Teaching Reasoning Models When to Stop

The next challenge for large reasoning models is not simply producing correct answers. It is deciding whether an answer is justified in the first place. A paper on arXiv examines how models behave when a prompt lacks information needed to reach a determinate solution, and finds a counterintuitive pattern: models may reason longer on unanswerable tasks than on answerable ones.

A mismatch with human reasoning

The authors compare model behavior with findings from a human study. When people encounter a task that cannot be solved from the available information, their reasoning effort is generally bounded by the effort spent on answerable tasks. Once the missing information becomes apparent, continuing to search for a solution has little value.

Large reasoning models show the opposite tendency. Faced with an underspecified prompt, they can produce an extended chain of thought and continue trying to construct an answer. This behavior treats additional reasoning as the default remedy, even when the problem is missing essential premises. The result is not only unreliable answering, but also unnecessary inference cost.

Rewarding an early solvability check

To address this issue, the researchers draw on a resource-rational view of human cognition. They introduce a GRPO reward designed to encourage the model to reason efficiently about whether a task contains enough information to be solved. The objective is not to make the model abstain more often in general, but to connect abstention with an earlier assessment of solvability.

The method was used to fine-tune several 4B reasoning models. The evaluation focused on three questions:

• Can the model recognize when a prompt is underspecified?
• Does better abstention damage performance on answerable tasks?
• Can the model reach its decision with less generated reasoning?

According to the reported results, abstention performance improved by 12.8% on average. The models retained their ability to answer solvable tasks, while their chains of thought became 44% shorter on average. These results suggest that better abstention does not necessarily require more elaborate reasoning. In many cases, the important capability is identifying the information gap before launching into a long search.

Why this matters for deployed systems

The work links reliability to inference efficiency. In production systems, avoiding prolonged reasoning on impossible or underspecified requests could reduce latency and computation. For users, a timely explanation that the prompt lacks necessary information is preferable to a confident-looking answer assembled from unsupported assumptions.

The available material does not establish how the method would perform across broader model sizes, task families, or real-world deployments. The reported findings come from several 4B models and should therefore be read as evidence for a training direction rather than a universal solution.

The central open question is calibration. A useful model must distinguish a genuinely unsolvable prompt from one that merely requires deeper reasoning. Rewarding efficiency may help establish that distinction, but excessive pressure to stop could also increase unnecessary abstentions. The broader lesson is that reasoning quality includes knowing when further reasoning is no longer productive.

arXiv