Tack, Laban, Neville: LLMs Falter on Evolving User Intent
TL;DR
- A new arXiv paper argues that strong static-benchmark scores do not transfer to conversations where user intent evolves across turns.
- The authors reuse existing single-turn benchmarks as multi-turn testbeds without new annotation, preserving each task's original evaluation protocol.
- Substantial drops appear across model families, though the abstract does not name specific models, benchmarks, or the size of the drops.
A new paper from Jihoon Tack, Philippe Laban and Jennifer Neville, posted to arXiv, argues that the way we evaluate LLMs is quietly missing one of the things users actually do most: change their mind mid-conversation. Their framework takes existing static benchmarks, the kind where the question is fully specified up front, and rewrites them into multi-turn dialogues where the user's intent is 'incrementally revealed, revised, and at times redirected mid-conversation,' while keeping the original evaluation protocol so scores stay comparable.
The headline finding, in the authors' own words, is that 'strong static-setting performance does not transfer to the evolving-intent setting, with substantial drops across model families.' That is an uncomfortable claim if it holds up, because it says the leaderboards buyers rely on to pick models are measuring a skill, answering a well-specified prompt once, that only partly overlaps with what a collaborative agent has to do in production: track a user who is still figuring out what they want.
The strategic read for anyone shipping agents on top of these models is that the gap between demo and deployment may be less about tool use or memory and more about intent tracking. A win on a fixed-intent test suite is not evidence that the model will hold together when a customer rephrases the goal on turn four. That is worth building your own evals around, especially for support, coding assistants, and any workflow where the user is thinking out loud.
The honest caveat is that the public abstract does not name the model families tested, the benchmarks converted, or the size of the drops observed, so 'substantial' is doing a lot of work until the full paper is read carefully. What the reporting does not give you at the abstract level is whether newer reasoning-tuned models close this gap or fall into it just as hard.
The upside, if the framework generalizes, is that model developers and application teams get a cheap way to stress-test intent tracking without paying for new annotation, since the authors explicitly reuse existing benchmarks as controlled testbeds. That is the part worth watching.
Originally reported by paper
Read the original article →Original headline: Study: LLMs Show Substantial Drops Across All Model Families When User Intent Evolves Mid-Conversation