d-OPSD: Self-Distillation from Future Answers
jiqizhixin · x · 2026-07-17
This paper introduces d-OPSD, an on-policy self-distillation framework for diffusion LLMs. The core idea: instead of only learning from past tokens, the model learns from its own generated "future answers," an approach the authors term suffix conditioning.
Method Highlights
- Supervision signals shift from the token level to the step level
- The training process aligns with the iterative mechanism of diffusion denoising
- The student model learns from "its own generated subsequent content"
Experimental Conclusions
- Across 4 reasoning benchmarks, d-OPSD outperforms both RLVR and SFT
- Training is highly efficient, requiring only about 10% of the optimization steps
The paper is titled Learning from the Self-future: On-policy Self-distillation for dLLMs.
More from Research
- Structural ensembles beat single predictions in TCR:pMHC generalization study — quaidmorris · 2026-07-22
- RSS launches under OMSF to push structural biology data modeling at scale — MoAlQuraishi · 2026-07-22
- enFoldX tops 8 neoantigen scans and an unseen-peptide benchmark — quaidmorris · 2026-07-22
- enFoldX reaches AUC 0.82 on human VDJdb and transfers to mouse at 0.76 — quaidmorris · 2026-07-22
- enFoldX gains accuracy as AF3 ensemble disagreement rises for non-binders — quaidmorris · 2026-07-22
- A 3D ray plot shows how hard this Jacobian counterexample is to read — moultano · 2026-07-22