![[2602.14451] Precedent-Informed Reasoning: Mitigating Overthinking in Large Reasoning Models via Test-Time Precedent Learning](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.14451] Precedent-Informed Reasoning: Mitigating Overthinking in Large Reasoning Models via Test-Time Precedent Learning
Summary
The paper introduces Precedent-Informed Reasoning (PIR) to enhance reasoning in Large Language Models (LLMs) by leveraging past cases, improving efficiency and accuracy.
Why It Matters
As LLMs become integral to various applications, optimizing their reasoning capabilities is crucial. This research addresses inefficiencies in LLMs by proposing a method that mimics human reasoning, potentially leading to more effective AI systems in real-world tasks.
Key Takeaways
- PIR reduces computational costs by minimizing redundant reasoning.
- Adaptive Precedent Selection (APS) identifies relevant precedents to guide reasoning.
- Test-time Experience Internalization (TEI) allows models to learn from precedents during operation.
- PIR improves accuracy-efficiency trade-offs across various reasoning tasks.
- The approach has implications for enhancing AI applications in complex problem-solving.
Computer Science > Artificial Intelligence arXiv:2602.14451 (cs) [Submitted on 16 Feb 2026] Title:Precedent-Informed Reasoning: Mitigating Overthinking in Large Reasoning Models via Test-Time Precedent Learning Authors:Qianyue Wang, Jinwu Hu, Huanxiang Lin, Bolin Chen, Zhiquan Wen, Yaofo Chen, Yu Rong, Mingkui Tan View a PDF of the paper titled Precedent-Informed Reasoning: Mitigating Overthinking in Large Reasoning Models via Test-Time Precedent Learning, by Qianyue Wang and 7 other authors View PDF HTML (experimental) Abstract:Reasoning in Large Language Models (LLMs) often suffers from inefficient long chain-of-thought traces with redundant self-exploration and validation, which inflate computational costs and even degrade performance. Inspired by human reasoning patterns where people solve new problems by leveraging past related cases to constrain search spaces and reduce trial-and-error, we propose Precedent Informed Reasoning (PIR) transforming LRMs'reasoning paradigm from exhaustive self-exploration to guided learning from precedents. PIR addresses two key challenges: what precedents to adopt and how to utilize them. First, Adaptive Precedent Selection (APS) constructs, for each question and LRM, a compact set of precedents that are both semantically related and informative for the model. It ranks examples by a joint score with semantic similarity and model perplexity, then adapts the amount of precedents to maximize perplexity reduction. Second, Test-time Experienc...