model-evaluation

• Five-model arXiv study says audit setup skews LLM bias tests
• ArXiv model claims 11–15% tail log-score gains; risk buyers get a new benchmark