--- title: "Your Reasoning Model Knows What Counts: Self-Guided" description: "Large Reasoning Models often produce long chain-of-thought traces full of detours, which raise inference cost and can hurt… 2 references." authors: ["Zi-Ao Ma", "Xian-Ling Mao", "Tian Lan", "Chen Xu", "Zhijing Wu"] year: 2026 venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)" source: "https://aclanthology.org/2026.acl-long.25/" doi: "10.18653/v1/2026.acl-long.25" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/ma-2026-your-reasoning-model-knows-what/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/ma-2026-your-reasoning-model-knows-what/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/57/) / Your Reasoning Model Knows What Counts: Self-Guided Chain-of-Thought Pruning for Efficient Reasoning # Your Reasoning Model Knows What Counts: Self-Guided Chain-of-Thought Pruning for Efficient Reasoning **Authors:** Zi-Ao Ma, [Xian-Ling Mao](https://md.research.elamur.ai/free/en/authors/xian-ling-mao/), [Tian Lan](https://md.research.elamur.ai/free/en/authors/tian-lan/), [Chen Xu](https://md.research.elamur.ai/free/en/authors/chen-xu/) and [Zhijing Wu](https://md.research.elamur.ai/free/en/authors/zhijing-wu/) · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2026](https://md.research.elamur.ai/free/en/venues/acl/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.acl-long.25/) · [PDF](https://aclanthology.org/2026.acl-long.25.pdf) · [DOI 10.18653/v1/2026.acl-long.25](https://doi.org/10.18653/v1/2026.acl-long.25) · [OpenAlex](https://openalex.org/W7166825786) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary Large Reasoning Models often produce long chain-of-thought traces full of detours, which raise inference cost and can hurt robustness. Earlier compression methods rely on length penalties, truncation heuristics, external verifiers or teacher-curated data. Because of this, they may cut steps that a particular model actually needs. SGP-CoT (Self-Guided Pruning) instead lets each model judge its own reasoning. The method works in three steps: 1. It samples several correct responses per prompt. 2. It splits each trace into reasoning units, using line-initial trigger tokens learned from frequency statistics. 3. It scores every unit with the model's own likelihoods. One score captures how much the unit supports the final answer, and another captures how much it supports local coherence. Low-impact units are removed. The pruned and original traces then form preference pairs for DPO training, so no external supervision is required. On AIME, GPQA-Diamond, MATH-500 and GSM8K, the tuned models give shorter answers with equal or better accuracy. The approach suits accuracy-sensitive deployments where cheaper inference matters. ## Key points - Experiments use DeepSeek-R1-Distill-Qwen-1.5B and 7B. Average end-to-end latency on one H800 GPU dropped from 640 s to 328 s (about 49%) and from 834 s to 470 s (about 44%), respectively. - LC-R1 and L1 compress more aggressively but lose more accuracy. SGP-CoT trades some compression for more consistent accuracy preservation. - About 1.5K preference pairs are enough. Scaling to 10K–30K adds little compression, and larger sets reduce accuracy on hard benchmarks, most visibly AIME 2024. - Ablations show that replacing DPO with plain SFT lengthens outputs on three benchmarks and lowers accuracy. Of the two scores, the answer-impact threshold matters more than the coherence threshold. - Impact scoring for a 7B–8B model costs under 10 H800 GPU hours. Code is available at github.com/HammerScholar/SGP-CoT. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions - [Beijing Institute of Technology](https://md.research.elamur.ai/free/en/universities/beijing-institute-of-technology/) — University, Beijing, China · 43 works · 3,652 citations 📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 3,486 works 🧭 Topics (2): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,685 works · [🪶 Efficient models & compression](https://md.research.elamur.ai/free/en/topics/efficient-models/) · 297 works ## Cite Cite as: Zi-Ao Ma, Xian-Ling Mao, Tian Lan, Chen Xu and Zhijing Wu. 2026. Your Reasoning Model Knows What Counts: Self-Guided Chain-of-Thought Pruning for Efficient Reasoning. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{ma2026your, title = {{Your Reasoning Model Knows What Counts: Self-Guided Chain-of-Thought Pruning for Efficient Reasoning}}, author = {Zi-Ao Ma, Xian-Ling Mao, Tian Lan, Chen Xu and Zhijing Wu}, booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, year = {2026}, doi = {10.18653/v1/2026.acl-long.25}, url = {https://doi.org/10.18653/v1/2026.acl-long.25} } ``` - [← previous Your Reasoning Benchmark May Not Test Reasoning: Revealing Perception Bottleneck in Abstract Reasoning Benchmarks](https://md.research.elamur.ai/free/en/papers/wang-2026-your-reasoning-benchmark-may-not/) - [next → Yoysef Meyer Yavets (1832–1914), Yiddish Translator of Sacred Jewish Books](https://md.research.elamur.ai/free/en/papers/hever-chybowski-2026-yoysef-meyer-yavets-18321914-yiddish/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/57/) ## References 2 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Pranjal Aggarwal and Sean Welleck. 2025. [L1: Controlling how long a reasoning model thinks with reinforcement learning](https://md.research.elamur.ai/free/en/papers/aggarwal-2025-l1-controlling-long-reasoning-model/). arXiv preprint arXiv:2503.04697. 2. While SGP-CoT demonstrates strong performance across diverse reasoning tasks, we acknowledge several limitations that suggest directions for future work: [find](https://scholar.google.com/scholar?q=While+SGP%2DCoT+demonstrates+strong+performance+across+diverse+reasoning+tasks%2C+we+acknowledge+several+limitations+that+suggest+directions+for+future+work%3A) *Updated 1 November 2026*