Title: Can Large Reasoning Models Be Interrupted?

Abstract: Practical deployments of Large Reasoning Models (LRMs) frequently involve reasoning tasks where prompts or environments are not static. This study challenges the assumption of a fixed world by assessing LRM resilience in two dynamic contexts: interruptions, which measure response accuracy when output generation is constrained by a budget, and dynamic context, which evaluates the model’s ability to adapt to changes occurring during inference. In benchmarks focused on mathematics and programming that demand extended reasoning chains, static evaluations tend to significantly overstate model robustness. Even leading LRMs, which demonstrate high precision in static conditions, exhibit unpredictable failures when subjected to interruptions or shifting contexts, with performance declining by as much as 60% if updates are introduced late in the reasoning process. Our investigation identifies several distinct failure patterns: "reasoning leakage," where the model integrates reasoning steps directly into the final answer upon interruption; "panic," where models under time pressure discard reasoning altogether and provide incorrect responses; and "self-doubt," where performance suffers as the model attempts to assimilate new information. Project Page: http://dynamic-lm.github.io/

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC