CORTEXA
← Browse
arxivcs.AIcs.LG2026-07-06

Rethinking On-Policy Self-Distillation for Thinking Models

Simran Kaur, Narutatsu Ri, Yinghui He, Liam Fowl, Sanjeev Arora

Self-distillation is a promising recipe for self-improvement in language models. In this setting, a model can serve as its own teacher when given privileged information, such as a solution to a math problem. This seems especially appealing for thinking models, which can use test-time reasoning to absorb the privileged information. Surprisingly, we show that privileged self-distillation degrades thinking models on long reasoning traces: across five Qwen3 and OLMo thinking models evaluated on AIME24, AIME25, and HMMT25, privileged-context distillation causes a relative drop of up to 17% in avg@16 accuracy. The degradation scales with the amount of privileged context withheld from the student and is most pronounced at long rollout budgets, where thinking models otherwise obtain their largest gains. This failure mode is not specific to self-distillation: on-policy distillation (OPD) improves thinking models, but privileged OPD reverses these gains. Our diagnostics link this failure mode to how privileged teacher context reshapes learning at high-entropy forking positions, where multiple continuations remain plausible and may lead to different reasoning paths. Privileged context lowers fork rates in thinking-model rollouts but not in instruction-model rollouts. This leads to an interesting dichotomy, where privileged context can help instruction-tuned models but hurts stronger thinking models. The effect is visible when the student begins a self-correction branch, where privileged OPD penalizes sampled reconsideration tokens that vanilla OPD supports. Thinking models trained with a privileged teacher produce fewer verification, backtracking, and hedging markers, even after length normalization. These findings indicate that self-distillation for strong thinking models requires attention to token-level signal, especially around correction and reasoning steps.

View free PDFSource page

Related papers

arxivcs.CVcs.AIcs.RO2026-07-24

SM4RT: Learning Structured Motion Geometry for 4D Reconstruction

Shing Ho J. Lin, Wenzhao Zheng, Dong Zhuo, Yuqi Wu, Jie Zhou, Jiwen Lu

Geometry Foundation Models (GFMs) have substantially advanced monocular 3D reconstruction, yet extending this capability to 4D dynamic understanding remains a fundamental challenge. Most existing motion perception methods (e.g., sparse tracking, dense point-wise flow) treat motio…

View free PDFSource page
arxivcs.AI2026-07-24Cited by 2

Explainable Reinforcement Learning for assisting Air Traffic Controllers

Anduel Mehmeti, Gabriella Gigante, Salvatore Venticinque

To effectively integrate AI into high-stakes, critical environments such as healthcare, autonomous driving, and aviation--and to advance toward higher levels of automation and seamless human-AI collaboration--building trust in AI-driven solutions is essential. Trust, in turn, is…

View free PDFSource page
arxivcond-mat.dis-nncond-mat.stat-mechcs.AInlin.CD2026-07-24

Multiplicity of Stable Attractors in Disordered Neural Models

Raffaele Marino, Roberto Livi, Antonio Politi

We show how large-deviation statistics allows one to obtain reliable estimates of the multiplicity of stable fixed-points in a model of neural ordinary differential equations previously employed in computational tasks. The result is obtained by developing a suitable perturbative…

View free PDFSource page