Consensus-based self-distillation extracts token-level training signals from LLM agreement
Researchers introduce CANON, a self-distillation method that extracts token-level supervision from consensus across multiple LLM solution samples, moving beyond existing approaches that treat agreement as a binary filter or scalar signal. The technique conditions a frozen model snapshot on majority answers to generate dense training targets without human labels, addressing a fundamental inefficiency in how current consensus-based methods leverage the information contained in agreeing outputs. This work matters for practitioners scaling reasoning models on unlabeled data, where label-free supervision remains a bottleneck for cost-effective improvement.
Modelwire context
ExplainerThe key distinction CANON draws is granularity: rather than asking whether a majority of sampled outputs agree on a final answer, it asks what the token-level distribution across agreeing outputs actually looks like, treating that distribution as a dense training target. That shift from answer-level to token-level supervision is where the efficiency claim lives.
This connects directly to the post-training calibration paper covered the same day ('Post-Training Shifts Confidence: A Three-Stage Analysis'), which mapped how SFT, RL, and on-policy distillation each shape confidence at different reasoning stages. That paper found on-policy distillation excels at pre-reasoning difficulty estimation, which is precisely the regime where CANON's richer consensus signal could improve the quality of what gets distilled. Both papers are circling the same practical problem: how to extract more usable supervision from model-generated outputs without human annotation. Together they suggest a convergence toward hybrid approaches that combine calibrated confidence signals with denser token-level targets.
Watch whether CANON's token-level gains hold when the frozen snapshot model and the sampling model diverge significantly in size, since that gap is where the consensus signal is most likely to degrade into noise rather than useful supervision.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsCANON
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Consensus as Privileged Context for Label-Free Self-Distillation”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.