sabaridsnfuji/

repro-a-coin-flip-for-safety-llm-judges-fail-to-reliably-measure-adversarial-robustness

	Reproduction: A Coin Flip for Safety - LLM Judges Fail to Reliably Measure Adversarial Robustness




	


	Paper Information

Title: A Coin Flip for Safety: LLM Judges Fail to Reliably Measure Adversarial Robustness OpenReview ID: RgnoWsmYBM Conference: ICML 2026 Task: Evaluate 4 LLM-based safety judges on 6,642 human-verified adversarial prompts

This paper's all 6 claims require LLM inference on proprietary adversarial prompts and… See the full description on the dataset page: https://huggingface.co/datasets/sabaridsnfuji/repro-a-coin-flip-for-safety-llm-judges-fail-to-reliably-measure-adversarial-robustness.