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

LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning

NeurIPS 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.

Proposes LYNX, an online cue-triggered early-exit method that trains a small MLP probe on hidden states at reasoning cue tokens using forced-exit labels and calibrates stop decisions with split conformal prediction, reusing one probe across benchmarks and temperatures.

Problem

Large reasoning models overthink, generating long chains of thought past the point of sufficiency, which wastes compute and can reduce accuracy; existing early-exit methods rely on decoding manipulation, verifiers, or hand-tuned thresholds and lack statistical guarantees.

Approach

Attach exit decisions to natural reasoning cue tokens ("hmm", "wait", "alternatively"), label cues by forcing an answer at each cue, train a probe on concatenated hidden states from three layers, and convert probe scores into a conformal stop/continue threshold with user-chosen confidence level c, calibrated once on MATH and reused unchanged elsewhere.

Key results

• On GSM8K with DeepSeek-R1-1.5B at c=0.80, LYNX improves accuracy from 75.97% to 76.80% while cutting tokens by 66.2%.
• On MATH-500 with DeepSeek-R1-1.5B at c=0.80, accuracy rises from 68.00% to 70.20% with 60.1% fewer tokens.
• On CommonsenseQA with DeepSeek-R1-1.5B at c=0.70, accuracy is 40.62% vs 40.54% baseline with 73.9% token reduction, showing zero-shot transfer from math training.

Surprising findings

• LYNX often improves accuracy over greedy baseline on GSM8K and MATH-500 by cutting off late-chain overthinking.
• At aggressive confidence levels (c=0.70), accuracy drops substantially on several benchmarks, e.g., AMC 2023 for DeepSeek-1.5B at T=0.6 falls 43% relative.

Limitations

• Conformal guarantees rely on exchangeability; the probe is calibrated on MATH and reused on other benchmarks, so coverage may shift under domain shift.
• Relies on hand-specified cue tokens, which may not exist in models without explicit thinking segments.
• Forced-exit dataset construction cost is not optimized.

Experimental setup

Datasets: MATH

Benchmarks: GSM8K MATH-500 AMC 2023 AIME 2024 CommonsenseQA

Metrics: pass@1 accuracy average generated tokens token reduction vs baseline

Baselines: DEER Think-or-Not vanilla CoT baseline

Builds on: DEER split conformal prediction

Closest papers by meaning

• Reasoning Models Better Express Their Confidence (NeurIPS 2025 · similarity 0.90)
• Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models (ICML 2026 · similarity 0.90)
• RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models (ICLR 2026 · similarity 0.90)
• Hint Tuning: Less Data Makes Better Reasoners (NeurIPS 2026 · similarity 0.90)
• When Confidence Rises Too Early: Detecting Shortcut Reasoning via Premature Answer Commitment (NeurIPS 2026 · similarity 0.89)
• Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.89)