LongChang Benchmark Tests Whether LLMs Can Discover Altered Physics Laws Alone
Zhen Yang · hf · 2026-10-28
Researchers from HKUST(GZ), Alibaba, ZJU and HKUST introduce LongChang, an automated framework that builds simulated worlds with altered physical laws and evaluates whether LLMs can notice unexpected behavior (like heavier objects falling slower) and discover the hidden law through free experimentation without being told what to investigate.
The evaluation finds only one model achieved limited success, suggesting current LLMs largely lack autonomous scientific discovery abilities.
More from AGI Musings
- AI labs are losing the narrative war and need a three-level cooperation vision — sudoraohacker · 2026-10-28
- The 'AI stink' will devalue video like it did text, argues viral thread — heyneighbor · 2026-10-28
- Miles Brundage: LLMs were 'born cheap' relative to human labor, driving rapid adoption — Miles_Brundage · 2026-10-28
- Top safety researcher admits misreading Anthropic report's 'AI safety is solved' headline — geoffreyirving · 2026-10-28
- Investor Rob Leclerc: AI Will Hit Hard Limits — "It Is Not a God" — robleclerc · 2026-10-28
- 'Superintelligence Is Here, You Can Build Anything': A Casual PSA — RileyRalmuto · 2026-10-28