Research map / Large language models / Reasoning and chain of thought

Are Large Reasoning Models Interruptible?

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Evaluates large reasoning models under mid-inference interruptions (hard/soft time constraints and update-driven context changes) on math and coding benchmarks, identifying reasoning leakage, panic, and self-doubt as failure modes.

Problem

Static evaluations of large reasoning models assume a frozen world, but real applications involve interruptions that force early answers, speed-up requests, or updated task specifications mid-reasoning; the robustness of LRMs under these dynamic conditions is unknown.

Approach

Formalizes interruptible reasoning as cutting a reasoning trace at relative positions and injecting hard stops, speedup directives, or update content, then measures interruption-conditioned accuracy and output length; it names and quantifies three failure modes (reasoning leakage, panic, self-doubt) and tests a training-free prompt-guidance postfix.

Key results

• Accuracy drops by up to 60% when updates are introduced late in the reasoning process.
• Under hard interrupts, early interruptions on AIME and LiveCodeBench can produce final answers up to ten times longer than full-thinking answers (reasoning leakage).
• Prompt guidance fully resolves self-doubt on GSM8K and MATH500 and narrows the gap on AIME and LiveCodeBench.

Surprising findings

• Under soft speedup instructions, panic (abandoning reasoning within a few tokens) accounts for over 90% of new errors on hard tasks.
• Small Qwen3-1.7B shows reasoning leakage via post-hoc reasoning after the answer even under hard interrupts.

Limitations

• Focuses on single interruption events in single-turn math and coding settings, not agentic or multi-turn workloads.
• Closed-source models cannot be evaluated faithfully because APIs do not support mid-reasoning intervention; GPT-5.4-Mini is only a proxy.
• Paper is diagnostic and proposes no training or control method to mitigate the failure modes.

Experimental setup

Datasets: GSM8K MATH-500 AIME-24 AIME-25 LiveCodeBench-v6

Benchmarks: GSM8K MATH-500 AIME-24 AIME-25 LiveCodeBench-v6

Metrics: pass@1 accuracy output length panic rate doubt rate

Baselines: Chain of Draft NoThinking full reasoning (no interruption) GPT-5.4-Mini (proxy)

Builds on: budget forcing NoThinking Chain of Draft

Closest papers by meaning

• Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · similarity 0.89)
• Answering the Unanswerable Is to Err Knowingly: Analyzing and Mitigating Abstention Failures in Large Reasoning Models (AAAI 2026 · similarity 0.88)
• CodeCrash: Exposing LLM Fragility to Misleading Natural Language in Code Reasoning (NeurIPS 2025 · similarity 0.88)
• On Reasoning Strength Planning in Large Reasoning Models (NeurIPS 2025 · similarity 0.87)
• AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (NeurIPS 2025 · similarity 0.87)
• Reasoning Structure of Large Language Models (ICML 2026 · similarity 0.87)