Research map / Large language models / Agents, code and math

Measuring Weak-to-Strong Legibility of Reasoning Models

NeurIPS 2026 · Large language models › Agents, code and math · approach: Inductive and abstract reasoning evaluation · 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 "transfer utility," a weak-model continuation metric for how legible reasoning traces from 12 open reasoning models are to weaker models, and shows it diverges from accuracy and efficiency and partly predicts weak-monitor precision.

Problem

Existing legibility metrics for reasoning traces focus on conciseness or local coherence and rely on strong judges, so they miss whether weaker models can use a strong model's reasoning for transfer, distillation, or oversight.

Approach

Measure legibility through interaction: feed weak models prefixes of a correct teacher trace at varying depths and score first-order transfer utility (mean weak accuracy across depth), second-order transfer utility (entropy of first-correct depth), and regression rate, alongside efficiency metrics (length, redundancy, backtracking).

Key results

• Overall accuracy is weakly negatively correlated with FOTU (Spearman rho = -0.35); GPT-OSS-120B (81% accuracy) ranks 11th of 12 on FOTU.
• On MATH, FOTU predicts Phi-3-Mini monitor precision (rho = +0.900 sabotage, +0.873 natural), but no FOTU signal survives length control on GPQA.
• Reward model scores show near-zero correlation with FOTU once restricted to common-correct problems (rho about -0.09 to -0.08).

Surprising findings

• The highest-accuracy and most token-efficient reasoners (e.g., GPT-OSS-120B) rank lowest on transfer utility.
• Reward models' moderate correlation with transfer utility vanishes when controlling for accuracy.

Limitations

• Single-shot settings only; does not cover multi-turn agentic deployments.
• Only open-weight reasoning models; proprietary traces are summaries and excluded.
• Legibility measured for model-to-model use, not human readers; backtracking relies on an LLM judge and redundancy threshold chosen manually.

Experimental setup

Datasets: MATH GPQA LSAT

Benchmarks: MATH GPQA LSAT

Metrics: first-order transfer utility second-order transfer utility regression rate token length redundancy backtracking count precision recall F1 Spearman correlation

Baselines: Qwen2.5-Math-PRM-7B Skywork-Reward-V2-Llama-3.1-8B Llama-3.1-8B-Instruct-RM-RB2

Builds on: prover-verifier games weak-to-strong generalization process reward models

Closest papers by meaning

• Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens (NeurIPS 2026 · similarity 0.91)
• The Quality-Utility Paradox: Why High-Reward Data Impairs Small Model Mathematical Reasoning (ICML 2026 · similarity 0.89)
• Are Language Models Efficient Reasoners? A Perspective from Logic Programming (NeurIPS 2025 · similarity 0.88)
• Reasoning Models Sometimes Output Illegible Chains of Thought (NeurIPS 2025 · similarity 0.88)
• Reasoning Structure of Large Language Models (ICML 2026 · similarity 0.88)
• Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Trace (NeurIPS 2026 · similarity 0.88)