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

TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 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.

Proposes TERMINATOR, a lightweight probe (one copied transformer layer plus a prediction head) trained on hindsight-optimal first-answer positions to early-exit LRM chain-of-thought at inference.

Problem

Large reasoning models overthink, generating many CoT tokens after the final answer has already first appeared; existing early-exit methods do not identify or exploit the optimal (hindsight) reasoning length for practical tasks.

Approach

Define hindsight-optimal reasoning length as the first logical arrival of the final answer in the CoT, automatically label tens of thousands of CoTs with an LRM-based extract-identify-verify pipeline, and train a token-level binary probe on final-layer hidden states to predict when to inject

Key results

• Reduces CoT length by 14%-55% on average across four datasets while outperforming prior early-exit methods.
• Halves average latency on MATH-500 for Qwen3-8B (32.68s to 14.10s) despite a 10.8% overhead of the probe per model compute, per the paper's latency analysis.
• Achieves best or second best performance on 28 of 32 metrics in the main results table.

Surprising findings

• Training on a single domain performs comparably to cross-domain training, with cross-domain giving only about 1% higher accuracy at slightly worse compression.
• Dynasor achieves aggressive token reduction but with significant accuracy degradation.

Limitations

• Assumes the earliest final answer exit is sufficient; the LRM may not commit to the exited answer in the full solution.
• Relies on the final answer appearing in the reasoning trace, which may not hold for open-ended generation.
• Labeling pipeline succeeds for only about 70-80% of traces and depends on an LRM judge; GPQA generalization is weak.

Experimental setup

Datasets: AIME (1983-2024) MATH OpenCoder-SFT OpenScience S1-K

Benchmarks: MATH-500 AIME 2025 HumanEval GPQA

Metrics: accuracy compression rate latency throughput Macro-F1

Baselines: Vanilla NoThinking DEER Dynasor Thought Calibration

Builds on: Thought Calibration DEER Dynasor Self-Certainty

Closest papers by meaning

• Dynamic Early Exit in Reasoning Models (ICLR 2026 · similarity 0.93)
• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.91)
• OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping (NeurIPS 2026 · similarity 0.91)
• Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.91)
• Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.91)
• Efficient Reasoning via Constrained Optimization in Latent Space (NeurIPS 2026 · similarity 0.90)