Treating the Symptom, Missing the Cause: Anatomy of a Scientific Reasoning Failure
| Source: arXiv AI
Tags: Claude Opus 4.6, Claude Sonnet 4.6, Scientific Agents, Causal Reasoning, AI for Biology
In synthetic biology-style tests, seven LLM agents fixed faulty systems more often than they diagnosed them; Claude Opus 4.6 rescued every organism but identified the cause in only 55% of cases.
Details
The authors build controlled synthetic regulatory systems where agents must experimentally find a hidden fault and rescue the resulting phenotype. Across seven LLMs, every model restored healthy phenotypes more often than it identified the fault. Claude Opus 4.6 and Sonnet 4.6 rescued every organism while diagnosing only 55% and 35% respectively. Agents encountered distinguishing evidence in 91% of relevant sessions and still often misdiagnosed. Probes show models read decisive evidence correctly and revise when the comparison is explicit, but fail to predict experiment outcomes under each competing cause. The authors argue scientific AI should be evaluated on causal conclusions, not only on whether interventions work.