EviTrack: New Framework for Sequential Prediction Under Delayed Disambiguation

EviTrack, a newly developed framework for test-time inference, focuses on sequential predictions in scenarios characterized by delayed disambiguation, where initial observations may be unclear and several latent explanations are feasible until enough evidence is gathered. Conventional methods relying on marginal inference often falter in these circumstances, either resolving uncertainty too soon or failing to adjust when relevant evidence emerges. Instead, EviTrack works with latent trajectories, keeping a collection of competing trajectory hypotheses and utilizing evidence and likelihood ratios to postpone decisions until data supports them. This approach is influenced by hypothesis management techniques in multiple hypothesis tracking and track-before-detect. To assess this framework, researchers created a controlled synthetic benchmark that clearly demonstrates delayed disambiguation. The paper can be found on arXiv with the identifier 2605.19283v1.