Research map / Large language models / Reasoning and chain of thought
Beyond Logits: Metastable Latent Dynamics for Sample-Efficient Best-of-N Selection in LLMs
Xinrong Li, Zidong Zhou, Keyu Shen, Wenhao Zhou, Shangqi Guo
Read the paperCheck your idea against 137,217 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 a best-of-N selection method for LLM reasoning based on metastable latent dynamics rather than output logits or sample agreement, aiming for sample-efficient verification.
Problem
Best-of-N selection for LLM reasoning needs reliable verification; logit-confidence proxies suffer calibration collapse and sample-agreement proxies are costly or brittle.
Approach
Analyzes latent (hidden-state) dynamics of candidate trajectories and uses metastability as a verification signal for selecting among N samples, instead of relying on output logit confidence or self-consistency.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select by LVE instead of Self-Consistency at N=3 | LVE · MATH500 (Qwen, N=3) | Pass@1 accuracy (%): 84.2 -> 87.4 | Improved |
| select by LogNorm-LVE instead of Self-Consistency at N=10 | LogNorm-LVE · MATH500 (Qwen, N=10) | Pass@1 accuracy (%): 87.6 -> 92.2 | Improved |
| select by LogNorm-LVE instead of Self-Consistency at N=5 | LogNorm-LVE · MATH500 (Qwen, N=5) | Pass@1 accuracy (%): 87.0 -> 89.0 | Improved |
| select by LogNorm-LVE instead of Self-Consistency at N=15 | LogNorm-LVE · MATH500 (Qwen, N=15) | Pass@1 accuracy (%): 88.0 -> 91.4 | Improved |
| select by Self-Certainty instead of random choice | Self-Certainty · MATH500 (Qwen, N=10) | Pass@1 accuracy (%): 81.4 -> 70.2 | Hurt |
| add length normalization (LogNorm-LVE vs raw LVE) (ablation) | LVE · MATH500 (Qwen, N=10) | Pass@1 accuracy (%): 90.4 -> 92.2 | Improved |
| add length normalization (LogNorm-LVE vs raw LVE) (ablation) | LVE · MATH500 (Llama, N=10) | Pass@1 accuracy (%): 85.4 -> 85.0 | No clear effect |
| add length normalization (LogNorm-LVE vs raw LVE) (ablation) | LVE · GPQA (Qwen, N=10) | Pass@1 accuracy (%): 55.6 -> 53.5 | Hurt |
| add raw LVE to a length-only feature for ranking candidates (ablation) | length-only classifier · MATH500 (Qwen) | AUROC: 0.517 -> 0.626 | Improved |
| score with LVE hidden-state replay instead of logits-based scoring | LVE · MATH500 (Llama, N=10) | scoring latency (ms): 1257 -> 870 | Improved |
| select by LVE instead of random choice | LVE · Average of four benchmarks (Qwen, N=10) | Pass@1 accuracy (%): 44.4 -> 56.5 | Improved |
| select by LVE instead of Self-Consistency on GPQA | LVE · GPQA (Qwen, N=10) | Pass@1 accuracy (%): 52.5 -> 55.6 | No clear effect |
Surprising findings
- Logit-confidence signals can suffer from calibration collapse, becoming misaligned with correctness.
Experimental setup
Baselines: logit confidence sample agreement
Closest papers by meaning
- Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.89)
- The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models (ICML 2026 · similarity 0.89)
- Scalable Best-of-N Selection for Large Language Models via Self-Certainty (NeurIPS 2025 · similarity 0.89)
- Inference Time Optimization with Confidence Dynamics (ICML 2026 · similarity 0.89)
- A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · similarity 0.88)
- A Simple Model of Inference Scaling Laws (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
