Two-thirds of teacher-student agreement masks trajectory failure in distillation
Researchers expose a critical blind spot in on-policy distillation, where student models learn from teacher feedback at the token level without accounting for whether the final answer was correct. The study reveals that two-thirds of agreement between student and teacher occurs on trajectories that ultimately fail, making naive imitation actively harmful. This outcome-confounding effect persists across multiple model pairs, suggesting that current distillation methods conflate local agreement with genuine learning signals. The finding reshapes how practitioners should interpret teacher supervision in reasoning tasks, requiring outcome-aware filtering rather than pointwise confidence matching.
Modelwire context
ExplainerThe paper doesn't just identify that on-policy distillation can fail; it quantifies the mechanism: two-thirds of student-teacher agreement happens on wrong trajectories. This means the problem isn't rare edge cases but structural to how token-level supervision works without outcome filtering.
This connects directly to the test-time adaptation work from the same day, which flagged how teacher-student frameworks can drift when supervision signals aren't properly validated. Here we see the inverse problem: even at training time, naive imitation from a teacher creates a false confidence in alignment. Both papers expose that proximity in model behavior (agreement, momentum stability) can mask actual performance degradation. The VAE soft-constraint work also shares this tension: local optimization metrics (latent capacity, pointwise confidence) can conflict with what actually matters downstream (disentanglement, correct answers).
If practitioners applying this finding to Qwen3-32B distillation report measurable accuracy gains by filtering training data to only outcome-correct trajectories, that validates the fix. If the same two-thirds failure rate appears in other model families (GPT, Claude, Llama) tested by independent groups within six months, that confirms this is a systematic property of on-policy distillation rather than a Qwen-specific artifact.
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.
MentionsQwen3-8B · Qwen3-32B · Qwen2.5-7B · Qwen2.5-32B
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.LG originally reported this story as “Outcome-Confounded Local Supervision in On-Policy 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.