Let the Results Speak: A Replication-First Paradigm for LLM Behavioral Benchmarking 事件

Let the Results Speak: A Replication-First Paradigm for LLM Behavioral Benchmarking arXiv:2605.27914v1 Announce Type: new Abstract: Subjective evaluation of LLM behavior -- empathy, restraint, calibrated emotional tone -- is hard. Human inter-rater agreement on such qualities saturates near rho ~ 0.45, and an LLM-as-judge proxy alone risks circularity: a judge sharing the target's training cohort cannot independently verify it. Anchoring validity to a single human-rater consensus does not exten