KAIST's TIDE Framework Finds 5 Hidden Bugs Where Reactive Agents Find One
TIDE is a new agent framework from KAIST that proactively hunts down multiple hidden problems in your documents and code — without being asked
Takeaways
- What it is: TIDE is a framework from KAIST that enables LLM agents to proactively discover multiple hidden problems in documents and code without any user prompt.
- Core mechanism: Two components work together — iterative discovery (finding problems in rounds, conditioning on what was already found) and thought templates (reusable schemas distilled from past solved cases).
- Key result: TIDE achieves up to 76% resolution coverage on workspace tasks with GPT-5 mini, vs. 49% for single-agent and 29% for parallel multi-agent baselines.
- Surprising finding: Running 10 parallel independent agents still performs worse than TIDE with just 2 iterations — more agents just rediscover the same obvious problems.
- Templates transfer: Templates built by one LLM (e.g., GPT) work comparably when used by a different one (e.g., Gemini), enabling cross-model reuse.
- No code release: The paper does not release code, datasets, or model weights; evaluation splits and template pools are not publicly available.
Most AI agents are fundamentally reactive: you ask, they answer. They fix the bug you point to, summarize the document you share, and schedule the meeting you request. But what about the five other bugs lurking in the same codebase? Or the conflicting spreadsheets quietly heading into a board presentation? A new paper from KAIST introduces TIDE (Template-guided Iterative Discovery and rEsolution), a framework designed to make agents proactively surface multiple hidden problems from context , without any user prompt to get them started.
The problem with waiting to be asked
Agents are widely deployed as assistants over documents, tools, and code. However, they typically act only on explicit user requests, which surface only the problems the user has noticed, while many other important problems coexist, hidden in plain sight, within the broader user context. The paper frames this gap with a vivid example: a budget approval given verbally but not yet recorded in writing, two copies of the same report with conflicting numbers both heading into an upcoming review, a recurring meeting the team has tacitly stopped attending. All of these are discoverable from existing documents and calendar entries , but only if the agent actually looks.
The challenge is harder than it sounds. The hidden problems within a given context are typically multiple in unknown number, and the most salient ones systematically overshadow subtler ones, so single-shot prediction leaves most problems undiscovered. Generic prompting offers no reusable prior on how contextual signals become evidence for a particular class of problem, leading predictions to drift toward generic or speculative claims.

Two mechanisms, one framework
TIDE combines two complementary ideas that attack the problem from different angles: iterative discovery and thought templates.
Iterative discovery surfaces a small batch of candidates per round while conditioning on what has already been found, so subsequent rounds extend coverage. Think of it as a directed search: instead of asking the model to list all problems at once (where it fixates on the obvious ones), TIDE asks it to find a few new problems each round, explicitly telling it what it already found so it cannot repeat itself. The loop runs up to T rounds and terminates early if a round returns nothing new.
Thought templates are reusable schemas distilled from previously solved cases that specify what contextual signals to attend to and how to connect them, anchoring each prediction in a recognizable problem class. Each template is a structured tuple with three fields: