Elamur Research⌕
☰
home / papers / 2026 / Thinking with Reasoning Skills: Fewer Tokens, More Accuracy

Thinking with Reasoning Skills: Fewer Tokens, More Accuracy

Paper · 2026 · Guangxiang Zhao, Qilong Shi, Xusen Xiao et al.

Summary

Reasoning LLMs such as OpenAI o1 and DeepSeek-R1 often produce thousands of intermediate tokens per query. Much of this is repeated trial-and-error, which drives up API cost and latency. Earlier efficiency methods like Chain-of-Draft, TALE and NoWait shorten the traces, but every query still starts from nothing, and accuracy often drops on hard problems. The authors propose Thinking with Reasoning Skills (TRS), a training-free, retrieval-based framework that works with black-box APIs. Offline, a stronger summarizer model turns full solution trajectories, whether successful or failed, into short structured "skill cards" indexed by keywords. Online, the system retrieves the most relevant cards for a new question and adds them to the prompt, so the model can follow a known strategy instead of rediscovering it. On math (DeepMath) and coding benchmarks, TRS uses fewer reasoning tokens while matching or beating standard CoT and compressed-prompt baselines. The gains are largest on harder problems and for weaker models. This matters for production deployments billed per token.

Key points

  • Skill cards:
    • Each card follows a fixed Trigger / Do / Avoid / Check / Risk format and comes with 10–20 retrieval keywords.
    • Failed trajectories become anti-pattern → correction cards.
    • Cards are stored as key–value pairs, with the original question plus keywords as the key.
  • Retrieval: BM25 is the default for math and hybrid sparse+dense retrieval for coding. A short gating instruction tells the model to ignore inapplicable or conflicting skills.
  • External transfer test:
    • Skills came from an AoPS-derived library of 7,616 contest problems and were evaluated on 120 questions (AIME 2024 I/II, AIME 2025, AIME 2026, HMMT November 2025).
    • Accuracy improved in 13 of 25 model–benchmark pairs and cost dropped in 20 of 25.
    • Doubao-1.8 gained +1.88 points with 2.8% lower cost; GPT-OSS-120B gained +0.50 with 6.5% lower cost.
  • Cross-model transfer: Skills distilled from one model help other models, though skills from the same source model usually give the best accuracy.
  • Ablations:
    • Swapping the summarizer to GPT-5-mini (500-question DeepMath subset) gave 89.4% on Doubao and 89.6% on OSS, still below full TRS on OSS.
    • The best injection prompt depends on the model: "Short" works best for Doubao and "Draft" for OSS.

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

🎓Tsinghua UniversityBeijing · China333 works · 87,019 citations🎓Peking UniversityBeijing · China244 works · 39,359 citations🎓University of Hong KongPok Fu Lam · China89 works · 11,774 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Guangxiang Zhao, Qilong Shi, Xusen Xiao, Xiangzheng Zhang, Tong Yang and Lin Sun. 2026. Thinking with Reasoning Skills: Fewer Tokens, More Accuracy. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.154

@inproceedings{zhao2026thinking,
  title = {{Thinking with Reasoning Skills: Fewer Tokens, More Accuracy}},
  author = {Guangxiang Zhao, Qilong Shi, Xusen Xiao, Xiangzheng Zhang, Tong Yang and Lin Sun},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.154},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.154}
}

References

23 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Anthropic. 2026. Introducing claude opus 4.6. https://www.anthropic.com/news/claude-opus-4-6. Accessed: 2026-02-15. ↗
  2. Xiang Chen, Lei Li, Ningyu Zhang, Xiaozhuan Liang, Shumin Deng, Chuanqi Tan, Fei Huang, Luo Si, and Huajun Chen. 2022. Retrieval-augmented prompt learning. arXiv preprint arXiv:2205.14704. — [arXiv:2205.14704](https://arxiv.org/abs/2205.14704) find
  3. DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  4. Google. 2025. A new era of intelligence with gemini 3. https://blog.google/ products-and-platforms/products/gemini/ gemini-3/. Accessed: 2026-02-15. ↗
  5. Zishan Guo, Hongyu Jin, Hongshen Tan, Xin Jiang, and Yuan-Fang Wang. 2023. Prompt-guided retrieval augmentation for non-knowledge-intensive tasks. In Findings of the Association for Computational Linguistics: ACL 2023. find
  6. Tingxu Han, Chunrong Fang, Shiyu Zhao, Shiqing Ma, Zhenyu Chen, and Zhenting Wang. 2024. Token-budget-aware LLM reasoning. Preprint, arXiv:2412.18547.
  7. Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Coconut: Reasoning in continuous latent space for efficient inference. arXiv preprint arXiv:2412.06769. find
  8. Junhao Hu, Wenrui Huang, Weidong Wang, Zhenwen Li, Tiancheng Hu, Zhixia Liu, Xusheng Chen, Tao Xie, and Yizhou Shan. 2025. Raas: Reasoning-aware attention sparsity for efficient llm reasoning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 2577–2590, Vienna, Austria. A… find
  9. OpenAI. 2024. Openai o1 system card. https://openai.com/index/openai-o1-system-card/. Accessed: 2026-02-15. ↗
  10. OpenAI. 2025. Gpt-5 system card. https://openai.com/index/gpt-5-system-card/. Accessed: 2026-02-15. ↗
  11. OpenAI. 2026. Openai api pricing. https://openai.com/api/pricing/. Accessed: 2026-02-15. ↗
  12. Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. arXiv preprint arXiv:2303.11366.
  13. Uptime Institute. 2025. Ai reasoning will take a toll on infrastructure footprint. https://uptimeinstitute.com/blog/ai-reasoning-will-take-a-toll-on-infrastructure-footprint. Accessed: 2026-02-15. ↗
  14. Chenlong Wang, Yuanning Feng, Dongping Chen, Zhaoyang Chu, Ranjay Krishna, and Tianyi Zhou. 2025. Wait, we don’t need to “wait”! removing thinking tokens improves reasoning efficiency. arXiv preprint arXiv:2506.08343. find
  15. Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291. find
  16. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V. Le, Ed H. Chi, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.
  17. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903.
  18. Mingyuan Wu, Jize Jiang, Haozhen Zheng, Meitang Li, Zhaoheng Li, Beitong Tian, Bo Chen, Yongjoo Park, Minjia Zhang, ChengXiang Zhai, and Klara Nahrstedt. 2025. Cache-of-thought: Master-apprentice framework for cost-effective vision language model reasoning. In Proceedings of the 2025 Conference on E… find
  19. Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of draft: Thinking faster by writing less. Preprint, arXiv:2502.18600. find
  20. Ling Yang, Zhaochen Yu, Tianjun Zhang, Shiyi Cao, Minkai Xu, Wentao Zhang, Joseph E. Gonzalez, and Bin Cui. 2024. Buffer of thoughts: Thought-augmented reasoning with large language models. arXiv preprint arXiv:2406.04271. find
  21. Ling Yang, Zhaochen Yu, Tianjun Zhang, Minkai Xu, Joseph E. Gonzalez, Bin Cui, and Shuicheng Yan. 2025. Supercorrect: Advancing small llm reasoning with thought template distillation and self-correction. In International Conference on Learning Representations (ICLR). find
  22. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. ReAct: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629.
  23. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhao-2026-thinking-reasoning-skills-fewer-tokens/