# Automated Concept Discovery for LLM-as-a-Judge Preference Analysis

arXiv: https://arxiv.org/abs/2603.03319 · HTML: https://mlsanity.com/paper/2603.03319/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 22 of its 28 references that are in the map.
- First posted on arXiv 9 Feb 2026.

## Builds on (most cited first, 22 of 22)

- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [A Survey on LLM-as-a-Judge](https://mlsanity.com/paper/2411.15594/) (2024)
- [Concept Bottleneck Models](https://mlsanity.com/paper/2007.04612/) (2020)
- [Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge](https://mlsanity.com/paper/2410.02736/) (2024)
- [Humans or LLMs as the Judge? A Study on Judgement Biases](https://mlsanity.com/paper/2402.10669/) (2024)
- [JudgeBench: A Benchmark for Evaluating LLM-based Judges](https://mlsanity.com/paper/2410.12784/) (2024)
- [Understanding Dataset Difficulty with V-Usable Information](https://mlsanity.com/paper/2110.08420/) (2021)
- [Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge](https://mlsanity.com/paper/2406.07791/) (2024)
- [Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judges](https://mlsanity.com/paper/2406.12624/) (2024)
- [Learning Multi-Level Features with Matryoshka Sparse Autoencoders](https://mlsanity.com/paper/2503.17547/) (2025)
- [A Holistic Approach to Unifying Automatic Concept Extraction and Concept Importance Estimation](https://mlsanity.com/paper/2306.07304/) (2023)
- [An Empirical Study of LLM-as-a-Judge for LLM Evaluation: Fine-tuned Judge Model is not a General Substitute for GPT-4](https://mlsanity.com/paper/2403.02839/) (2024)
- [Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions](https://mlsanity.com/paper/2504.15236/) (2025)
- [Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment Dataset](https://mlsanity.com/paper/2507.09650/) (2025)
- [What's In My Human Feedback? Learning Interpretable Descriptions of Preference Data](https://mlsanity.com/paper/2510.26202/) (2025)
- [Annotation alignment: Comparing LLM and human annotations of conversational safety](https://mlsanity.com/paper/2406.06369/) (2024)
- [Disentangling Dense Embeddings with Sparse Autoencoders](https://mlsanity.com/paper/2408.00657/) (2024)
- [Explaining Datasets in Words: Statistical Models with Natural Language Parameters](https://mlsanity.com/paper/2409.08466/) (2024)
- [Position: Use Sparse Autoencoders to Discover Unknowns](https://mlsanity.com/paper/2506.23845/) (2025)
- [AI vs. Human Judgment of Content Moderation: LLM-as-a-Judge and Ethics-Based Response Refusals](https://mlsanity.com/paper/2505.15365/) (2025)
- [Sparse Autoencoders are Topic Models](https://mlsanity.com/paper/2511.16309/) (2025)
- [Uncovering Factor Level Preferences to Improve Human-Model Alignment](https://mlsanity.com/paper/2410.06965/) (2024)

## Built on by (most cited first, 1 of 1)

- [Automatic Hard Example Synthesis with Multi-Level Agentic Data Curation](https://mlsanity.com/paper/2607.14256/) (2026)

## Abstract (arXiv)

Large Language Models (LLMs) are increasingly used as scalable evaluators of model outputs, but their preference judgments exhibit systematic biases and can diverge from human evaluations. Prior work on LLM-as-a-judge has largely focused on a small, predefined set of hypothesized biases, leaving open the problem of automatically discovering unknown drivers of LLM preferences. We address this gap by studying several embedding-level concept extraction methods for analyzing LLM judge behavior. We compare these methods in terms of interpretability and predictiveness, finding that sparse autoencoder-based approaches recover substantially more interpretable preference features than alternatives while remaining competitive in predicting LLM decisions. Using over 27k paired responses from multiple human preference datasets and judgments from three LLMs, we analyze LLM judgments and compare them to those of human annotators. Our method both validates existing results, such as the tendency for LLMs to prefer refusal of sensitive requests at higher rates than humans, and uncovers new trends across both general and domain-specific datasets, including biases toward responses that emphasize concreteness and empathy in approaching new situations, toward detail and formality in academic advice, and against legal guidance that promotes active steps like calling police and filing lawsuits. Our results show that automated concept discovery enables systematic analysis of LLM judge preferences without predefined bias taxonomies.