A single masked Riemannian flow-matching model over paired CLIP embeddings gives calibrated aleatoric retrieval entropy and epistemic marginal-typicality scores without retraining the frozen encoder.
Elements of information theory (wiley series in telecom- munications and signal processing)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
GeoFlowVLM: Geometry-Aware Joint Uncertainty for Frozen Vision-Language Embedding
A single masked Riemannian flow-matching model over paired CLIP embeddings gives calibrated aleatoric retrieval entropy and epistemic marginal-typicality scores without retraining the frozen encoder.