#14 · ML / AI methods

Interventions probe whether biomedical model features and adaptations are causal

Across these three preprints, the dominant thread is controlled testing of what drives model behavior, rather than whether a headline metric improves. In molecular encoders, block-level interventions on Uni-Mol and MolCLR are compared against random-direction controls, showing that linear probes do not establish causal roles. SafeDelta applies a split-control criterion to decide when few-shot adaptation actually helps single-cell perturbation prediction. ECG-RSI pursues accuracy gains in ECG model question answering through verified waveform synthesis and feedback-guided interventions. The ECG item is less about causal inference than about intervention-driven improvement, so the shared thread is partial. What is new is the insistence on explicit control conditions before attributing gains or mechanisms to a model component or adaptation step.