We introduce PhysClosureBench to test whether AI agents can orchestrate complete physics investigations. The benchmark comprises 50 expert‑verified tasks with clearly defined research goals, spanning strong‑field, attosecond and related areas. Each task provides the objective and essential scientific constraints, leaving the agent to design experiments, generate data and validate evidence.
Local scientific performance is measured by the Atomic Score, a 20‑criterion metric with a maximum of 10 points. Overall research closure is assessed by the RC0‑RC4 hierarchy, which checks if indispensable evidence and scope conditions are satisfied.
Among seven frontier LLMs, Claude‑Opus‑5 achieves the highest mean Atomic Score (6.29/10) but reaches RC4 in only 32.7% of attempts. Across the benchmark, 36.2% of attempts scoring at least 8/10 still fall short of RC4, and every model attains RC4 on far fewer tasks in all three runs than it does in at least one run. Case‑level analysis shows that misinterpretation of generated results and unresolved physical applicability limits often prevent full closure.
These results highlight a persistent gap between agents’ local scientific achievements and their ability to achieve complete research closure.
Review