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

Dynamic Early Exit in Reasoning Models

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 1 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.

Proposes DEER, a training-free method that makes large reasoning models self-truncate long chain-of-thought by inducing trial answers at transition points and exiting early when confidence exceeds a threshold.

Problem

Overthinking in long chain-of-thought of large reasoning models wastes compute and can reduce accuracy; static early-exit heuristics are suboptimal because the optimal exit point varies per problem.

Approach

At reasoning transition points (linguistic markers like "Wait" or high-entropy steps), the model is prompted to produce a trial answer; its geometric-mean token confidence decides whether to terminate thinking. DEER-PRo runs multiple answer-inducing prompts and calibrates confidence with a MAD penalty.

Key results

• DEER improves accuracy by 0.9 to 4.8 points while reducing sequence length by 19.1% to 42.9% versus vanilla models on the main benchmarks.
• Abstract reports length reduction of 19.1% to 80.1% and accuracy gains of 0.3% to 5.0% across 10 benchmarks and 11 LRMs.
• Early-exit accuracy stays between 88% and 98% per the paper, and DEER's probe-free confidence reaches about 95% accuracy on MATH-500 versus about 80% for a trained probe in a concurrent study.

Surprising findings

• Nearly all baselines fail on QwQ-32B due to invalid end-of-thinking delimiter generation, while DEER still reduces length by 19.1%.
• Smaller models generate longer reasoning with more overthinking, so DEER yields greater length reduction for them.

Limitations

• Confidence threshold lambda and entropy threshold were chosen empirically; robustness is shown only within a range (0.9-0.97 for the main model).
• Relies on the model reliably emitting transition markers or ending thinking; invalid </think> delimiters hurt some models.
• Added latency from trial answer inducing, especially for code tasks with lengthy trial answers, reduces efficiency gains without branch-parallel decoding.

Experimental setup

Datasets: GSM8K MATH-500 AMC 2023 AIME 2024 AIME 2025 OlympiadBench GPQA Diamond HumanEval BigCodeBench LiveCodeBench

Benchmarks: GSM8K MATH-500 AMC 2023 AIME 2024 AIME 2025 OlympiadBench GPQA Diamond HumanEval BigCodeBench LiveCodeBench

Metrics: accuracy token count compression rate pass@1 latency

Baselines: Vanilla TCC CoD NoThinking Dynasor-CoT SEAL

Builds on: Dynasor-CoT Chain-of-Draft Token-Conditional Control SEAL

Closest papers by meaning

• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.93)
• A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.93)
• Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.92)
• Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning (NeurIPS 2025 · similarity 0.92)
• Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.92)
• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.91)