A paper on HuggingFace documents a phenomenon that anyone working with agents has always suspected: when the user changes their mind mid-conversation, models lose the thread. The problem lies in intent tracking. The model has access to context, but doesn’t know how to recalculate it when the objective shifts.

Researchers took known tasks, evaluated in single-turn form, and transformed them into multi-turn conversations where user intent reveals gradually, gets revised, sometimes gets rerouted halfway through. The evaluation protocol stayed the same, so existing benchmarks serve as a test bed without new annotations.

The result is clear: models that perform well in the static setting collapse when intent evolves. The drop is consistent across different model families. Performance on fully-specified tasks, the metric the entire industry measures itself by, doesn’t predict how a model fares when the user reconsiders what they asked for.

For anyone using agents in daily work, this explains a frustrating and previously unexplainable pattern. You ask for one thing, the model starts well. You add a constraint, correct a premise, change a detail. After two or three of these adjustments, the agent convinces itself of an objective that isn’t what you asked for anymore. It stays consistent step by step, but globally off track.

As we covered on July 25 with ICAE-Bench and WorkBuddy, benchmarks are learning to test agents on vague and incomplete intents. This paper goes a step further: it shows that even when the model starts with the right information, it can’t recalibrate when that information changes.

The practical takeaway: if your agentic workflow involves iteration with course corrections, the model alone isn’t enough. You need an explicit mechanism that tracks the current objective and reinjects it at every turn.

In detail

The paper starts from a simple observation. Models are almost always evaluated on closed tasks: one request, one response, one score. But when you use an agent, conversation works differently. You start with a rough idea, refine it, sometimes flip it. The real user specifies intent little by little, and initial specifications are almost always incomplete or wrong.

To measure this gap, researchers built a framework that takes tasks from existing benchmarks and transforms them into multi-turn conversations. The simulated user reveals intent incrementally: first a vague request, then a correction, then a revision, sometimes a complete redirection. The evaluation protocol stays identical to the original, so you can directly compare static and dynamic performance on the same task.

The result is a widespread drop. Models that score high in static tests lose ground substantially in the dynamic scenario, and the pattern repeats across multiple architectural families. Performance on single-turn, the metric that dominates public discourse about models, is blind to this dimension.

A commenter on HuggingFace, who works with coding agents on long projects, adds a useful detail. The model builds a compressed representation of the project that can drift toward assumptions introduced in earlier iterations. The agent stays consistent with accumulated assumptions, but aligns less and less with the original objective. It’s a subtle failure: step by step everything checks out, but overall you’re lost.

For those building agentic workflows, the implication is operational. If the session involves corrections and revisions, conversation memory alone won’t maintain alignment. You need explicit objective state, updated at every course change and reinjected into context at each turn. The playbook From endless chat to fresh session moves in this direction: compact the conversation into a clean starting context, without losing decisions and constraints.

What remains open. The paper describes the phenomenon and measures it, but solutions are still vague. Explicit intent-tracking mechanisms (structured memory, reinjected objective state) are reasonable proposals, but untested in this work. The evaluation framework, though, is immediately reusable: whoever builds a test bed with evolving intents can measure their agent on a dimension that standard benchmarks ignore.