LLMs Get Lost in Evolving User Intent
Jihoon Tack, Philippe Laban, Jennifer Neville
cs.LG
2026-07-23
A framework converts any verifiable single-turn benchmark into multi-turn conversations where intent evolves (reveal, revise, switch); frontier models drop sharply, exposing a gap static evaluation cannot see.
In real collaboration, users almost never specify intent up front. They disclose, revise, and sometimes pivot mid-conversation. Yet most LLM evaluation and training still runs in single-turn, fully-specified settings, leaving a fundamental question open: can models track and act on a user's intent as it shifts across a multi-turn conversation?
This is hard to test because multi-turn benchmarks are expensive and resist clean ground truth. Every new task needs hand-authored conversations and labels, trajectories diverge, and existing benchmarks fall back on LLM-as-judge, whose reliability is itself an open question.
Rather than build multi-turn environments from scratch, the authors start from existing verifiable single-turn benchmarks, which already provide scalable, verifiable supervision. The key move is to anchor the source task's user intent at the final turn, then retrospectively construct the preceding turns.
Intent is formalized as a structured state evolving along three transitions:
The last turn must land exactly on the source intent, so the final action can be scored by the source dataset's original verifier, with no new annotation and no LLM judge. Earlier turns are synthesized under controlled dynamics: counterfactual argument values for revision, predecessor tasks for switch, recursively chainable to arbitrary depth.
Four domains are tested: math (GSM8K), text-to-SQL (BIRD), search (BrowseComp+), coding (SWE-Bench Verified). The main experiment fires each transition twice, six transitions over seven turns.
Static performance largely fails to transfer. GPT-5.5 drops from 99.0% on GSM8K single-turn to 80.5% after six transitions. Drops are broad across families and steepest in coding and search.
| Model | GSM8K single to evolve | BIRD single to evolve | SWE-Bench single to evolve |
| GPT-5.5 | 99.0 to 80.5 | 80.0 to 71.0 | 86.0 to 80.0 |
| DeepSeek V3.2 | 96.5 to 78.5 | 76.0 to 53.0 | 76.0 to 76.0 |
| Kimi K2.5 | 97.0 to 75.5 | 75.0 to 66.0 | 82.0 to 70.0 |
| Mistral Large 3 | 95.5 to 73.5 | 61.0 to 56.0 | 56.0 to 0.0 |
Several findings stand out. On SWE-Bench, several agents (GPT-5.1 and others) routinely exhaust the 100-call per-turn budget, stuck in extended thinking until timeout. More counter-intuitively, the multi-turn setting grants more total tool calls, yet search and coding degrade the most. A close look at SWE shows the model spends most of its budget on exploration (sed, grep, find) rather than execution (pytest, applyedits), with fewer than 4 of 100 calls per turn being execution-related. More tool access does not buy better intent adaptation; accumulated context and tool traces become distractors.
Function switch is the hardest transition, worsening when composed with others. A turn-wise intent-tracking analysis shows reveal and revision tracking stay near-perfect while switch degrades markedly across multiple occurrences. Two simple memory mechanisms (prompt recap and oracle recap) recover some ground; oracle recap lifts GPT-5.5 from 65% to 75% under function switch, still short of the 80% single-turn figure.
The paper exposes a capability gap that static evaluation cannot see: maintaining a belief over what the user currently wants, and acting on it, across turns. For agent builders, this means do not trust single-turn benchmark scores; the real bottleneck in long-horizon collaboration is intent tracking, and more tool calls do not fix it. The framework can also mass-produce training data from existing benchmarks; the authors train Qwen3-4B with GRPO for under 50 steps on the evolving version of GSM8K, lifting evolving-intent accuracy from 64.0 to 76.0 with no single-turn loss.
The authors concede: the framework models only intent evolution, not finer user variation like persona, style, or typos, so synthesized dialogue is stylistically uniform; each turn carries a single intent transition, whereas real users may revise and switch in the same turn; the verifier is exact only at the final turn, with intermediate trajectories not directly scored. My own read: the SWE-Bench column showing Mistral Large 3 and GPT-5.1 at 0.0 is partly budget-exhaustion timeouts, and the paper does not fully separate intent-tracking failure from a single agent run simply crashing; the sample sizes (50 for SWE, 100 for BIRD) are small.