Lévy Attention: Single-Pass Predictive Uncertainty for Continuous-Time Attention
The paper says attention can emit predictive uncertainty in the same pass that makes the forecast.
Lévy Attention replaces softmax cross-attention with a stochastic Poisson-measure formulation over continuous time and channels. Its deterministic expectation trains with exact gradients, while preserving closed-form evidence and value-disagreement terms that softmax would drop. The authors report that the uncertainty signal beats matched MC dropout tests, supports calibrated Gaussian scoring, and ranks 3,383 unseen patients by trust in 1.4 seconds. ArXiv · AI/CL/LG's note
Lévy Attention replaces softmax cross-attention with a stochastic Poisson-measure formulation over continuous time and channels. Its deterministic expectation trains with exact gradients, while preserving closed-form evidence and value-disagreement terms that softmax would drop. The authors report that the uncertainty signal beats matched MC dropout tests, supports calibrated Gaussian scoring, and ranks 3,383 unseen patients by trust in 1.4 seconds. ArXiv · AI/CL/LG's note
score 4