INFUSER-Qwen3-8B-base

INFUSER-Qwen3-8B-base

This repository contains the INFUSER-trained checkpoint based on Qwen/Qwen3-8B-Base.

INFUSER is an iterative co-training framework introduced in INFUSER: Influence-Guided Self-Evolution Improves Reasoning. It co-evolves a Generator that drafts questions from unstructured documents and a Solver (this model) that improves by training on them. Unlike standard difficulty-based methods, INFUSER rewards the generator using an optimizer-aware influence score to create an adaptive curriculum.

• Paper: arXiv:2606.09052
• Project Page/Blog: https://y-agent.github.io/infuser
• Code: GitHub Repository

Summary

• Base model: Qwen/Qwen3-8B-Base
• Method: INFUSER
• Data repository: Siyuc/infuser-data

Evaluation

Released Checkpoint Scores

Comparison Summary

Category and overall means are computed over the same benchmark groups. R-Few (paper) and SPICE (paper) are self-reported values from their original papers, so missing categories are shown as -.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Siyuc/INFUSER-Qwen3-8B-base")
tokenizer = AutoTokenizer.from_pretrained("Siyuc/INFUSER-Qwen3-8B-base")

Citation

@misc{chen2026infuser,
  title        = {INFUSER: Influence-Guided Self-Evolution Improves Reasoning},
  author       = {Siyu Chen and Miao Lu and Beining Wu and Heejune Sheen and Fengzhuo Zhang and Shuangning Li and Zhiyuan Li and Jose Blanchet and Tianhao Wang and Zhuoran Yang},
  year         = {2026},
  eprint       = {2606.09052},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url          = {https://arxiv.org/abs/2606.09052}
}