# Superficial Self-Improved Reasoners Benefit from Model Merging

arXiv: https://arxiv.org/abs/2503.02103 · HTML: https://mlsanity.com/paper/2503.02103/

## Key facts

- Cited by 16 papers in the map (data through 14 Oct 2026).
- Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning.
- Builds on 27 of its 78 references that are in the map.
- First posted on arXiv 3 Mar 2025.

## Placed in

- LLM reasoning / [Test-time compute and efficient thinking](https://mlsanity.com/lr/test-time-compute-and-efficient-thinking/)
- Synthetic data and self-improvement for LLMs / [Self-rewarding alignment and self-training](https://mlsanity.com/synth/self-rewarding-alignment-and-self-training/)

## Builds on (most cited first, 27 of 27)

- [Llama 2: Open Foundation and Fine-Tuned Chat Models](https://mlsanity.com/paper/2307.09288/) (2023)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Let's Verify Step by Step](https://mlsanity.com/paper/2305.20050/) (2023)
- [OpenAI o1 System Card](https://mlsanity.com/paper/2412.16720/) (2024)
- [Self-Rewarding Language Models](https://mlsanity.com/paper/2401.10020/) (2024)
- [Large Language Models Can Self-Improve](https://mlsanity.com/paper/2210.11610/) (2022)
- [MathQA: Towards Interpretable Math Word Problem Solving with Operation-Based Formalisms](https://mlsanity.com/paper/1905.13319/) (2019)
- [Reinforced Self-Training (ReST) for Language Modeling](https://mlsanity.com/paper/2308.08998/) (2023)
- [Scaling Relationship on Learning Mathematical Reasoning with Large Language Models](https://mlsanity.com/paper/2308.01825/) (2023)
- [The Curse of Recursion: Training on Generated Data Makes Models Forget](https://mlsanity.com/paper/2305.17493/) (2023)
- [Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models](https://mlsanity.com/paper/2312.06585/) (2023)
- [CodeT5+: Open Code Large Language Models for Code Understanding and Generation](https://mlsanity.com/paper/2305.07922/) (2023)
- [Is synthetic data from generative models ready for image recognition?](https://mlsanity.com/paper/2210.07574/) (2022)
- [V-STaR: Training Verifiers for Self-Taught Reasoners](https://mlsanity.com/paper/2402.06457/) (2024)
- [Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge](https://mlsanity.com/paper/2407.19594/) (2024)
- [On the Stability of Iterative Retraining of Generative Models on their own Data](https://mlsanity.com/paper/2310.00429/) (2023)
- [The Curious Decline of Linguistic Diversity: Training Language Models on Synthetic Text](https://mlsanity.com/paper/2311.09807/) (2023)
- [How Bad is Training on Synthetic Data? A Statistical Analysis of Language Model Collapse](https://mlsanity.com/paper/2404.05090/) (2024)
- [MathGenie: Generating Synthetic Data with Question Back-translation for Enhancing Mathematical Reasoning of LLMs](https://mlsanity.com/paper/2402.16352/) (2024)
- [Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling](https://mlsanity.com/paper/2408.16737/) (2024)
- [Self-Consistency Preference Optimization](https://mlsanity.com/paper/2411.04109/) (2024)
- [CrossGET: Cross-Guided Ensemble of Tokens for Accelerating Vision-Language Transformers](https://mlsanity.com/paper/2305.17455/) (2023)
- [A Closer Look at How Fine-tuning Changes BERT](https://mlsanity.com/paper/2106.14282/) (2021)
- [GameArena: Evaluating LLM Reasoning through Live Computer Games](https://mlsanity.com/paper/2412.06394/) (2024)
- [Nepotistically Trained Generative-AI Models Collapse](https://mlsanity.com/paper/2311.12202/) (2023)

## Built on by (most cited first, 16 of 16)

- [Two Heads are Better Than One: Test-time Scaling of Multi-agent Collaborative Reasoning](https://mlsanity.com/paper/2504.09772/) (2025)
- [SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs](https://mlsanity.com/paper/2510.05069/) (2025)
- [How to Protect Yourself from 5G Radiation? Investigating LLM Responses to Implicit Misinformation](https://mlsanity.com/paper/2503.09598/) (2025)
- [Mitigating Forgetting Between Supervised and Reinforcement Learning Yields Stronger Reasoners](https://mlsanity.com/paper/2510.04454/) (2025)
- [Behavior Knowledge Merge in Reinforced Agentic Models](https://mlsanity.com/paper/2601.13572/) (2026)
- [LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models](https://mlsanity.com/paper/2507.14204/) (2025)
- [Theoretical Modeling of Large Language Model Self-Improvement Training Dynamics Through Solver-Verifier Gap](https://mlsanity.com/paper/2507.00075/) (2025)
- [Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner](https://mlsanity.com/paper/2506.01301/) (2025)
- [Assessing and Mitigating Medical Knowledge Drift and Conflicts in Large Language Models](https://mlsanity.com/paper/2505.07968/) (2025)
- [Stabilizing Self-Consuming Diffusion Models with Latent Space Filtering](https://mlsanity.com/paper/2511.12742/) (2025)
- [Growing Through Experience: Scaling Episodic Grounding in Language Models](https://mlsanity.com/paper/2506.01312/) (2025)
- [What Makes LLMs Effective Sequential Recommenders? A Study on Preference Intensity and Temporal Context](https://mlsanity.com/paper/2506.02261/) (2025)
- [CopT: Contrastive On-Policy Thinking with Continuous Spaces for General and Agentic Reasoning](https://mlsanity.com/paper/2605.20075/) (2026)
- [Post-Hoc Merging is Not Enough: Many-Shot Model Merging with Loss-Gap Balancing](https://mlsanity.com/paper/2606.16501/) (2026)
- [CHORUS: Complementary Experts for High-Coverage Testbench Stimulus Generation](https://mlsanity.com/paper/2608.10090/) (2026)
- [The evolution of sex for artificial intelligence: a population-genetic framework for multigenerational model populations](https://mlsanity.com/paper/2609.18560/) (2026)

## Abstract (arXiv)

As scaled language models (LMs) approach human-level reasoning capabilities, self-improvement emerges as a solution to synthesizing high-quality data corpus. While previous research has identified model collapse as a risk in self-improvement, where model outputs become increasingly deterministic, we discover a more fundamental challenge: the superficial self-improved reasoners phenomenon. In particular, our analysis reveals that even when LMs show improved in-domain (ID) reasoning accuracy, they actually compromise their generalized reasoning capabilities on out-of-domain (OOD) tasks due to memorization rather than genuine. Through a systematic investigation of LM architecture, we discover that during self-improvement, LM weight updates are concentrated in less reasoning-critical layers, leading to superficial learning. To address this, we propose Iterative Model Merging (IMM), a method that strategically combines weights from original and self-improved models to preserve generalization while incorporating genuine reasoning improvements. Our approach effectively mitigates both LM collapse and superficial learning, moving towards more stable self-improving systems.