route. It collects trajectories the model generated itself and separates those that reach a verified correct answer within a generation budget from all the remaining ones. The contrast between their internal activations yields a steering vector. A frozen copy of the base model applies that vector at each prediction position, and the student learns from next-token distributions on prefixes it generated.

According to the authors, outcome verification is used to build and calibrate the direction, but distillation needs neither problem-specific reference text nor updates to the teacher's parameters. Only the distilled student is used at inference.

The reported results are the paper's own. The authors say ACSD gave the highest mean accuracy over four mathematical benchmarks on each of five models among the methods they evaluated. On DeepSeek-R1-0528-Qwen3-8B, they report 71.9% mean mathematical accuracy and 70.9% on LiveCodeBench v6 pass@12, against 69.0% and 66.3% for the reference-conditioned OPSD baseline. They also say contrasts among correct trajectories can support distillation, that extracted directions can be reused across mathematical training datasets, and that ACSD kept late-position logit-update magnitudes more stable than OPSD on fixed student trajectories.

What is confirmed is limited to what the authors state in the abstract. We found no independent coverage, replication or peer review, and only the abstract page was reviewed here, not the full paper, so details such as the compared methods and experimental settings are unverified. The headline numbers are for a named model and specific benchmarks, mostly mathematical plus one coding test, and do not show how the method performs elsewhere.

Why it matters: if the results hold, the approach would let developers turn answer-checking signals into dense training guidance without writing worked reference solutions for every problem. That remains a claim to be tested by outside groups. This item rests on a single source, the authors' own preprint.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.