Frozen Judges, Moving Agents: Version-Dependent LLM-Judge Error and the Limits of Judge-Assisted Agent Evaluation
cs.LG
Submitted: 2026-09-28
Updated: 2026-09-29
Sources
- From Confident Closing to Silent Failure: Characterizing False Success in LLM Agents
- SWE-Bench+: Enhanced Coding Benchmark for LLMs
- PPI++: Efficient Prediction-Powered Inference
- GAUGE: When Not to Trust LLM-as-a-Judge in User-Simulated Evaluation of Task-Oriented Agents
- Bias and Uncertainty in LLM-as-a-Judge Estimation
- What Does an LLM-Agent Leaderboard Rank Actually Compare?
- Who Drifted: the System or the Judge? Anytime-Valid Attribution in LLM Evaluation Pipelines
- AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectories
- Self-Preference Bias in LLM-as-a-Judge
- An Illusion of Progress? Assessing the Current State of Web Agents
- Large Language Model Critics for Execution-Free Evaluation of Code Changes
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks