Source-linked AI summary
Mind the Sim2Real Gap in User Simulation for Agentic Tasks
Xuhui Zhou, Weiwei Sun, Qianou Ma, Yiqing Xie, Jiarui Liu, Weihua Du, Sean Welleck, Yiming Yang, Graham Neubig, Sherry Tongshuang Wu, Maarten Sap
TL;DR
LLM user simulators are widely used in interactive evaluation, but their fidelity to real human behavior and judgment is insufficiently verified. The paper formalizes this Sim2Real gap, introduces USI, and compares 31 simulators with 451 people across 165 τ-bench tasks. Simulators diverge behaviorally, inflate evaluation quality, and fail to reproduce the full signal captured by human feedback.
Problem
Interactive benchmarks rely on LLMs both to generate user turns and evaluate agents, despite limited evidence that these simulations faithfully reflect real users.
Method
The paper introduces a taxonomy and User-Sim Index, then compares 31 LLM simulators with 451 real participants on the same 165 τ-bench tasks.
Results
The best simulator scores 76.0 USI versus 92.7 for humans; simulators are more cooperative and positive, while binary rewards miss human-perceived quality.
Takeaways & Limitations
Simulator-based evaluations should be checked against human interactions and report behavior, outcomes, and user ratings separately rather than assuming replacement.
Takeaways & Limitations
Findings may not generalize beyond airline and retail tasks, the mostly English-speaking US participant population, role-playing setup, or the single-agent study.
Abstract
from arXiv · showhide
As NLP evaluation shifts from static benchmarks to multi-turn interactive settings, LLM-based simulators have become widely used as user proxies, serving two roles: generating user turns and providing evaluation signals. Yet, these simulations are frequently assumed to be faithful to real human behaviors, often without rigorous verification. We formalize the Sim2Real gap in user simulation and present the first study running the full $τ$-bench protocol with real humans (451 participants, 165 tasks), benchmarking 31 LLM simulators across proprietary, open-source, and specialized families using the User-Sim Index (USI), a metric we introduce to quantify how well LLM simulators resemble real user interactive behaviors and feedback. Behaviorally, LLM simulators are excessively cooperative, stylistically uniform, and lack realistic frustration or ambiguity, creating an "easy mode" that inflates agent success rates above the human baseline. In evaluations, real humans provide nuanced judgments across eight quality dimensions while simulated users produce uniformly more positive feedback; rule-based rewards are failing to capture rich feedback signals generated by human users. Overall, higher general model capability does not necessarily yield more faithful user simulation. These findings highlight the importance of human validation when using LLM-based user simulators in the agent development cycle and motivate improved models for user simulation.
1 Introduction
The paper examines whether LLM user simulators faithfully reproduce human interaction and evaluation behavior, and whether rule-based rewards capture human judgments. It formalizes these gaps and studies them on τ-bench using real participants and 31 simulators.
- The study asks whether simulated users behave like people, provide comparable quality judgments, and can be replaced by rule-based rewards.These questions address simulator optimization, benchmark validity, and the richness of human feedback.
- The authors introduce a taxonomy spanning four behavioral dimensions and two evaluation dimensions, aggregated into the 0–100 User-Sim Index.USI is intended to audit simulator fidelity before deployment.
- 451 real participants completed 165 τ-bench tasks, enabling comparison with 31 proprietary, open-source, and specialized LLM simulators.The study replaces τ-bench’s simulated user with humans while benchmarking simulator families.
- 76.0 was the best simulator USI, versus 92.7 for humans, indicating a substantial Sim2Real gap.The result is reported as the headline alignment comparison.
- LLM simulators are more uniform, cooperative, and information-complete, while evaluators overrate quality and binary rewards miss nuanced human judgments.GPT-5.1 overestimates human-likeness by 55% and overall score by 18% of the rating scale.
2 Related Work
Prior work has used LLMs in interactive user simulation and questioned their fidelity, but systematic assessment against real users remains limited. The paper positions its contribution as directly measuring this unresolved Sim2Real gap.
- LLM user simulators now appear in interactive benchmarks for training and evaluating conversational agents.Earlier approaches included rule-based and statistical dialogue-policy models.
- No prior benchmark had assessed how faithfully its LLM simulations reflect real users across the full interactive evaluation loop.This motivates measuring both simulator behavior and evaluation alignment.
- Existing work has examined isolated aspects of human-behavior fidelity, including persona consistency, dialogue quality, and task-oriented conversation.The cited studies do not provide a comprehensive validation of simulated users against real users.
- Prior findings also suggest that stronger assistants do not necessarily produce better user simulators.The paper reports this counterintuitive pattern as related evidence about simulator capability.
3 Framework
The framework separates behavioral and evaluative Sim2Real gaps, measures alignment across defined dimensions, and aggregates them into USI while retaining component scores. It combines distributional behavior metrics, outcome calibration, and feedback agreement.
- Taxonomy: The taxonomy covers four behavioral dimensions—communication style, information pattern, clarification behavior, and error reaction—and two evaluation dimensions, ECE and Eval.These dimensions represent the two roles played by LLM simulators: interaction and assessment.
- Behavioral metrics: Behavioral alignment uses Sørensen–Dice coefficients over lexical, structural, and behavioral features extracted from simulator and human interactions.The comparison is between population-level simulator distributions and aggregate human distributions.
- Behavioral metrics: Dice scores range from 0 to 100, with 100 indicating perfect human alignment, and dimension scores average constituent metrics.A zero-zero comparison is assigned a perfect score.
- Outcome calibration: ECE compares simulator and human success across five equal-width task-difficulty bins, with lower values indicating better outcome calibration.Task difficulty is fixed using mean simulator success across tasks, independent of the scored simulator.
- Evaluative alignment: Eval is defined as (1 − MAE) × 100, measuring agreement between simulated feedback and human judgments across survey dimensions.The survey includes task success and interaction-quality ratings.
- User-Sim Index: USI aggregates D1–D4, (1−ECE) × 100, and Eval into a composite alignment score, but its components are reported separately.The framework warns that a simulator may align in one role while diverging in another.
4 Experimental Setup: τ-bench
The experiment instantiates the framework on τ-bench, comparing human and LLM users under a controlled agent-and-task setup. It collects interaction traces, success judgments, quality ratings, and simulator alignment scores across broad model categories.
- τ-bench: τ-bench jointly studies an LLM user simulator, a tool-augmented customer-service agent, and a binary automatic reward.Its airline and retail tasks specify customer goals, constraints, databases, and applicable policies.
- Human study: Human annotators role-played customers on the same 165 tasks with the same agent used in τ-bench evaluation.Their interactions provide human-agent traces for behavioral comparison.
- Human study: Post-task surveys collected five-way task-success judgments and six interaction-quality ratings, including efficiency, human-likeness, flow, and reuse.Ratings were recorded on short ordinal scales.
- Human study: Three independent annotation batches support human–human agreement estimates and test stability across annotator pools.The design prioritizes task-distribution breadth with three annotators per task.
- Models: 31 simulators span 18 proprietary, 9 open-source, and 4 specialized models, all interacting with the same agent on the same tasks.The specialized models were fine-tuned for user simulation or human-like behavior, yet did not close the gap.
- Aggregation: USI combines four behavioral scores, outcome calibration, and evaluative alignment, with mean ± std reported across three human batches.Models without survey data use the five-component version excluding Eval.
5 RQ1: The Behavioral Gap
LLM user simulators diverge from real humans across all four behavioral dimensions, creating an easier evaluation setting that can inflate apparent agent competence.
- LLM simulators diverge significantly from humans across all four behavioral dimensions.
- Communication style (D1): 1.0% of GPT-4o turns are short versus 29.0% for humans, while 49.0% are polite versus 15.3% for humans.These differences reflect excessive verbosity and uniform politeness relative to real customers.
- Information pattern (D2): UserLM-8b includes nearly twice as many identifier-like tokens per turn as humans, indicating more front-loaded information.
- Clarification behavior (D3): GPT-4o expresses uncertainty in 14.6% of turns versus 7.3% for humans, whereas UserLM-8b shows 3.0% uncertainty and 10.6% certainty markers versus 1.0% for humans.Models therefore miscalibrate clarification behavior in opposite directions.
- Error reaction (D4): GPT-4o and CoSER pivot more than humans, at 19.1% and 16.5% versus 8.4%, while using less accusatory language.Because rates cover all user turns, this dimension indicates an interaction-level association rather than a causal response to verified errors.
- Top simulators yield agent success rates up to 77.8%, exceeding the human baseline of 63.6%.Volunteered information, cooperation through errors, and absent frustration remove conversational challenges such as ambiguity and incomplete information.
6 RQ2 & RQ3: The Evaluation Gap
LLM-based evaluators and binary rewards disagree with human judgments in different ways: evaluators miscalibrate quality assessments, while rewards fail to capture human-perceived quality.
- GPT-5.1 rates human-likeness 1.11 points higher and reuse intent 0.83 points higher than humans, but task completion 0.15 points lower.This reflects leniency on interaction quality alongside conservatism on task completion.
- LLM evaluators inflate perceived interaction quality while underestimating actual task success, preventing reliable substitution for human assessment.
- 70.6% of reward=0 interactions are judged successful by humans, while 33% of reward=1 interactions are judged unsuccessful or only partially successful.Binary rewards also cannot distinguish full completion, partial completion, or an appropriate policy refusal.
- Reward=0 and reward=1 distributions are nearly indistinguishable for overall score, efficiency, and human-like qualities.The disconnect extends across all eight quality dimensions and indicates that binary reward is orthogonal to human-perceived quality.
- Overall, neither LLM evaluation nor automatic binary reward adequately captures human-perceived interaction quality.
7 Discussion and Limitations
The discussion recommends validating simulator behavior and conclusions against real users while preserving simulation’s speed and repeatability. It also identifies boundaries from prompting, task/domain coverage, participant diversity, and study design.
- Prompting: Prompting moved some behaviors toward humans but reduced USI from 70.9 to 64.6 and worsened ECE from 0.18 to 0.29.The experiment suggests that improving some behavioral features can move others further from human behavior, though it does not show prompting cannot close the gap.
- Recommendations: The study recommends reporting simulated behavior, task outcomes, and user ratings separately, then checking whether rankings and conclusions persist on a smaller human task sample.This preserves broad simulator comparisons while testing representative easier and harder tasks across domains with real users.
- Interpreting simulator results: Simulator agreement can raise confidence that results are not tied to one model or prompt, but cannot replace human checks for claims about real users.Simulation-based success demonstrates compatibility with a simulator’s choices about information sharing, error reactions, and success criteria, not necessarily with people.
- Scope and generalizability: Broad gaps appeared across airline and retail tasks and 31 models, but findings may not extend to coding, web, research, clinical, or open-ended social interactions.Participants were mostly English-speaking US role-players, while real users vary in language, background, expertise, patience, and willingness to correct systems.
- Measurement and study limitations: Results are constrained by approximate word matching, a single agent model, role-playing participants, three conversations per task, and possible τ-bench issues.The study notes that D4 can capture reactions unrelated to agent errors and does not establish how often that measurement is correct.
8 Conclusion
The study compares LLM simulators with people on the same interactive tasks and finds gaps in both user behavior and evaluation. It concludes that simulation remains useful for scalable experiments, but claims about real users require human verification.
- Study conclusion: 31 LLM simulators and 451 people were compared on the same 165 τ-bench tasks with the task and agent held fixed.The comparison covers both the interactor and evaluator roles played by simulators.
- User behavior: As interactors, LLMs are more uniform and cooperative, provide more information upfront, ask for clarification differently, and push back less when interactions go wrong.These behavioral differences can change how often the agent succeeds.
- Evaluation: As evaluators, examined LLMs often give more favorable ratings than people, while τ-bench’s automatic binary reward does not capture the full range of human judgments.The study therefore identifies gaps in both simulated feedback and automatic evaluation.
- Implication: Simulators remain valuable for fast, repeatable experiments and broad comparisons, but strong simulator performance does not establish equally strong performance with people.Claims about real users need separate human verification.
A.1 Annotator Demographics, Annotation Interface, and Annotation Quality Control
The appendix describes the human annotation sample, interaction interface, quality-control judge, and survey criteria used to collect and filter interaction traces.
- Annotation interface: The web interface displayed task instructions and role-playing guidelines beside a chat panel, and participants ended interactions with /stop before completing a post-task survey.Participants sent one message at a time while role-playing the specified user.
- Quality control: The LLM judge was calibrated against 51 author-labeled interactions before filtering the final dataset.The calibration set supplied a human ground truth for quality-control decisions.
- Quality control: The LLM judge reached κ=0.6 agreement with human labels and 0.94 precision, but showed conservative bias with low false positives and high false negatives.Among interactions accepted by the judge, 94% were confirmed as passing by humans.
- Annotation criteria: The survey asked annotators to score instruction following, natural role-play, avoidance of unsupported facts, conversation fidelity, and authentic human-like responses on a 0–100 scale.The quality-control prompt also required responsible and thoughtful survey responses.
A.2 Model details
The study evaluates 31 LLM user simulators across proprietary, open-source, and specialized categories, using 165 τ-bench tasks and three independent batches per model.
- Evaluation setup: Each of 31 simulators was evaluated on all 165 τ-bench tasks in three independent batches, with mean batch scores used for final model comparisons.The evaluation covers the complete task set for every model.
- Proprietary models: The proprietary category contains 18 models from the GPT, Claude, and Gemini families, accessed through their respective APIs with default parameters.The listed models span multiple generations and model sizes within these families.
- Open-source models: The open-source category contains nine models, including DeepSeek, Llama, Qwen, GPT-oss, MiniMax, and Kimi systems.The group includes models ranging from 7B to 235B-scale systems.
- Specialized models: The specialized category contains four models fine-tuned for user simulation or human-like behavior: CoSER-8B, UserLM-8B, HumanLike-7B, and HumanLM-opinion.These models are distinguished by their simulation- or human-behavior-specific training focus.
A.3 Survey (τ-bench)
The survey defines task success, efficiency, clarification experience, human-likeness, interaction smoothness, overall performance, reuse intention, and open-ended behavioral feedback, alongside transparent behavioral metrics. Its validation evidence shows that persona prompting can improve some measures while worsening others, and that simulators differ from humans in information sharing, uncertainty, and error reactions.
- Survey measures: The survey asks users to judge task completion, efficiency, clarification quantity and effort, human-likeness, interaction smoothness, overall performance, and reuse intention.It also collects specific examples of successful behavior, errors, rule violations, forgotten information, and problematic messages.
- Operational metrics: Behavioral metrics quantify communication style, information patterns, clarification behavior, and error-reaction-like language using turn-level lexical and regex indicators.Metrics include verbosity, short-turn frequency, politeness, front-loading, uncertainty, pushback, clarification, information-seeking, emotion, accusation, and pivoting.
- Persona-prompting ablation: Persona prompting moves some metrics toward humans but overshoots others, lowering overall fidelity rather than consistently closing the behavioral gap.The prompt improves politeness and front-loading, but worsens pivoting, leaves emotion unchanged, collapses D4, and degrades calibration.