Rewarding Efficient Reasoning Improves Abstention on Underspecified Tasks in Reasoning Models

Abstract:While modern large reasoning models (LRMs) excel at providing correct answers in many tasks, we provide additional evidence for the observation that they often struggle with a critical capability: knowing when to abstain from answering. We analyze this gap by comparing LRM behavior to results from a human study, revealing that human reasoning effort on unanswerable tasks is upper-bounded by answerable tasks, whereas LRMs waste computational resources by generating longer Chains of Thought (CoTs) on unanswerable than on answerable prompts. To overcome this inefficiency, we take inspiration from a resource-rational perspective on human cognition and introduce a novel GRPO reward that encourages efficient reasoning about whether the task contains all the information needed to solve it. Fine-tuning several 4B LRMs with this reward leads to human-like abstention performance gains (+12.8% on average) while retaining answering capabilities and boosting the models' efficiency (44% shorter CoTs on average).

Computation and Language (cs.CL); Machine Learning (cs.LG)

arXiv:2609.20846 [cs.CL]

(or arXiv:2609.20846v1 [cs.CL] for this version)

https://doi.org/10.48550/arXiv.2609.20846 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.