BenchBench-Refusals: some models over-refuse half of benign biology questions

Benchling released BenchBench-Refusals, a benchmark measuring whether frontier models can distinguish harmful from benign biology requests, finding some models over-refuse more than half of ordinary biological terms.

2026-11-07 ~ 2026-11-07 · 2 related posts

• New benchmark shows some LLMs refuse over 50% of legitimate biology queries — nlarusstone · 2026-11-07

1 near-duplicate retellings: nlarusstone