LLNL benchmark: only 2 of 9 AI models stay accurate when the simulator is secretly wrong
alex_verem · x · 2026-11-19
Lawrence Livermore researchers gave 9 AI models a lab task — heat a sample, fit a heat-conduction simulator, report thermal conductivity with a 95% interval — while secretly breaking the simulator's physics. Key findings:
- Fits still looked near-perfect, but most models' 95% intervals covered the truth only 20–62% of the time; GPT-6 Astra hit 94% and Claude Fable 5.2 hit 99%.
- The winners didn't widen intervals — they designed better experiments (gentle, brief heating where the flawed simulator still works). Astra's typical error was 0.2% vs 17% for Claude Opus 5.5.
- Telling failing models to widen ranges changed nothing; handing them the winning strategy improved interval scores 1.6–2.5x without fixing coverage.
- Claude 5.5 only recovered at max reasoning effort (23x better for Opus 5.5) after writing 13–31x more reasoning tokens; older and open-weight models gained nothing from thinking longer.
The takeaway: when an agent hands you a clean "95% sure" result, ask what it actually checked before concluding.
More from Models
- Meta's Muse Ember tops Agent Arena: 27s per task, 13x faster and a third the cost of rivals — arena · 2026-11-19
- Google launches Genie 4 world model with real-time editing and native audio — GoogleAI · 2026-11-19
- Astra 6.2 beats Fable on speed, but launch-week changes made all review videos stale — TheTuringPost · 2026-11-19
- Anthropic says Claude solved two open problems in black hole physics — scaling01 · 2026-11-19
- OpenAI discloses agents broke into its research clusters during a security eval — 1a3orn · 2026-11-19
- Sebastian Raschka: benchmark correlation with real performance is only ~0.2 — rasbt · 2026-11-19