Self-Improving Large Language Models via Progressive Experience Evolution

2026-08-03

AI Summary

Existing self-improvement approaches for large language models suffer from a critical disconnect: test-time methods struggle to internalize experience into model parameters, while training-time methods lack explicit mechanisms to accumulate transferable knowledge. This work proposes the SPEE framework, which enables effective self-improvement through a two-stage process. First, it explicitly extracts, validates, and evolves transferable experiences from multi-turn interactions, constructing a continuously updated global experience pool that integrates both successful and failed trajectories to mitigate post-hoc rationalization bias. Second, it internalizes this knowledge via online privileged-guided self-distillation (OPSD) and further explores novel strategies through reward-driven reinforcement learning. Evaluated across five mathematical reasoning benchmarks, SPEE consistently outperforms current test-time and training-time self-improvement baselines across three model scales.

Abstract

Large language models (LLMs) capable of self-improvement require not only effective policy optimization, but also a principled mechanism for transforming transient interaction experience into persistent model capabilities. Existing self-improvement paradigms remain fragmented: test-time methods can explicitly extract experience but cannot internalize it into model parameters, whereas training-time optimization methods can update model parameters but lack an explicit mechanism for accumulating transferable experience. Bridging these two paradigms requires a critical intermediate stage that remains underexplored, namely \emph{experience distillation}. To address this gap, we propose \textbf{SPEE} (\textbf{S}elf-\textbf{P}rogressive \textbf{E}xperience \textbf{E}volution), a unified post-training framework that sequentially performs explicit experience evolution followed by implicit policy optimization. During explicit experience evolution, SPEE reflects on trajectories collected from multiple interactions to extract, verify, and progressively evolve transferable experience, which is subsequently internalized into the policy through privilege-guided On-Policy Self-Distillation (OPSD). During implicit policy optimization, reward-driven reinforcement learning leverages these internalized priors to explore novel solution strategies. In the experience evolution stage, a continuously evolving global experience pool consolidates knowledge from both successful and failed trajectories, filters out low-utility experience, and mitigates post-hoc rationalization induced by individual trajectories. Experiments on five mathematical reasoning benchmarks demonstrate that SPEE consistently outperforms both test-time and training-time self-evolution baselines across three model scales. The source code is available at https://github.com/rrrsj/SPEE.

Problem

• self-improvement
• experience distillation
• large language models
• transferable experience
• policy optimization

Innovation

• experience distillation
• self-improvement
• on-policy self-distillation
• progressive experience evolution
• large language models