# Diagnosing and Mitigating Thinking Collapse in On-Policy Self-Distillation

arXiv: https://arxiv.org/abs/2607.10805 · HTML: https://mlsanity.com/paper/2607.10805/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning.
- Builds on 24 of its 36 references that are in the map.
- First posted on arXiv 12 Jul 2026.

## Placed in

- LLM reasoning / [Prompting and chain-of-thought reasoning](https://mlsanity.com/lr/prompting-and-chain-of-thought-reasoning/)
- LLM alignment / RLHF / preference optimisation / [Process reward models for reasoning](https://mlsanity.com/alignment/process-reward-models-for-reasoning/)
- Knowledge distillation / [On-policy and self-distillation for LLMs](https://mlsanity.com/distill/on-policy-and-self-distillation-for-llms/)

## Builds on (most cited first, 24 of 24)

- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters](https://mlsanity.com/paper/2408.03314/) (2024)
- [On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes](https://mlsanity.com/paper/2306.13649/) (2023)
- [Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models](https://mlsanity.com/paper/2601.18734/) (2026)
- [Reinforcement Learning via Self-Distillation](https://mlsanity.com/paper/2601.20802/) (2026)
- [Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe](https://mlsanity.com/paper/2604.13016/) (2026)
- [Self-Distillation Enables Continual Learning](https://mlsanity.com/paper/2601.19897/) (2026)
- [Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes](https://mlsanity.com/paper/2603.25562/) (2026)
- [Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs](https://mlsanity.com/paper/2605.00674/) (2026)
- [Self-Distilled RLVR](https://mlsanity.com/paper/2604.03128/) (2026)
- [TIP: Token Importance in On-Policy Distillation](https://mlsanity.com/paper/2604.14084/) (2026)
- [Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?](https://mlsanity.com/paper/2603.24472/) (2026)
- [Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision](https://mlsanity.com/paper/2604.12002/) (2026)
- [The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes](https://mlsanity.com/paper/2605.11182/) (2026)
- [Purified OPSD: On-Policy Self-Distillation Without Losing How to Think](https://mlsanity.com/paper/2607.02234/) (2026)
- [Revisiting Overthinking in Long Chain-of-Thought from the Perspective of Self-Doubt](https://mlsanity.com/paper/2505.23480/) (2025)
- [Rebellious Student: Reversing Teacher Signals for Reasoning Exploration with Self-Distilled RLVR](https://mlsanity.com/paper/2605.10781/) (2026)
- [Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information](https://mlsanity.com/paper/2605.11609/) (2026)
- [Think Dense, Not Long: Dynamic Decoupled Conditional Advantage for Efficient Reasoning](https://mlsanity.com/paper/2602.02099/) (2026)
- [Less is More: Early Stopping Rollout for On-Policy Distillation](https://mlsanity.com/paper/2605.27028/) (2026)
- [Better, Faster: Harnessing Self-Improvement in Large Reasoning Models](https://mlsanity.com/paper/2605.24998/) (2026)
- [Reason-KE++: Aligning the Process, Not Just the Outcome, for Faithful LLM Knowledge Editing](https://mlsanity.com/paper/2511.12661/) (2025)

## Built on by (most cited first, 5 of 5)

- [Agentic Reinforcement Learning with Observation-Calibrated Self-Distillation](https://mlsanity.com/paper/2608.04788/) (2026)
- [Privileged Likelihood Is Not Automatically Value: Three Checks for Token Credit in On-Policy Self-Distillation](https://mlsanity.com/paper/2608.09263/) (2026)
- [VISTA: Verifier-Informed Student-to-Teacher Adaptation for On-Policy Self-Distillation](https://mlsanity.com/paper/2608.28306/) (2026)
- [iSDFT: Information-Proximal Self-Distillation for Continual Learning in LLMs](https://mlsanity.com/paper/2609.24646/) (2026)
- [The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models](https://mlsanity.com/paper/2610.02191/) (2026)

## Abstract (arXiv)

On-Policy Self-Distillation (OPSD) has emerged as a crucial paradigm for enhancing and aligning Large Language Models (LLMs). However, in complex reasoning tasks, OPSD paradoxically degrades downstream performance. In this paper, we systematically investigate this pathology and identify a severe optimization trap we define as Thinking Collapse -- a sharp decline in the model's native intermediate reasoning behavior, measured by epistemic-token density (ET per 1k). Through entropy-based gradient masking and token-level target analysis, we show that this collapse is triggered by aggressive teacher gradients at high-student-entropy decision forks, where student epistemic tokens are frequently suppressed into teacher non-epistemic targets and are highly concentrated in high pointwise student-teacher divergence regions. To resolve this optimization pathology, we propose Adaptive Dual-Perspective OPSD (AD-OPSD), a robust control framework that dynamically moderates the self-distillation objective. AD-OPSD selectively anchors high-suppression-risk sandboxed tokens to a reference prior derived from the frozen base model via an asymmetrical pointwise divergence gate, preserving native thinking capacity while retaining OPSD's error-correcting power. Extensive experiments across competitive mathematical benchmarks show that AD-OPSD improves over standard OPSD by up to +4.1% absolute average accuracy across diverse model scales and datasets. Further analysis demonstrates that AD-OPSD mitigates thinking collapse and generalizes robustly to different post-training paradigms.