Image de couverture de livre

Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Feuer, Benjamin, Goldblum, Micah, Datta, Teresa, Nambiar, Sanjana, Besaleli, Raz, Dooley, Samuel, Cembalest, Max, Dickerson, John P.
Format:	Preprint
Publié:	2024
Sujets:	Machine Learning Artificial Intelligence
Accès en ligne:	Acceder al recurso
Tags:	Ajouter un tag Pas de tags, Soyez le premier à ajouter un tag!

Documents similaires

When Judgment Becomes Noise: How Design Failures in LLM Judge Benchmarks Silently Undermine Validity
par: Feuer, Benjamin, et autres
Publié: (2025)

FineGRAIN: Evaluating Failure Modes of Text-to-Image Models with Vision Language Model Judges
par: Hayes, Kevin David, et autres
Publié: (2025)

Towards Provably Unbiased LLM Judges via Bias-Bounded Evaluation
par: Feuer, Benjamin, et autres
Publié: (2026)

LiveBench: A Challenging, Contamination-Limited LLM Benchmark
par: White, Colin, et autres
Publié: (2024)

TuneTables: Context Optimization for Scalable Prior-Data Fitted Networks
par: Feuer, Benjamin, et autres
Publié: (2024)

A Simple Baseline for Predicting Events with Auto-Regressive Tabular Transformers
par: Stein, Alex, et autres
Publié: (2024)

Strategies for Increasing Corporate Responsible AI Prioritization
par: Wang, Angelina, et autres
Publié: (2024)

Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive
par: Pal, Arka, et autres
Publié: (2024)

When Do Neural Nets Outperform Boosted Trees on Tabular Data?
par: McElfresh, Duncan, et autres
Publié: (2023)

Measuring Style Similarity in Diffusion Models
par: Somepalli, Gowthami, et autres
Publié: (2024)

Transformers Boost the Performance of Decision Trees on Tabular Data across Sample Sizes
par: Jayawardhana, Mayuka, et autres
Publié: (2025)

Commercial LLM Agents Are Already Vulnerable to Simple Yet Dangerous Attacks
par: Li, Ang, et autres
Publié: (2025)

Cascade: Token-Sharded Private LLM Inference
par: Thomas, Rahul, et autres
Publié: (2025)

WILDCHAT-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training
par: Feuer, Benjamin, et autres
Publié: (2025)

When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges
par: Darshan, Parth, et autres
Publié: (2026)

Attribution Quality in AI-Generated Content:Benchmarking Style Embeddings and LLM Judges
par: Abbas, Misam
Publié: (2025)

vTune: Verifiable Fine-Tuning for LLMs Through Backdooring
par: Zhang, Eva, et autres
Publié: (2024)

Dynamic Delayed Tree Expansion For Improved Multi-Path Speculative Decoding
par: Thomas, Rahul, et autres
Publié: (2026)

Large Language Models Must Be Taught to Know What They Don't Know
par: Kapoor, Sanyam, et autres
Publié: (2024)

KestRel: Relational Verification Using E-Graphs for Program Alignment
par: Dickerson, Robert, et autres
Publié: (2024)

The Lie Derivative for Measuring Learned Equivariance
par: Gruver, Nate, et autres
Publié: (2022)

Adaptive Retention & Correction: Test-Time Training for Continual Learning
par: Chen, Haoran, et autres
Publié: (2024)

The No Free Lunch Theorem, Kolmogorov Complexity, and the Role of Inductive Biases in Machine Learning
par: Goldblum, Micah, et autres
Publié: (2023)

Linting Style and Substance in READMEs
par: Mynampaty, Hima, et autres
Publié: (2026)

Participative Management--Style and Substance.
par: Goldberg, Robert L.
Publié: (1979)

RAG-E: Quantifying Retriever-Generator Alignment and Failure Modes
par: Randl, Korbinian, et autres
Publié: (2026)

Ensemble Monitoring for AI Control: Diverse Signals Outweigh More Compute
par: Koran, Eugene, et autres
Publié: (2026)

Multi-Token Prediction via Self-Distillation
par: Kirchenbauer, John, et autres
Publié: (2026)

Unveiling Substance Use Patterns and Dysfunctional Family Dynamics Among Adults in Rehabilitation
par: Bagsit, Micah Esther, et autres
Publié: (2026)

Privacy-Preserving Mechanisms Enable Cheap Verifiable Inference of LLMs
par: Pal, Arka, et autres
Publié: (2026)

Knowing What You Know Is Not Enough: Large Language Model Confidences Don't Align With Their Actions
par: Pal, Arka, et autres
Publié: (2025)

Preserving Cross-Modal Consistency for CLIP-based Class-Incremental Learning
par: Chen, Haoran, et autres
Publié: (2025)

Postpartum Obesity in Cuba: Risk Outweighs Response
par: Calixto Orozco-Muñoz
Publié: (2017)

LiveMathematicianBench: A Live Benchmark for Mathematician-Level Reasoning with Proof Sketches
par: He, Linyang, et autres
Publié: (2026)

Authorship and the Politics and Ethics of LLM Watermarks
par: Räz, Tim
Publié: (2024)

Interpreting Chekhov’s Prose by Leonard A.Polakiewicz. Ars Rossica. Brookline: Academic Studies Press, 2024. 432 pp. $129.00. ISBN 979‐8‐88719‐566‐7
par: Robin Feuer Miller
Publié: (2025)

How Model Size, Temperature, and Prompt Style Affect LLM-Human Assessment Score Alignment
par: Jung, Julie, et autres
Publié: (2025)

Judging Against the Reference: Uncovering Knowledge-Driven Failures in LLM-Judges on QA Evaluation
par: Lee, Dongryeol, et autres
Publié: (2026)

On Evaluating LLM Alignment by Evaluating LLMs as Judges
par: Liu, Yixin, et autres
Publié: (2025)

Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation Is Wasteful
par: Marek, Martin, et autres
Publié: (2025)