SAGE: Multi-Agent Self-Evolution for LLM Reasoning

TL;DR

SAGE introduces a multi-agent self-evolution framework for LLM reasoning that co-evolves agents to generate, plan, solve, and verify tasks, significantly improving performance on reasoning benchmarks without extensive human-labeled data.

Contribution

The paper proposes a novel multi-agent self-evolution framework for LLM reasoning that reduces reliance on labeled data and enhances stability and performance in multi-step reasoning tasks.

Findings

Achieved 8.9% improvement on LiveCodeBench with Qwen-2.5-7B.

Achieved 10.7% improvement on OlympiadBench.

Demonstrated stable self-training with quality control mechanisms.

Abstract

Reinforcement learning with verifiable rewards improves reasoning in large language models (LLMs), but many methods still rely on large human-labeled datasets. While self-play reduces this dependency, it often lacks explicit planning and strong quality control, limiting stability in long-horizon multi-step reasoning. We present SAGE (Self-evolving Agents for Generalized reasoning Evolution), a closed-loop framework where four agents: Challenger, Planner, Solver, and Critic, co-evolve from a shared LLM backbone using only a small seed set. The Challenger continuously generates increasingly difficult tasks; the Planner converts each task into a structured multi-step plan; and the Solver follows the plan to produce an answer, whose correctness is determined by external verifiers. The Critic scores and filters both generated questions and plans to prevent curriculum drift and maintain training signal quality, enabling stable self-training. Across mathematics and code-generation benchmarks, SAGE delivers consistent gains across model scales, improving the Qwen-2.5-7B model by 8.9% on LiveCodeBench and 10.7% on OlympiadBench.