BARREL: Boundary-Aware Reasoning for Factual and Reliable LRMs
🤖 AI Summary
Large reasoning models (LRMs) exhibit strong performance in mathematical and logical reasoning but often suffer from overconfidence due to ill-defined cognitive boundaries, leading to erroneous outputs on uncertain queries—particularly manifesting as “last-step random guessing” and “recursive indecision spirals.” This work formally defines and models these two pathological overthinking patterns for the first time. We propose Boundary-Aware Reasoning via Regularized Learning (BARREL), a concise reasoning framework incorporating a learnable cognitive boundary mechanism that enables models to actively identify and abstain from answering unknown questions. BARREL integrates knowledge distillation, controllable reasoning-path constraints, uncertainty-aware loss, and boundary regularization. Evaluated on DeepSeek-R1-Distill-Llama-8B, BARREL improves factual reliability from 39.33% to 61.48%, achieving accuracy competitive with models fine-tuned on R1-generated data. Our contributions include: (i) the first formal characterization of overthinking pathologies in LRMs; (ii) a principled, boundary-aware reasoning framework; and (iii) empirical validation demonstrating substantial gains in reliability without sacrificing accuracy.
Application Category
📝 Abstract
Recent advances in Large Reasoning Models (LRMs) have shown impressive capabilities in mathematical and logical reasoning. However, current LRMs rarely admit ignorance or respond with"I don't know". Instead, they often produce incorrect answers while showing undue confidence, raising concerns about their factual reliability. In this work, we identify two pathological reasoning patterns characterized by overthinking that contribute to the overconfident and incorrect answers: last-minute guessing and second-thought spiraling. To address these issues, we propose BARREL-a novel framework that promotes concise and boundary-aware factual reasoning. Our experiments show that BARREL-training increases the reliability of DeepSeek-R1-Distill-Llama-8B from 39.33% to 61.48%, while still achieving accuracy comparable to models finetuned on reasoning data generated by R1. These results demonstrate that our pilot study is inspiring to build more reliable and factual System 2 LRMs.
Problem
Research questions and friction points this paper is trying to address.
LRMs often produce incorrect answers with undue confidence
Identify pathological reasoning patterns causing overconfident responses
Propose BARREL to enhance factual reliability in LRMs
Innovation
Methods, ideas, or system contributions that make the work stand out.
Boundary-aware reasoning framework BARREL
Addresses overthinking in LRMs
Improves reliability without accuracy loss
🔎 Similar Papers
No similar papers found.
