Lost in Simulation: LLM-Simulated Users are Unreliable Proxies for Human Users in Agentic Evaluations

Abstract: Agentic benchmarks increasingly rely on LLM-simulated users to scalably evaluate agent performance, yet the robustness, validity, and fairness of this approach remain unexamined. Through a user study with participants across the United States, India, Kenya, and Nigeria, we investigate whether LLM-simulated users serve as reliable proxies for real human users in evaluating agents on τ-Bench retail tasks. We find that user simulation lacks robustness, with agent success rates varying up to 9 percentage points across different user LLMs. Furthermore, evaluations using simulated users exhibit systematic miscalibration, underestimating agent performance on challenging tasks and overestimating it on moderately difficult ones. African American Vernacular English (AAVE) speakers experience consistently worse success rates and calibration errors than Standard American English (SAE) speakers, with disparities compounding significantly with age. We also find simulated users to be a differentially effective proxy for different populations, performing worst for AAVE and Indian English speakers. Additionally, simulated users introduce conversational artifacts and surface different failure patterns than human users. These findings demonstrate that current evaluation practices risk misrepresenting agent capabilities across diverse user populations and may obscure real-world deployment challenges.

• The paper demonstrates that LLM-simulated users fail to reliably emulate human interactions in agentic benchmarks.
• It employs a stratified study comparing agent performance using success rates and calibration errors across diverse demographics and simulation models.
• Findings reveal notable behavioral deviations, group-specific biases, and miscalibration risks that could jeopardize real-world agent deployment.

Evaluating the Reliability of LLM-Simulated Users in Agentic Benchmarks

Introduction

The increasing deployment of LLM-driven agents in real-world, multi-turn scenarios such as retail assistance and travel management necessitates effective evaluation protocols to measure agent robustness and user experience. Their evaluation has trended toward replacing human users in benchmarking pipelines with LLM-simulated user models for scalability. The paper "Lost in Simulation: LLM-Simulated Users are Unreliable Proxies for Human Users in Agentic Evaluations" (2601.17087) addresses critical concerns about the reliability, validity, and fairness of such simulation-based protocols, presenting thorough empirical evidence that challenges the fidelity of current simulated user paradigms.

Methodology

The study adapts the $\tau$-Bench [yao2025taubench]—a widely used agentic benchmark centered on retail customer service scenarios—to systematically contrast agent performance with LLM-simulated users and real human users. Human users were recruited from the United States (with explicit stratification by dialect: Standard American English and African American Vernacular English, and by age ranges), India, Kenya, and Nigeria, ensuring coverage across major English-speaking populations and demographic slices relevant for fairness analyses.

Agents (GPT-4o) interacted with both simulated and human users to complete tasks stratified by difficulty. Success rates and a calibration metric, Expected Calibration Error (ECE), were used to quantify the accuracy with which simulated user outcomes reflect real human performance across demographic groups and task complexities.

Behavioral Differences between Simulated and Human Users

The investigation reveals pronounced conversational artifacts in simulated user interactions, notably increased question-asking and exaggerated politeness, diverging from the more direct communication patterns observed in human users.

Figure 1: Conversational snippets highlight increased question-asking and politeness in simulated user interactions versus more concise responses from human participants.

This behavioral gap exposes the risk of agentic evaluations training models to expect and optimize for non-representative user behaviors, potentially diminishing real-world performance.

Robustness of Simulated User Evaluation

Success rates of agentic benchmarks vary substantially with the choice of user simulation LLM. Performance fluctuates by up to 9 percentage points across user models, even when the agent LLM is held constant. For example, using Sonnet 4.5 as the user model yields a 75.9% success rate versus 67.8% for GPT-4o and 67.0% for Sonnet 3.7, exposing the brittleness and lack of robustness in simulation-based evaluation pipelines. This sensitivity highlights the necessity for multi-model reporting and increased scrutiny in simulation infrastructure design.

Validity: Calibration to Human Users

Calibration analyses reveal substantial misalignment between agent performance with simulated and human users. With US participants, agents achieve a 45.2% task success rate, while calibration error ($ECE_\text{Human--LLM}$) reaches 15.1, indicating systematic miscalibration.

Figure 2: Success rate and calibration error between simulated and real users show simulated users underestimate agent performance on hard tasks and overestimate on moderate ones for US participants.

Most notably, simulated users systematically underestimate agent performance on the most challenging tasks (human user success rates: 30.8% vs. much lower simulated rates) and overestimate performance on moderately difficult tasks (human: 39%).

Fairness: Demographic Biases and Group-Specific Calibration

Performance and calibration gaps compound for marginalized groups, particularly African American Vernacular English (AAVE) speakers. Agent success rates for AAVE participants drop to 39.4% (compared to 50.6% for SAE speakers), and calibration error inflates to 20.3 versus 11.7 for SAE. These disparities intensify with age, with agent success for AAVE, 55+ registering nearly 19 percentage points below the SAE, 55+ baseline.

Figure 3: Success rate and miscalibration persist for SAE, 18-34, illustrating moderate but non-trivial misalignment even for the best-calibrated groups.

Figure 4: Success rate trajectory for United States SAE speakers reveals age-associated variance, with calibration error fluctuating across age groups.

Cross-country analysis (India, Kenya, Nigeria) identifies further group-specific reliability gaps, with simulated users poorly calibrated to Indian and AAVE participants ($ECE_\text{Human--LLM}$ 18.9 for both), while Nigerian and Kenyan users exhibit moderate calibration gaps. These findings contradict any assumption of homogeneity in simulation-to-human transfer performance, revealing LLM-based user simulation to be more reliable for specific Anglo-centric dialects and routinely misaligned for others.

Error Type and Attribution Analysis

Failure mode analyses indicate that simulated users prompt more agent-driven failures (agent responsible: 48.9% vs. human condition: 24.5%), whereas human conversations are predominantly user-driven (user responsible: 62.2% vs. simulation: 40%). Simulated user conversations surface frequent argument errors and output omissions not observed in human interactions, inadvertently obfuscating authentic deployment challenges where user ambiguity, misunderstanding, and partial compliance are prevalent.

Implications for Agentic Benchmark Design

The strong empirical evidence for robustness, validity, and fairness failures in LLM-based user simulation necessitates a reevaluation of agentic benchmarking protocols. The results indicate that optimizing agents wholly or predominantly for simulated users may obscure systematic weaknesses, demographic disparities, and real-world deployment risk. The observed behavioral artifacts and group-specific calibration failures underscore the unilateral nature of simulated benchmarks and highlight the need for multi-model simulation, direct validation against diverse real user data, and behavioral interventions to correct simulation artifacts.

Prospects for Future Research

The research implies several avenues for advancement:

• Multilingual and Cross-cultural Extension: Given the restriction to English, extending analyses to multilingual, culturally variable settings is critical for global deployment.
• Agent Diversity: The impact of agent model performance and interaction style on calibration gaps warrants systematic investigation.
• Domain Transferability: Exploring how simulation reliability generalizes to domains such as healthcare and education, with more complex, longer-form tasks.
• Behavioral and Demographic Prompting: The mixed efficacy of behavioral constraints and demographic prompting merits further work to develop robust, controllable simulation pipelines.

Conclusion

This paper provides comprehensive evidence that LLM-simulated users are insufficiently robust, valid, or fair proxies for real human users in agentic benchmarks. Simulated user-based evaluation pipelines risk misrepresenting agent capabilities and exacerbating deployment inequities, particularly for marginalized communities and underrepresented dialects. For agentic benchmark designers, the findings underscore the urgency of integrating direct human validation, behavioral audit mechanisms, and demographic representativity—concrete steps critical for closing the simulation–reality gap in interactive AI evaluation.

Figure 5: User study chat interface exemplifies the real-world multi-turn environment used for benchmarking agent performance with diverse human populations.