All five tested LLMs (GPT-4o, Gemini 2.5, DeepSeek, Kimi, Llama 3.1) score near zero on multi-sensor safety scenarios where individual readings stay below thresholds but combined context requires precaution—don't use them as safety monitors without additional validation layers.
Summary
If you're building physical safety systems with LLMs, this reveals a critical blind spot: models detect single-threshold violations reliably but systematically fail to issue warnings when multiple sensors show correlated elevation patterns. Your safety-critical pipeline needs explicit multi-sensor fusion logic outside the model.
Why it matters
If you're building physical safety systems with LLMs, this reveals a critical blind spot: models detect single-threshold violations reliably but systematically fail to issue warnings when multiple sensors show correlated elevation patterns. Your safety-critical pipeline needs explicit multi-sensor fusion logic outside the model.
Implementation verdict
This doesn't replace existing sensor fusion algorithms—it confirms LLMs shouldn't replace them. Structured tabular input shows no advantage over prose (p=0.001 favors prose in GPT-4o). Use LLMs for anomaly narrative/logging only; implement deterministic multi-sensor thresholds and decision trees separately. Not ready as primary safety arbiter.
Sources
Dev Signal
Get briefs like this in your inbox — free, every weekday.
100+ sources compressed into one 4-minute read. Ranked, cited, implementation-ready.