Entropy Trajectory Shape Predicts LLM Reasoning Reliability: Monotonicity as a Diagnostic Signal

Forum topic · 小凯 · 2026-03-21

Summary

This article explains a research finding that the shape of an LLM's entropy trajectory during chain-of-thought (CoT) reasoning, rather than the total entropy drop, can predict whether the model will arrive at a correct answer. The core concept is Entropy-Trajectory Monotonicity: when entropy decreases (or stays flat) at every reasoning step, accuracy is significantly higher than when entropy occasionally rises. Experiments on GSM8K with Qwen2.5-7B-Instruct showed monotonic chains reached 68.8% accuracy versus 28.6% for chains with two violations (OR=2.50, p=0.0005). Mistral-7B replicated the result with a 34.7 percentage-point gap. Total entropy reduction correlated only weakly (r=-0.06, p=0.31). Token-level confidence is unreliable: Expected Calibration Error drifts from 0.186 to 0.312 across steps. The method costs roughly 1/8 of self-consistency while improving accuracy by 5.8 points on the 73.7% of cases it covers, enabling lightweight reliability filtering for medical, legal, and edge deployments.

Key points

Core finding: Entropy-Trajectory Monotonicity (ETM) — whether entropy monotonically decreases step by step during CoT reasoning — is a strong predictor of correctness, regardless of how much entropy drops in total.
Qwen2.5-7B-Instruct on GSM8K: monotonic chains (0 violations) achieved 68.8% accuracy; one violation dropped accuracy to 50.8%; two violations to 28.6%. Odds ratio 2.50, p=0.0005.
Mistral-7B replication: 72.3% accuracy on monotonic chains vs 37.6% on non-monotonic chains, a 34.7 percentage-point gap (OR=4.33), confirming generality across models.
Shape beats magnitude: total entropy reduction correlated only r=-0.06 (p=0.31) with accuracy, so a small but smooth drop outperforms a large but volatile one.
Token confidence is unreliable: Expected Calibration Error (ECE) drifted from 0.186 at step 1 to 0.312 later in the chain; calibration worsens as reasoning deepens, making raw probability an unsafe reliability signal.
Cost advantage: The ETM-based method requires roughly 1/8 of the tokens of self-consistency (≈1500 vs 40,000 tokens per query), while delivering +5.8 percentage-point accuracy on the 73.7% of cases the filter can score.
Applications: lightweight reliability filtering for high-stakes domains (medical, legal, financial), edge/mobile inference, and real-time assistants that can abstain when the trajectory is non-monotonic.
Future directions: real-time mid-chain correction when entropy rises, adaptive stopping when entropy stabilizes, training-time rewards that encourage monotonic trajectories, and extension to multimodal reasoning.

Background concepts

Entropy in information theory: H = −Σ p(x) · log p(x); higher entropy means a more diffuse probability distribution over next tokens.
Chain-of-Thought (CoT): prompting the model to emit intermediate reasoning steps (Wei et al., NeurIPS 2022) so errors can be localized.
Self-consistency: sampling many CoT chains (e.g., 40) and majority-voting on final answers; effective but expensive (Wang et al., ICLR 2023).

Why monotonicity matters

The authors argue that monotonic descent is a structural, global property of the reasoning process rather than a self-reported confidence value. A non-monotonic step reflects the model reconsidering or backtracking — a frequent symptom of an earlier error that propagates. Cognitive-science findings about human problem solving show an analogous pattern: smooth, coherent reasoning correlates with correct answers more than the sheer quantity of evidence considered.

Practical recipe

1. Generate a small number (e.g., 1–few) of CoT completions for the query. 3. Compute entropy of the next-token distribution at each step. 4. If entropy is non-increasing across all steps, accept the answer with high confidence; otherwise flag for review or fall back to a heavier method.

References

Zhao, X., et al. (2026). Entropy trajectory shape predicts LLM reasoning reliability. arXiv:2603.18940.
Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 35, 24824–24837.
Wang, X., et al. (2023). Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR.
Shannon, C. E. (1948). A Mathematical Theory of Communication. Bell System Technical Journal, 27(3), 379–423.
Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting Hallucinations in Large Language Models Using Semantic Entropy. Nature, 630, 625–630.