Multitask learning models commonly assume that task identity is available at inference time, which limits their applicability in real-world scenarios where such information may be missing or unreliable. This paper addresses task-agnostic inference in multitask binary classification, where both the task and its corresponding label must be inferred simultaneously under uncertainty. We propose an uncertainty-aware inference framework based on intuitionistic fuzzy sets (IFS) to model task-related evidence. For each task, membership, non-membership, and hesitation degrees are computed to explicitly represent supportive, contradictory, and ambiguous evidence, respectively. A hesitation-driven training objective is introduced to suppress interference from irrelevant tasks and improve robustness in task-agnostic settings. Experiments on the GLUE benchmark demonstrate that the proposed approach consistently improves performance on low-resource tasks while maintaining competitive accuracy on high-resource tasks. Beyond predictive performance, the proposed framework provides interpretable task-level evidence, offering insights into decision-making under task ambiguity. These results indicate that intuitionistic fuzzy inference provides an effective and interpretable alternative to conventional task routing strategies for uncertainty-aware multitask learning.
Sơn et al. (Fri,) studied this question.