CHIMERA-4B-RL
This model was introduced in the paper CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning.
Authors: Xinyu Zhu, Yihao Feng, Yanchao Sun, Xianzhi Du, Pingzhi Li, Olli Saarikivi, Yun Zhu, Yu Meng.
Description
CHIMERA-4B-RL is CHIMERA-4B-SFT further trained with reinforcement learning on the CHIMERA dataset.
CHIMERA is a compact synthetic reasoning dataset comprising 9K samples designed for generalizable cross-domain reasoning. It provides rich, long Chain-of-Thought (CoT) trajectories across 8 major scientific disciplines. Despite its modest size, post-training a 4B model on this data allows it to approach or match the reasoning performance of significantly larger models like DeepSeek-R1 and Qwen3-235B.
Results
| Model | GPQA-D | AIME 24 | AIME 25 | AIME 26 | HMMT Feb 25 | HMMT Nov 25 | HLE |
|---|---|---|---|---|---|---|---|
| Qwen3-4B-Thinking-2507 | 65.8 | 81.6 | 81.0 | 80.8 | 59.2 | 57.3 | 7.3 |
| CHIMERA-4B-SFT | 68.8 | 86.5 | 79.8 | 80.3 | 63.1 | 66.3 | 9.0 |
| CHIMERA-4B-RL | 70.1 | 86.9 | 80.7 | 82.7 | 65.7 | 67.0 | 9.0 |
Citation
@article{zhu2026chimera,
title={CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning},
author={Zhu, Xinyu and Feng, Yihao and Sun, Yanchao and Du, Xianzhi and Li, Pingzhi and Saarikivi, Olli and Zhu, Yun and Meng, Yu},
journal={arXiv preprint arXiv:2603.00889},
year={2026}
}