New benchmark shows some LLMs refuse over 50% of legitimate biology queries
nlarusstone · x · 2026-11-07
Benchling released BenchBench-Refusals, a new biosecurity benchmark measuring how well LLMs distinguish harmful biology requests from legitimate ones — what the author calls "stuff every biologist knows," now with numbers.
Key findings:
- Some models refuse more than 50% of legitimate biology queries, over-blocking normal academic work
- Refusal rate is not correlated with model quality: some of the strongest models are the best calibrated, answering correctly while still refusing harmful requests
- The scatter plot separates models into broad, selective, misplaced, and limited refusal quadrants; some smaller models barely refuse anything
The takeaway: biosafety guardrail calibration is a training-alignment choice, not a byproduct of raw model capability.
More from Research
- Google: stale TODO comments derail 1 in 7 coding agent sessions — rseroter · 2026-11-07
- JHU paper: treating demographic annotation disagreement as the target for valid LLM annotation — krisgligoric · 2026-11-07
- VLM video annotation shortcuts: highest-accuracy configs can yield wrong conclusions, JHU study finds — krisgligoric · 2026-11-07
- Applied Compute shares recipe for production RL with continual learning — ypatil125 · 2026-11-07
- Muir linear attention streams 10M tokens in 25MB state, trains 2.7x faster than Transformer at 32K — _akhaliq · 2026-11-07
- ICLR slammed for poor score discrimination: 4-point scale squeezes papers into one narrow band — TuhinChakr · 2026-11-07