Elamur Research⌕
☰
home / papers / 2026 / CRISP: Compressed Reasoning via Iterative Self-Policy Distillation

CRISP: Compressed Reasoning via Iterative Self-Policy Distillation

Paper · 2026 · Hejian Sang, Yuanda Xu, Zhengze Zhou et al.

Summary

Reasoning models often produce very long chains of thought, even for simple problems. This raises inference cost and can introduce mistakes. CRISP shortens reasoning without labelled answers, length budgets or difficulty estimators. The model acts as its own teacher: the same network, given an extra instruction to be brief, produces target logits. The uninstructed student then generates rollouts on math prompts from DAPO-17K, and training minimizes per-token reverse KL toward the teacher on those rollouts. The teacher's weights are refreshed from the student at fixed intervals, so compression can deepen over time.

Because reverse KL is mode-seeking, the compression adapts to difficulty. Easy problems lose most of their tokens, while hard ones keep the steps they need. On Qwen3-14B, Qwen3-8B and DeepSeek-R1-Distill-Llama-8B, outputs became much shorter with accuracy kept or improved. Out-of-domain skills were also preserved. The approach matters because it needs only prompts and a brevity instruction, so it could apply where reliable verifiers do not exist.

Key points

  • On Qwen3-14B (30K budget, mean@8), reasoning length fell by 56% on MATH-500, 38% on AIME 2024 and 32% on AIME 2025.
    • MATH-500 accuracy rose from 93.0 to 96.3.
    • AIME 2024 stayed within roughly one point of the base model, while AIME 2025 dropped modestly.
  • On DeepSeek-R1-Distill-Llama-8B, accuracy improved on all five benchmarks while responses got shorter. Training used only math, yet GPQA-Diamond and MMLU scores held steady.
  • Reverse KL is essential. With forward KL, Qwen3-8B collapsed within about 100 steps: accuracy fell to near zero and length grew to the token cap.
  • Short training rollouts are enough. A 1k-token cap still gave 29.8% MATH-500 compression with no accuracy loss. The 8k default worked best, at 31.6% compression.
  • Code is available at github.com/HJSang/OPSD_Reasoning_Compression. The authors note the method depends on how well the base model follows instructions.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 19,503 worksArtificial Intelligence · 18,805 worksArts and Humanities · 17,959 worksLanguage and Linguistics · 17,789 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,948 works

Cite

Cite as Hejian Sang, Yuanda Xu, Zhengze Zhou, Ran He, Zhipeng Wang and Jiachen Sun. 2026. CRISP: Compressed Reasoning via Iterative Self-Policy Distillation. arXiv:2603.05433.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2603.05433 · arXiv 2603.05433

@misc{sang2026crisp,
  title = {{CRISP: Compressed Reasoning via Iterative Self-Policy Distillation}},
  author = {Hejian Sang, Yuanda Xu, Zhengze Zhou, Ran He, Zhipeng Wang and Jiachen Sun},
  year = {2026},
  doi = {10.48550/arxiv.2603.05433},
  eprint = {2603.05433},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2603.05433}
}

References

6 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. P. Aggarwal and S. Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697, 2025.
  2. Y. Xu, H. Sang, Z. Zhou, R. He, and Z. Wang. Overconfident errors need stronger correction: Asymmetric confidence penalties for reinforcement learning. arXiv preprint arXiv:2602.21420, 2026. — [arXiv:2602.21420](https://arxiv.org/abs/2602.21420) find
  3. Yinger Zhang, Shutong Jiang, Renhao Li, Jianhong Tu, Yang Su, Lianghao Deng, Xudong Guo, Chenxu Lv, and Junyang Lin. Deepplanning: Benchmarking long-horizon agentic planning with verifiable constraints. arXiv preprint arXiv:2601.18137, 2026.
  4. B. Hou, Y. Zhang, J. Ji, Y. Liu, K. Qian, J. Andreas, and S. Chang. Thinkprune: Pruning long chain-of-thought of llms via reinforcement learning. arXiv preprint arXiv:2504.01296, 2025.
  5. W. Chen, J. Yuan, T. Jin, N. Ding, H. Chen, Z. Liu, and M. Sun. The overthinker’s diet: Cutting token calories with difficulty-aware training. arXiv preprint arXiv:2505.19217, 2025. W.-R. Chen, V. Kothapalli, A. Fatahibaarzi, H. Sang, S. Tang, Q. Song, Z. Wang, and M. Abdul-Mageed. Distilling the es… find
  6. D. Rein, B. L. Hou, A. C. Stickland, J. Petty, R. Y. Pang, J. Dirani, J. Michael, and S. R. Bowman. Gpqa: A graduate-level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022, 2023. I. Shenfeld, M. Damani, J. Hübotter, and P. Agrawal. Self-distillation enables continual learning. arXiv prepr…

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sang-2026-crisp-compressed-reasoning-iterative-self/