The researchers released RLRT (RLVR with Reversed Teacher) on GitHub, along with OPSD and self-distilled RLVR baselines, as a testbed for single-turn math reasoning. The release is meant to offer a unified view of self-distillation in LLM post-training. The authors also say RLRT has been accepted to NeurIPS 2026 after an oral presentation at a COLM workshop.
We've open-sourced RLRT (RLVR with Reversed Teacher), together with representative OPSD and self-distilled RLVR baselines, as a testbed for studying self-distillation in single-turn (math) reasoning.
💻 Code: https://github.com/beanie00/self-distillation-analysis
🌐 Project page: https://beanie00.github.io/self-distillation-analysis/rlrt/
Feedback and PRs welcome! We hope it offers a unified view of self-distillation in LLM post-training and a solid foundation for future research.
RLRT has also been accepted to #NeurIPS2026, following its selection for an oral presentation at the COLM workshop! 🎉 TL;DR: RLRT (RLVR with Reversed Teacher) reverses self-distillation on correct rollouts by reinforcing self-driven tokens that diverge from the teacher, turning disagreement into valuable exploration. Glad to share this with my co-first author @jiwonjeon_jen! 😍 (Code will be released soon)
Source: Jeonghye Kim✈️ COLM 2026 · x.comPublished · added here
