ALICE: a foundation model for in-context, zero-shot mutual information estimation
eurecom-probai · hf · 2026-09-30
ALICE is a foundation model for mutual information (MI) estimation that removes per-distribution training.
- Existing neural estimators need large data, retraining per distribution, and fixed data types; ALICE trains only on a family of synthetic distributions.
- It estimates rectified-flow velocity fields in-context: conditioned on samples of an unseen distribution, it predicts the field, then computes MI via a fixed identity integrating the squared joint-conditional difference.
- Validated on a standard benchmark and applied zero-shot to biology, genetics, and neuroscience—first single model to close the gap with per-distribution neural estimators while natively supporting varying dimensionality and sample sizes.
More from Research
- Liquid AI's LongevityBench: compact LFMs beat frontier models on several aging tasks — JosephJacks_ · 2026-10-01
- AlphaFold Database adds predicted structures for 2,800+ viruses, 30% interactions undocumented — AllThingsApx · 2026-10-01
- Nominate papers worth keeping alive as AI enables continuously updated research — ChenhaoTan · 2026-10-01
- Microsoft's ML System Forecasts Grid Risks from Space Weather for 66,935 US Substations — Microsoft Research · 2026-10-01
- Empirical test: Isolation Forest performs best with max_samples=1.0 on IDS data — Ragno_ · 2026-09-30
- Google Figures Out How to Watermark AI-Designed Proteins for Biosecurity — Ars Technica AI · 2026-09-30