When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents

1 INTRODUCTION

• INTENTFLUX converts executable source tasks into controlled multi-turn edits while retaining original graders, making stale intent measurable through task success.Variants are replaced values; decoys are withdrawn constraints retained only when obeying them changes the source-task grader’s verdict.Source: §1 · pp. 1–2
• On 627 source tasks, mean task score falls from 0.476 (easy) to 0.384 (hard), a 19.3% relative reduction.Across eight models, evolving-dialogue performance is lower at every difficulty; a turn-matched control scores 0.734 versus 0.778 single-turn and 0.469 for drift.Source: §1 · p. 2
• On GENERAL-TEST, Deep Agents scores 0.354 and OpenHarness 0.392, compared with 0.367 for bare multi-turn execution; compression alone does not reliably recover performance.The comparison uses rolling summarization and two-phase compaction, respectively.Source: §1 · p. 2
• On GENERAL-TEST, STATEFORGE raises mean task score from 0.367 to 0.467 by maintaining an explicit active state before generation.The tracker removes superseded intent items and conclusions derived from them before supplying the active state to the base agent.Source: §1 · p. 2
• Ground-truth final-state input raises GENERAL-TEST score to 0.549, still below the clean single-turn score of 0.778.This indicates that state-estimation errors are material but explain only part of the remaining gap.Source: §1 · p. 2
• On-policy distillation improves a 2B tracker from 0.266 to 0.445, while thinking distillation improves a standalone 35B agent from 0.296 to 0.461.A 9B tracker is statistically indistinguishable from the 122B reference.Source: §1 · p. 2

Evolving intent and memory updates

• Evolving-intent and memory benchmarks study revisions, retractions, instruction retention, or current-versus-superseded values; INTENTFLUX adds heterogeneous executable tasks and grader-consequential decoys.The cited settings include EvolIF, InterruptBench, and Tack et al. (2026).Source: §2 · p. 2

Context management and state tracking

• Context methods compress or retrieve histories, whereas STATEFORGE folds edits into active requirements and tracks their dependencies in executable settings.The paper relates this to entity and belief tracking while targeting open-ended task requirements.Source: §2 · p. 2

Learning state maintenance

• The paper adapts rationale distillation and on-policy distillation to train either the tracker sub-role or a full agent on executable stale-intent failures.Source: §2 · p. 3

3.1 EVOLVING INTENT STATES

• A task’s intent state is a finite set of atomic natural-language goals, transformed by ADD and DELETE edits into the final active state G⋆.The oracle applies the edit sequence, and the source-task grader evaluates the candidate output against G⋆.Source: §3.1 · p. 3
• A VARIANT is a plausible value later replaced, while a DECOY is a plausible intent item later withdrawn; retained decoys change the source grader’s verdict if obeyed.Task success tests stale-state use, but aggregate failure alone does not attribute an error to a particular stale item.Source: §3.1 · p. 3

3.2 SOURCE TASKS AND CONSTRUCTION

• The General track spans code, math, SQL, data-to-text, summarization, and tool use; VitaBench supplies interactive delivery, in-store, OTA, and cross-domain tasks.The 627-task calibration pool, separate 502-task training pool, 125-task GENERAL-TEST set, and 100-task INTERACTIVE-TEST set serve distinct experiments.Source: §3.2 · p. 3
• Source problems are decomposed into ordered shards that populate target goals, with behavior-changing variants and plausible decoys annotated under semantic-dependency constraints.A downstream audit recovers the annotated final intent for 95% of GENERAL-TEST.Source: §3.2 · pp. 3–4

3 THE INTENTFLUX BENCHMARK

• IntentFlux turns verifiable source tasks into dialogues; the agent sees the dialogue, while an oracle derives the final active intent for the original grader.The source-task grader evaluates the final output or action.Source: §3 · p. 3, Figure 1 · p. 4

3.3 CONTROLLED STALE-INFORMATION LOAD

• Difficulty strata use construction budgets b ∈ {1, 3, 5}, jointly increasing superseded values, withdrawn items, and edit turns rather than isolating their separate effects.After the final edit, the simulator adds only a neutral request, with no explicit restatement of G⋆.Source: §3.3 · p. 4

4 EVALUATION PROTOCOL

• Each evaluation couples a user simulator, tested model, and native evaluator; the simulator interleaves feasible edits across active-goal and decoy queues.DeepSeek-V4-Flash serves as simulator and rubric judge when LLM judgment is needed, separate from the tested agent.Source: §4 · p. 4
• The intent-drift gap (IDG) is the case-paired clean–drift difference in full-credit alignment, capturing total loss associated with recovering intent from dialogue rather than isolating one cause.Clean and drift conditions share the source case, final intent, grader, and tested model; a turn-matched no-drift control tests whether extra turns alone account for the loss.Source: §4 · p. 5
• CAR counts only full task-native credit across the entire evaluation set, while mean score preserves partial credit on continuous-output tasks.Mean score averages task-native metrics and is used only for within-benchmark comparisons under fixed task composition.Source: §4 · p. 5

5 MEASURING INTENT DRIFT

• Increasing the joint variant/decoy budget lowers both mean score and CAR, but does not isolate the separate effects of variants and decoys.The easy, medium, and hard budgets are 1, 3, and 5; added edits also lengthen dialogues.Source: §5 · p. 5
• Every tested model has positive IDG at every difficulty stratum, and the loss generally grows with joint stale-information load.DeepSeek V4 Pro is the exception to the general easy-to-medium increase in the point estimate.Source: §5 · p. 5, Table 1 · p. 5

5.1 TASK-FAMILY RESPONSES AND A TURN-MATCHED CONTROL

• LiveCodeBench, GSM8K, and HumanEval show the largest declines, BFCL rises, and SummHay remains near a low-score floor.The equal-family macro average also falls from .497 to .433 to .385 across easy, medium, and hard.Source: §5.1 · pp. 5–6
• 0.265 separates the turn-matched drift score of .469 from the no-drift control score of .734, showing turn count alone does not explain the loss.The control removes revisions while matching dialogue length, so it does not isolate wording or edit-type effects.Source: §5.1 · p. 6

5.2 INTERACTIVE TRANSFER

• 0.103 lower rubric score in the no-restate interactive condition contrasts with clean; binary CAR remains unchanged across non-oracle conditions.Explicitly restating final intent in the diagnostic condition masks the rubric-score gap.Source: §5.2 · p. 6

6 FROM HISTORY COMPRESSION TO EXPLICIT STATE FOLDING

• Deep Agents and OpenHarness score 0.354 and 0.392, respectively, versus 0.367 for bare multi-turn execution on GENERAL-TEST.Compressed histories can retain both a superseded value and its replacement, leaving the generator to determine which remains active.Source: §6 · p. 6
• State folding removes superseded intent and conclusions dependent on it while preserving information that remains valid after an edit.Unlike history management, it asks which retained information is still valid.Source: §6 · p. 6

7 STATEFORGE: STATE FOLDING AND TRANSFER

• The tracker removes or replaces superseded intent items and invalidates conclusions derived from them before supplying the active state to generation.A separate format tracker and final gate handle output conventions.Source: §7 · pp. 6–7

7.1 STATE FOLDING REPAIRS INTENT DRIFT

• 0.100 mean-score gain (95% CI [0.030, 0.171]) for STATEFORGE over bare multi-turn execution on GENERAL-TEST.The scores are 0.467 for STATEFORGE and 0.367 for bare multi-turn; the ground-truth-state score is 0.549, still below the clean 0.778.Source: Table 2 · p. 7, §7.1 · p. 7

7.2 TRACKER CAPACITY AND TRANSFER

• A 9B tracker is statistically indistinguishable from the 122B reference, while OPD raises the 2B tracker from 0.266 to 0.445.The 122B tracker scores 0.428 under the same OPD evaluation protocol.Source: §7.2 · p. 8

7.3 FROM STATE TRACKING TO AGENT BEHAVIOR

• Thinking distillation raises a standalone 35B agent from 0.296 to 0.461 on GENERAL-TEST without an external harness.The result does not establish generalization to every update regime; the clean score decreases from .745 to .718.Source: §7.2 · p. 8, §7.3 · p. 8
• The tracker and full-agent paths test modular state maintenance and internalized state updates as distinct deployment choices.The OPD tracker stack and standalone thinking-distillation setting use different configurations and evaluation protocols.Source: §7.3 · p. 8, Figure 3 · p. 8

7.4 WHAT THE ABLATION ESTABLISHES

• Matched recap comparisons do not establish explicit-state representation superiority; their paired 95% confidence intervals include zero at n = 125.The oracle comparison shows positive paired difference from the full harness, indicating headroom from exact state estimation within the history-preserving interface.Source: §7.4 · p. 9

8.1 WHAT THE BENCHMARK MEASURES

• The calibration measures a joint stale-information stress response, not separate effects of variants, decoys, or dialogue length.Case-paired clean–drift gaps remain positive across models and strata, while sensitivity varies by task family.Source: §8.1 · p. 9
• Trace inspection finds omitted deleted goals and, in BFCL, REPLACE failures that bind a new value to the wrong function-call slot.These observations motivate tracking intent-item identity and dependencies rather than relying only on retrieval over earlier utterances.Source: §8.1 · p. 9

8.2 SIMULATOR AND EVALUATOR VALIDITY

• In an audit of 100 sampled simulator utterances, fidelity is 96% and coherence is 97%.The audit indicates invented and contradictory intent changes are uncommon in the sampled dialogues.Source: §8.2 · p. 9
• On VitaBench, providing the complete final intent in closing turns removes the measured partial-credit drift gap.The main interactive condition does not restate the final intent, and the restating condition is a separate diagnostic rather than a difficulty factor.Source: §8.2 · p. 9

8.3 LIMITATIONS

• INTENTFLUX targets executable or decomposable intent, and some comparisons and transfer-path scaling remain limited or unresolved.The external-harness comparison and INTERACTIVE-TEST each use one tested model; the component-matched ablation does not isolate the explicit state representation as the sole source of system-level gain.Source: §8.3 · p. 9

• INTENTFLUX makes obsolete intent grader-consequential and measures whether agents follow current intent after earlier information is superseded or withdrawn.Its construction yields a controlled difficulty response.Source: §9 · p. 10
• History-management harnesses leave substantial errors, while STATEFORGE improves performance by maintaining an explicit active state.Source: §9 · p. 10
• OPD further improves a small tracker, while thinking distillation begins to internalize state-folding behavior into the agent.Source: §9 · p. 10

A BENCHMARK CONSTRUCTION AND EVALUATION DETAILS

• Tables 3, 4, 10, and 11 identify the source-set and paper-strata, task-native-grader, turn-matched-control, and VitaBench interactive-transfer comparisons.The table captions specify their scope but do not provide result values.Source: Table 3 · p. 12, Table 4 · p. 12, Table 10 · p. 14
• Source problems are decomposed into ordered atomic shards, with behavior-changing variants, grader-consequential decoys, dependency checks, and human inspection.A downstream GENERAL-TEST audit recovers the annotated G⋆ for 95% of cases.Source: A BENCHMARK CONSTRUCTION AND… · p. 12

B DETAILED BENCHMARK RESULTS

• Tables 5–9 cover joint variant/decoy calibration, paired clean–drift results, per-family response, cross-model validation, and task-family full-credit IDG.The calibration budgets jointly raise available alternatives and withdrawn decoys, while mean dialogue length increases from 13.2 to 37.1 turns.Source: Table 5 · p. 13, Table 6 · p. 13, Table 7 · p. 13

C CONTROLS AND INTERACTIVE TRANSFER

• The turn-matched no-drift control states final requirements first and uses no-change confirmations, with identical turn counts and comparable message length to drift.It is closer to single-turn performance on most families, but does not isolate message order, edit wording, variants, and decoys separately.Source: C CONTROLS AND INTERACTIVE TRANSFER · p. 14
• The aligned no-restate condition scores lower than clean, while default closing restatement masks that difference; all three non-oracle conditions have CAR .190.This result is specific to partial-credit rubric evaluation in the single-model setup.Source: C CONTROLS AND INTERACTIVE TRANSFER · p. 14

D STATEFORGE DIAGNOSTICS AND TRAINING DETAILS

• Tables 12 and 13 cover component ablation and tracker-scale comparisons under the unified re-run protocol.Source: Table 12 · p. 15, Table 13 · p. 15
• The oracle row shows material headroom from exact state estimation within the history-preserving harness, but does not establish that tracking dominates the clean single-turn gap.The matched recap and full-harness comparisons are not separated at this sample size.Source: D STATEFORGE DIAGNOSTICS AND TRAINING… · p. 15
• The modular-adaptation scores pair a frozen 122B base agent with a 2B tracker, and the before–after change is attributable to OPD training within that stack.These full-episode scores use the tracker-training evaluation protocol rather than the GENERAL-TEST task-score protocol.Source: D STATEFORGE DIAGNOSTICS AND TRAINING… · p. 15
• The 35B thinking-distilled v4 model scores .461 on the drift set and .444 under avg@3 decoding, versus .296 for the base model.On the clean condition, the distilled model scores .718 versus .745 for the base model.Source: D STATEFORGE DIAGNOSTICS AND TRAINING… · p. 15

E QUALITY-CONTROL RUBRICS

• Across 100 sampled edit-phase utterances, simulator fidelity is 96% and coherence is 97%.The fidelity rubric checks scheduled edits and active requirements; the coherence rubric checks reference resolution, active-state consistency, and dialogue flow.Source: E QUALITY-CONTROL RUBRICS · p. 15