Title: TIDE: Proactive Multi-Problem Discovery via Template-Guided Iteration

Abstract:

AI agents are increasingly utilized as assistants across documents, codebases, and various tools. However, their operational mode is typically reactive, responding solely to explicit user prompts. This approach limits discovery to issues the user has already identified, leaving numerous other significant problems obscured within the broader context—often unnoticed and with their total quantity unknown beforehand. We address this gap by defining the challenge of uncovering multiple latent issues embedded in the context. This task requires not only identifying these coexisting problems but also grounding them in evidence and pairing them with actionable solutions.

To tackle this, we present TIDE, a framework driven by iterative processes and guided by templates, incorporating two synergistic mechanisms. First, we introduce iterative discovery. Inspired by the limitation of single-pass predictions—which tend to fixate on the most prominent cases and produce generalized claims—our method reveals a limited batch of candidates in each round. By conditioning on findings from previous iterations, subsequent rounds progressively expand the scope of coverage. Second, we employ thought templates: reusable schemas derived from previously resolved cases. These templates delineate which contextual signals to prioritize and how to link them, thereby anchoring each prediction within a distinct and recognizable problem category.

We evaluated TIDE in two practical environments: personal workspaces and software repositories. Testing across four different model backbones demonstrated that TIDE achieves significant improvements over both single-shot methods and parallel multi-agent baselines in terms of task coverage, problem identification, and resolution efficacy.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC