Evidence-Bound Reasoning: Neuro-Semantic Verification of Biomedical AI in Glioblastoma Radiogenomics
The paper separates whether a biomedical AI claim is evidence-consistent from whether the underlying prediction is good.
The authors build a framework that turns glioblastoma MRI radiomics into evidence records and machine-checkable claims. In their tests, the verifier hit 100% exact-set accuracy on a 6,620-claim corruption benchmark, while predictive performance could degrade to chance without breaking verification. A small LLM pilot had GPT-5.6 Sol reproduce all 72 prespecified atomic claims, with deterministic verification recovering all expected conditions. External MGMT prediction was weak, which the paper treats as a stress test rather than the main result. ArXiv · AI/CL/LG's note
The authors build a framework that turns glioblastoma MRI radiomics into evidence records and machine-checkable claims. In their tests, the verifier hit 100% exact-set accuracy on a 6,620-claim corruption benchmark, while predictive performance could degrade to chance without breaking verification. A small LLM pilot had GPT-5.6 Sol reproduce all 72 prespecified atomic claims, with deterministic verification recovering all expected conditions. External MGMT prediction was weak, which the paper treats as a stress test rather than the main result. ArXiv · AI/CL/LG's note
score 4