# Self-Improvement Can Self-Regress: The Rise-and-Collapse Failure Mode of LLM Self-Training

arXiv: https://arxiv.org/abs/2606.21090 · HTML: https://mlsanity.com/paper/2606.21090/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning.
- Builds on 29 of its 39 references that are in the map.
- First posted on arXiv 17 Jun 2026.

## Placed in

- LLM reasoning / [Reinforcement learning for reasoning](https://mlsanity.com/lr/reinforcement-learning-for-reasoning/)

## Builds on (most cited first, 29 of 29)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Evaluating Large Language Models Trained on Code](https://mlsanity.com/paper/2107.03374/) (2021)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks](https://mlsanity.com/paper/1703.03400/) (2017)
- [Overcoming catastrophic forgetting in neural networks](https://mlsanity.com/paper/1612.00796/) (2016)
- [Program Synthesis with Large Language Models](https://mlsanity.com/paper/2108.07732/) (2021)
- [Gradient Episodic Memory for Continual Learning](https://mlsanity.com/paper/1706.08840/) (2017)
- [The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery](https://mlsanity.com/paper/2408.06292/) (2024)
- [STaR: Bootstrapping Reasoning With Reasoning](https://mlsanity.com/paper/2203.14465/) (2022)
- [Measuring Coding Challenge Competence With APPS](https://mlsanity.com/paper/2105.09938/) (2021)
- [Self-Rewarding Language Models](https://mlsanity.com/paper/2401.10020/) (2024)
- [Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models](https://mlsanity.com/paper/2401.01335/) (2024)
- [Training Language Models to Self-Correct via Reinforcement Learning](https://mlsanity.com/paper/2409.12917/) (2024)
- [DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining](https://mlsanity.com/paper/2305.10429/) (2023)
- [Meta-SGD: Learning to Learn Quickly for Few-Shot Learning](https://mlsanity.com/paper/1707.09835/) (2017)
- [Population Based Training of Neural Networks](https://mlsanity.com/paper/1711.09846/) (2017)
- [Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models](https://mlsanity.com/paper/2312.06585/) (2023)
- [Iterative Reasoning Preference Optimization](https://mlsanity.com/paper/2404.19733/) (2024)
- [Differentiable Expected Hypervolume Improvement for Parallel Multi-Objective Bayesian Optimization](https://mlsanity.com/paper/2006.05078/) (2020)
- [Pass@k Training for Adaptively Balancing Exploration and Exploitation of Large Reasoning Models](https://mlsanity.com/paper/2508.10751/) (2025)
- [ReST meets ReAct: Self-Improvement for Multi-Step Reasoning LLM Agent](https://mlsanity.com/paper/2312.10003/) (2023)
- [Can Large Reasoning Models Self-Train?](https://mlsanity.com/paper/2505.21444/) (2025)
- [Prosperity before Collapse: How Far Can Off-Policy RL Reach with Stale Data on LLMs?](https://mlsanity.com/paper/2510.01161/) (2025)
- [Multi-Objective Hyperparameter Optimization in Machine Learning -- An Overview](https://mlsanity.com/paper/2206.07438/) (2022)
- [Exploration vs Exploitation: Rethinking RLVR through Clipping, Entropy, and Spurious Reward](https://mlsanity.com/paper/2512.16912/) (2025)
- [Progress or Regress? Self-Improvement Reversal in Post-training](https://mlsanity.com/paper/2407.05013/) (2024)
- [Flexible Entropy Control in RLVR with a Gradient-Preserving Perspective](https://mlsanity.com/paper/2602.09782/) (2026)

## Built on by (most cited first, 4 of 4)

- [Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops](https://mlsanity.com/paper/2607.07663/) (2026)
- [Training AI Scientists to Replicate Research](https://mlsanity.com/paper/2608.13331/) (2026)
- [Question Begets Question: Self-Evolving Curriculum for Reinforcement Fine-Tuning on Competition Mathematics](https://mlsanity.com/paper/2608.01522/) (2026)
- [When Updating Stops Being Learning: Rethinking LLM Self-Evolution via learnable information gain](https://mlsanity.com/paper/2609.36535/) (2026)

## Abstract (arXiv)

Self-improvement can self-regress. In REINFORCE post-training for code, a model can quickly improve on its optimized metric and then collapse within the same training campaign. We study this in a controlled multi-seed testbed using Qwen-2.5-3B and Qwen-2.5-7B, trained on competitive-programming tasks with binary CodeGrader reward across 10 sequential 20-step campaigns. Across campaigns, pass@1 shows a robust rise-then-collapse pattern: it peaks within tens of gradient steps and then falls back, sometimes to near zero. This is not cross-task catastrophic forgetting, but within-task policy over-optimization on a fixed distribution; KL- and EWC-style constraints do not prevent it. We ask where the control loop should sit. We compare three levels: CARE, a between-campaign memory mechanism with a capability posterior, transfer gate, and regression-aware belief revision; ES, a within-campaign early-stop rule that rolls forward the peak checkpoint and sets the next budget to peak_step+3; and GRPO, which changes the RL update using group-relative reward normalization. The answer is regime-dependent. On Qwen-2.5-3B, where naive REINFORCE is fragile, CARE v2 nearly doubles end-of-chain pass@1 from 4.9% to 9.5%, with paired bootstrap 95% CI [+0.4,+8.9] and gains in 4/5 seeds. On Qwen-2.5-7B, CARE reaches parity with naive REINFORCE, 13.8% vs. 11.8%, while ES reaches 22.2% [14.1,28.0]. Out-of-the-box GRPO reaches 20.7% [15.7,25.1], nearly matching REINFORCE+ES. GRPO raises the floor but does not remove the cliff. Its 7B gain mainly comes from better between-campaign carryover, while the within-campaign peak-to-end gap remains about 17 points under both REINFORCE and GRPO. GRPO+ES gives mixed evidence: 2/3 seeds improve, but one final cliff lowers the mean to 17.0% [0.0,28.1]. A Gemma-3-4B pilot shows the same signature, suggesting the phenomenon is not limited to Qwen.