Research map / Large language models / Alignment and preferences

Bridging Human and LLM Judgments: Understanding and Narrowing the Gap

NeurIPS 2025 · Large language models › Alignment and preferences · approach: LLM-as-judge and feedback data · 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 Bridge, a statistical framework that models a shared latent human preference score and LLM deviations via covariates, fit with a "logit trick" ordinal logistic procedure, to refine LLM judge ratings and test human-LLM discrepancies.

Problem

LLM-as-a-judge ratings diverge systematically from human judgments, and no prior framework jointly calibrates LLM scores to humans and formally tests which features cause the gap.

Approach

Assume both humans and LLMs are ordinal-logit responses to a shared latent preference, with LLM latent score a linear transform (beta Z + gamma^T X) of the human score; estimate LLM latent scores from judge output probabilities (logit trick) and fit human-side parameters by MLE, yielding asymptotically normal estimates for inference on gamma.

Key results

• Across six LLM judges and both datasets, Bridge (ordinal) and LogReg consistently outperform raw LLM probabilities on cross-entropy, calibration error, and accuracy, with the largest gains on BigGen Bench.
• Longer responses receive systematically lower scores from LLM judges than from humans (negative text-length gamma for all judges on both datasets, e.g., -0.83 for GPT-4.1-nano on BGB).
• Human annotators reward creativity/engagement more than LLM judges on Chatbot Arena (e.g., -1.27 for GPT-4.1-nano, -0.32 for GPT-4.1).

Measured effects reported in the paper

Surprising findings

• LLM judges favor brevity relative to humans, contrary to the common view that LLM judges are biased toward longer responses.
• For technical queries on Chatbot Arena, human-LLM divergences were not statistically significant.

Limitations

• Vulnerable to model misspecification (unrealistic distributional assumptions or omitted covariates).
• Constructing informative covariates X requires domain knowledge; analysis uses observational data so estimates are associations, not causal effects.
• Restricted to ordinal discrete ratings in main experiments; continuous or unordered extensions are in appendix; reduced power on subsets.

Experimental setup

Datasets: BigGen Bench Chatbot Arena (arena-human-preference-100k)

Benchmarks: BigGen Bench Chatbot Arena

Metrics: cross-entropy loss calibration error accuracy KL divergence p-value (FDR-adjusted)

Baselines: Raw LLM probabilities LogReg (multinomial logistic regression on raw probabilities) In-context learning judge

Builds on: Platt scaling Bradley-Terry model ordinal logistic regression AlpacaEval 2.0 Prometheus

Closest papers by meaning

• Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.92)
• From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation (NeurIPS 2026 · similarity 0.92)
• Maximizing Signal in Human-Model Preference Alignment (AAAI 2025 · similarity 0.91)
• Polyrating: A Cost-Effective and Bias-Aware Rating System for LLM Evaluation (ICLR 2025 · similarity 0.91)
• Validating LLM-as-a-Judge Systems under Rating Indeterminacy (NeurIPS 2025 · similarity 0.91)
• Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.91)