CRISP: Compressed Reasoning via Iterative Self-Policy Distillation
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2026
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866911586423144448 |
|---|---|
| author | Sang, Hejian Xu, Yuanda Zhou, Zhengze He, Ran Wang, Zhipeng Sun, Jiachen |
| author_facet | Sang, Hejian Xu, Yuanda Zhou, Zhengze He, Ran Wang, Zhipeng Sun, Jiachen |
| contents | Reasoning models think out loud, but much of what they say is noise. We introduce CRISP (Compressed Reasoning via Iterative Self-Policy Distillation), a method that teaches models to reason more concisely by distilling their own concise behavior back into themselves. The entire approach reduces to one idea: condition the same model on a ''be concise'' instruction to obtain teacher logits, and minimize per-token reverse KL on the student's own rollouts. No ground-truth answers, no token budgets, no difficulty estimators. Just self-distillation. Yet this simplicity belies surprising sophistication: CRISP automatically compresses easy problems aggressively while preserving the deliberation needed for hard ones. On Qwen3-8B and Qwen3-14B, we achieve 57--59% token reduction on MATH-500 while improving accuracy by 9--16 points absolute. On AIME 2024, the 14B model gains 10 points with 41% compression. Ablations show that qualitative conciseness instructions outperform explicit token targets, and periodic teacher refreshes yield a broad stable regime. The method generalizes across model families -- DeepSeek-R1-Distill-Llama-8B improves accuracy by up to 5 points with 17--32% compression -- and transfers beyond math to multi-step agentic planning (DeepPlanning), reducing token usage by 42--51% while preserving planning quality. Code is available at https://github.com/HJSang/OPSD_Reasoning_Compression. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_05433 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | CRISP: Compressed Reasoning via Iterative Self-Policy Distillation Sang, Hejian Xu, Yuanda Zhou, Zhengze He, Ran Wang, Zhipeng Sun, Jiachen Machine Learning Reasoning models think out loud, but much of what they say is noise. We introduce CRISP (Compressed Reasoning via Iterative Self-Policy Distillation), a method that teaches models to reason more concisely by distilling their own concise behavior back into themselves. The entire approach reduces to one idea: condition the same model on a ''be concise'' instruction to obtain teacher logits, and minimize per-token reverse KL on the student's own rollouts. No ground-truth answers, no token budgets, no difficulty estimators. Just self-distillation. Yet this simplicity belies surprising sophistication: CRISP automatically compresses easy problems aggressively while preserving the deliberation needed for hard ones. On Qwen3-8B and Qwen3-14B, we achieve 57--59% token reduction on MATH-500 while improving accuracy by 9--16 points absolute. On AIME 2024, the 14B model gains 10 points with 41% compression. Ablations show that qualitative conciseness instructions outperform explicit token targets, and periodic teacher refreshes yield a broad stable regime. The method generalizes across model families -- DeepSeek-R1-Distill-Llama-8B improves accuracy by up to 5 points with 17--32% compression -- and transfers beyond math to multi-step agentic planning (DeepPlanning), reducing token usage by 42--51% while preserving planning quality. Code is available at https://github.com/HJSang/OPSD_Reasoning_Compression. |
| title | CRISP: Compressed Reasoning via Iterative Self-Policy Distillation |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2603.05433 |