Research map / Large language models / Alignment and preferences

VibeCheck: Discover and Quantify Qualitative Differences in Large Language Models

ICLR 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.

Introduces VibeCheck, which iteratively discovers and quantifies LLM-distinguishing "vibes" (qualitative output axes) using an LLM proposer and an ensemble of LLM judges, scored for well-definedness, separability, and user alignment.

Problem

Pairwise comparison of LLMs relies on correctness metrics and predefined criteria, missing subjective qualitative differences (vibes) that shape user preference and that are hard to discover and quantify.

Approach

Automatically propose axes of difference between two models' outputs with an LLM, validate each axis with LLM judges for inter-judge agreement (Cohen's kappa), separability (model-matching accuracy), and predictiveness of preference, and iterate on misclassified samples to find new vibes.

Key results

• On Chatbot Arena Llama-3-70b vs GPT-4/Claude-3-Opus, VibeCheck (3 iterations) reaches 80.34% model-matching accuracy vs 72.10% for predefined vibes, with preference prediction of 59.34% vs 61.11% for predefined vibes.
• Vibes predict model identity at 80% and human preference at 61% overall on Chatbot Arena.
• On MATH CoT, GPT-4o vs Llama-3-405b vibes reach 97.09% model-matching and 72.79% preference prediction accuracy; on COCO captions, GPT-4V vs Gemini-1.5-Flash reach 99.13% and 89.02%.

Measured effects reported in the paper

Surprising findings

• Predefined vibes achieve higher preference prediction accuracy (61.11%) than VibeCheck's discovered vibes on the overall Chatbot Arena set (59.34% at 3 iterations).
• VibeCheck's vibes miss human-found insights such as GPT fabricating facts and literal interpretation of questions, which the authors attribute to the proposer's limitations.

Limitations

• Hard to disentangle whether a vibe causally drives human preference versus confounders such as accuracy.
• LLM judges (GPT-4o-mini) are often incorrect and have biases such as self-preference; vibe discovery may miss relevant differences.
• Results vary across runs due to stochastic LLM discovery, hurting reproducibility.

Experimental setup

Datasets: HC3 Chatbot Arena (Llama-3-70b battles) CNN/DailyMail MATH COCO captions (ShareGPT-4V)

Benchmarks: HC3 Chatbot Arena (Llama-3-70b battles) CNN/DailyMail MATH COCO captions

Metrics: model-matching accuracy preference prediction accuracy Cohen's kappa separability score

Baselines: Predefined vibes (GPT-4o-generated 10 axes) Human-discovered vibes (HC3)

Builds on: AutoSxS LLMComparator Zhong et al. automated data analysis (set A contains more X)

Closest papers by meaning

• MM-RLHF: The Next Step Forward in Multimodal LLM Alignment (ICML 2025 · similarity 0.87)
• Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation (NeurIPS 2025 · similarity 0.86)
• Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.86)
• Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) (NeurIPS 2025 · similarity 0.86)
• Paloma: A Benchmark for Evaluating Language Model Fit (NeurIPS 2024 · similarity 0.86)
• Maximizing Signal in Human-Model Preference Alignment (AAAI 2025 · similarity 0.86)