HKUST's LongChang Catches AI Agents Inventing Broken Physics That Never Existed
A new benchmark drops LLMs into simulated worlds with rewritten physics and asks them to notice something is off. Only one model does.
- LongChang tests whether LLMs can notice and explain altered physics with no task or hints given.
- Only GPT-6-Astra discovered hidden laws, scoring 16.7% while 11 other frontier models scored 0%.
- Altered laws included heavier objects falling slower, red balls attracting, and oscillating gravity strength.
- Explorer agents manipulate objects, build custom tools, and write reports judged by separate LLM agents.
- In normal-physics control worlds, about one third of reports falsely claimed something was wrong.
- Code and 540 run records are available at EnVision-Research/LongChang.
Researchers at HKUST (Guangzhou), Alibaba, and Zhejiang University have introduced LongChang, an automated benchmark for open-ended scientific discovery. An explorer model enters a simulated glass room, manipulates objects, receives textual observations, and tries to identify hidden changes to the laws of physics without an assigned task or hints.
Among 12 evaluated language models, the benchmark reports a 16.7% discovery rate for GPT-6-Astra and 0% for the other 11. In control worlds governed by ordinary physics, roughly one-third of reports still claimed an anomaly. The results expose two risks for autonomous agents: overlooking contradictory evidence and inventing explanations for changes that never occurred.
The task begins before the hypothesis
Many scientific-reasoning evaluations provide a hypothesis, dataset, or experimental goal. LongChang starts earlier by testing whether a model can recognize an unexpected event, decide that it merits investigation, design experiments, and support an explanation with evidence.
| Measure | Reported result |
|---|---|
| Explorer models | 12 language models |
| Evaluation protocol | Three runs per model per scene |
| Total runs | 540 |
| Highest discovery rate | GPT-6-Astra at 16.7% |
| Other discovery rates | Eleven models at 0% |
| Normal-physics controls | About one-third of reports alleged an anomaly |
| Judge model | GPT-6.1-Sol |
Inside the glass room
Each explorer inhabits a room containing objects such as pendulums, balances, and colored balls. It can push, drag, rotate, and lift objects, then read text describing the result. The model can also propose new measurement tools, which a separate agent reviews before installation. The prompt supplies neither a research question nor notice that the physics has changed.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
