Sequential prediction is challenging in regimes of delayed disambiguation, where early observations are ambiguous and multiple latent explanations remain plausible until sufficient evidence accumulates. Standard approaches based on marginal inference struggle in this setting, either collapsing uncertainty prematurely or failing to recover once informative evidence arrives. We introduce EviTrack, a test-time inference framework that operates over latent trajectories rather than marginal states. EviTrack maintains a set of competing trajectory hypotheses and applies evidence- and likelihood-ratio-based selection to delay commitment until supported by data, drawing inspiration from hypothesis management in multiple hypothesis tracking and track-before-detect. To evaluate this setting, we construct a controlled synthetic benchmark with known latent ground truth that explicitly exhibits delayed disambiguation. At matched inference budget, EviTrack substantially outperforms sampling-based baselines, achieving faster post-disambiguation recovery. These results show that, in delayed disambiguation regimes, moderate trajectory-level selection is more effective than increasing sampling coverage, highlighting selection over sampling as a key principle for reliable sequential inference.

In partial multi-label learning (PML), each training example is associated with a set of candidate labels, among which only some labels are valid. As a common strategy to tackle PML problem, disambiguation aims to recover the ground-truth labeling information from such inaccurate annotations. However, existing approaches mainly rely on heuristics or ad-hoc rules to disambiguate candidate labels, which may not be universal enough in complicated real-world scenarios. To provide a principled way for disambiguation, we make a first attempt to explore the probabilistic graphical model for PML problem, where a directed graph is tailored to infer latent ground-truth labeling information from the generative process of partial multi-label data. Under the framework of stochastic gradient variational Bayes, a unified variational lower bound is derived for this graphical model, which is further relaxed probabilistically so that the desired prediction model can be induced with simultaneously identified ground-truth labeling information. Comprehensive experiments on multiple synthetic and real-world data sets show that our approach outperforms the state-of-the-art counterparts.

Here we describe the additional details of FlaMBé's curation including structured guidelines for each annotation task, corpus curation, and file assembly. All manual curation in FlaMBé was conducted by three annotators who have doctorate level expertise in computational biology. For named entity tagging annotations a set of structured guidelines were followed to ensure consistency. The guidelines given to reviewers are in the annotator guidelines section below. B.1 Tissue and cell type entities Generally, all terms, related synonyms, and text entities that can be mapped to an entry from the tissue, organ, body part, fluid, and cell type branches of the NCI thesaurus were labeled. Instead of a rigid vocabulary fixed on exact matches of NCIThesaurus (NCIT) terms and synonyms, annotators were encouraged to tag any word with the same meaning as an ontology term. For example, "Pancreatic ductal adenocarcinoma" describes cancer of the pancreas, which can be related back to the NCI Thesaurus, and thus was tagged as a "TISSUE". An initial set of rules was provided to each annotator. When one annotator encountered a corner case (e.g., "is neuron a tissue or cell type?") all annotators discussed, reached a consensus, then added the corner case to the set of annotation rules.