Caijun Xu

01 / Research agenda

Self-evolving LLMs.

I study how language models can continually improve through reasoning, interaction, feedback, and the data they generate.

Reasoning & RL

Improving multi-step reasoning capabilities through scalable reinforcement learning and exploration.

RLVRReasoning

Data synthesis

Creating scalable synthetic data, adaptive environments, and curricula for continual model improvement.

SynthesisEnvironments

Verifier & feedback

Designing reliable evaluation and feedback signals that close the loop between generation and learning.

VerifiersFeedback

02 / Updates

Recent news

All updates

May 27, 2026	📄 New preprint: DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes — learning to recover from weak-model failures without a stronger teacher.
May 04, 2026	📄🎉 SCALER: Synthetic Scalable Adaptive Learning Environment for Reasoning — accepted to ACL 2026 Findings.
Oct 20, 2024	📄🎉 Exploring High-Order User Preference with Knowledge Graph for Recommendation — accepted to CIKM 2024 (short paper).
Dec 03, 2023	🥉 Won ICPC Bronze Medal (Jinan).
Nov 05, 2023	🥈 Won CCPC Silver Medal (Harbin).
Oct 22, 2023	🥉 Won ICPC Bronze Medal (Xi’an).

03 / Selected work

Publications

Full list

Preprint

DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes

Caijun Xu, Changyi Xiao, Zhongyuan Peng, and Yixin Cao

arXiv preprint arXiv:2605.28421, May 2026

Abs arXiv Bib PDF

DenoiseRL is a reinforcement learning framework that replaces external supervision with recovery-oriented optimization over failures from weak models. It learns directly from incorrect reasoning traces by conditioning the policy on noisy prefixes and optimizing the on-policy continuation, improving reasoning performance, exploration efficiency, and self-corrective behavior without stronger teacher models or heavily curated difficult datasets.
@article{xu2026denoiserl, title = {{DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes}}, author = {Xu, Caijun and Xiao, Changyi and Peng, Zhongyuan and Cao, Yixin}, journal = {arXiv preprint arXiv:2605.28421}, year = {2026}, month = may, url = {https://arxiv.org/abs/2605.28421}, }

ACL’26 Findings

SCALER: Synthetic Scalable Adaptive Learning Environment for Reasoning

Caijun Xu, Changyi Xiao, Zhongyuan Peng, Xinrun Wang, and Yixin Cao

In Findings of the Association for Computational Linguistics: ACL 2026, May 2026

arXiv Bib

@inproceedings{xu2026scaler,
  title = {{SCALER: Synthetic Scalable Adaptive Learning Environment for Reasoning}},
  author = {Xu, Caijun and Xiao, Changyi and Peng, Zhongyuan and Wang, Xinrun and Cao, Yixin},
  booktitle = {Findings of the Association for Computational Linguistics: {ACL} 2026},
  year = {2026},
  url = {https://arxiv.org/abs/2601.04809},
}

CIKM’24

Exploring High-Order User Preference with Knowledge Graph for Recommendation

Caijun Xu, Fuwei Zhang, Zhao Zhang, Fuzhen Zhuang, and Rui Liu

In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, May 2024

Bib

@inproceedings{xu2024exploring,
  title = {{Exploring High-Order User Preference with Knowledge Graph for Recommendation}},
  author = {Xu, Caijun and Zhang, Fuwei and Zhang, Zhao and Zhuang, Fuzhen and Liu, Rui},
  booktitle = {Proceedings of the 33rd {ACM} International Conference on Information and Knowledge Management},
  pages = {4138--4142},
  year = {2024},
  url = {https://doi.org/10.1145/3627673.3679921},
}

04 / Contact

Let’s exchange ideas.

I’m always glad to discuss self-evolving LLMs, reasoning, reinforcement learning, and synthetic data.