INFUSER: Influence-Guided Self-Evolution Improves Reasoning

INFUSER is an innovative co-training framework designed for self-evolving language models, featuring a Generator that creates questions and answers from unstructured documents, and a Solver that learns from these materials. It employs a novel optimizer-aware influence score to enhance the Generator's performance, utilizing a dual-normalized variant of GRPO (DuGRPO) for training. On the Qwen3-8B-Base model, INFUSER achieves over 20% relative improvement on Olympiad and SuperGPQA benchmarks, demonstrating its effectiveness in adaptive curriculum learning and offering a flexible approach for practitioners aiming to enhance reasoning capabilities in LLMs.