ALICE: In-context, Zero-shot, Mutual Information Estimation
A single synthetic-trained model estimates mutual information on unseen data without fitting a new neural estimator each time.
ALICE is presented as an in-context estimator for mutual information when samples are scarce and per-distribution training is impractical. The paper says it estimates rectified-flow velocity fields from samples of an unseen distribution, then computes MI through a fixed identity comparing joint and conditional fields. The authors report validation on a standard benchmark and applications in biology, genetics, and neuroscience data not seen during training. They claim the same model supports different data dimensionalities and sample counts while closing the gap with separately trained neural estimators. HF Daily Papers' note
ALICE is presented as an in-context estimator for mutual information when samples are scarce and per-distribution training is impractical. The paper says it estimates rectified-flow velocity fields from samples of an unseen distribution, then computes MI through a fixed identity comparing joint and conditional fields. The authors report validation on a standard benchmark and applications in biology, genetics, and neuroscience data not seen during training. They claim the same model supports different data dimensionalities and sample counts while closing the gap with separately trained neural estimators. HF Daily Papers' note
score 4