NEX: Neuron Explore-Exploit Scoring for Label-Free Chain-of-Thought Selection and Model Ranking

2026-02-05

AI Summary

This work addresses the selection bottleneck faced by large language models in multi-path reasoning or checkpoint merging, where supervision signals from the target distribution are absent. The authors propose the first white-box, unsupervised exploration-exploitation (E-X) scoring framework grounded in dynamic neuron activation patterns. By modeling MLP neuron activations via a two-state hidden Markov model and leveraging sparse activation caching, the method integrates neuron reuse analysis with a Good-Mass Fraction metric to construct a label-free quality scoring mechanism. Requiring only minimal unlabeled data, this approach accurately predicts downstream performance, ranks model variants across multiple reasoning benchmarks and the Qwen3 merged model family, and demonstrates strong validity and interpretability through human evaluation and causal transfer experiments.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Relational Probabilistic Models
• Natural Language Processing: (Large) Language Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Large language models increasingly spend inference compute sampling multiple chain-of-thought traces or searching over merged checkpoints. This shifts the bottleneck from generation to selection, often without supervision on the target distribution. We show entropy-based exploration proxies follow an inverted-U with accuracy, suggesting extra exploration can become redundant and induce overthinking. We propose NEX, a white-box label-free unsupervised scoring framework that views reasoning as alternating E-phase (exploration) and X-phase (exploitation). NEX detects E-phase as spikes in newly activated MLP neurons per token from sparse activation caches, then uses a sticky two-state HMM to infer E-X phases and credits E-introduced neurons by whether they are reused in the following X span. These signals yield interpretable neuron weights and a single Good-Mass Fraction score to rank candidate responses and merged variants without task answers. Across reasoning benchmarks and Qwen3 merge families, NEX computed on a small unlabeled activation set predicts downstream accuracy and identifies better variants; we further validate the E-X signal with human annotations and provide causal evidence via"Effective-vs-Redundant"neuron transfer.

Problem

• chain-of-thought selection
• model ranking
• label-free evaluation
• reasoning efficiency
• unsupervised scoring

Innovation

• Explore-Exploit Scoring
• Label-Free Model Selection
• Chain-of-Thought Ranking
• Sparse Neuron Activation
• White-Box Interpretability