pith. sign in

arxiv: 2606.20708 · v1 · pith:SRKDPVZLnew · submitted 2026-06-16 · 💻 cs.AI · cs.CL· cs.HC

Simulated Customers Never Walk Away: Decision Fidelity of LLM User Simulators Measured Against Real Purchase Outcomes

Liang Chen This is my paper

Pith reviewed 2026-06-27 01:36 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.HC
keywords LLM simulationdecision fidelityuser simulatorsconversational agentssales automationdisengagementfidelity evaluationpurchase outcomes
0
0 comments X

The pith

LLM customer simulators match real buyers but push non-buyers toward purchase.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Existing tests of LLM user simulators check only whether they sound like humans in assigned roles, but leave untested whether they decide like humans whose motivation is real and can fade. The paper defines decision fidelity as the match between simulated and real decision-state dynamics, measured on 2,790 actual sales conversations that include verified payment outcomes. Simulators reproduce the states of eventual buyers almost exactly yet shift eventual non-buyers markedly closer to buying, halving expressed resistance and doubling deliberation while never fabricating a purchase. The same pattern appears across different model families and survives explicit instructions that the simulator may disengage. Because real non-buyers say no and stop while simulated ones keep asking about price, any benchmark or training run that uses these simulators will overstate agent performance exactly on the customers who matter most.

Core claim

Simulators exhibit a disengagement deficit: they reproduce the decision states of eventual buyers with near-zero bias but inflate those of eventual non-buyers toward the purchase frame by a depth bias of +0.40, halving resistance from 25.1 percent to 13.5 percent and nearly doubling deliberation from 21.9 percent to 40.1 percent while producing no purchases. The deficit replicates across model families and is not removed by prompting the simulator that it may disengage.

What carries the argument

The teacher-forced probe protocol that holds conversation context and measurement instrument fixed while comparing simulated decision states to verified real purchase outcomes.

If this is right

  • Benchmarks of sales or persuasion agents will overstate funnel progress on the subset of customers who would actually walk away.
  • Training loops that rely on these simulators will reinforce the same over-engagement pattern.
  • Simple prompting fixes that allow disengagement reduce marginal bias but leave the outcome-conditioned contrast largely intact.
  • The deficit appears consistently across multiple model families, suggesting it is not an isolated implementation flaw.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same bias may appear in any simulation setting where real users can quit, such as customer support or negotiation dialogues.
  • Explicitly modeling decaying motivation rather than relying on role instructions could reduce the deficit.
  • Agents trained or evaluated on current simulators may under-deliver when deployed against real users who retain the option to disengage.

Load-bearing premise

The fixed-context probe protocol and the set of 2,790 sampled real conversations capture the simulators' own decision dynamics without introducing selection effects or measurement artifacts.

What would settle it

Re-running the identical teacher-forced probes on a new collection of real sales conversations from a different product category or customer segment and finding that the depth bias for non-buyers is no longer near 0.40.

Figures

Figures reproduced from arXiv: 2606.20708 by Liang Chen.

Figure 1
Figure 1. Figure 1: The teacher-forced decision-fidelity protocol. At each probe, the real user and the simulator continue the identical real prefix ht, and both continuations are scored by the same instrument Φ; the paired depth bias D is aggregated per conversation and contrasted across real-outcome strata. Context-dependent and judge-dependent biases cancel by construction (Prop. 1); only outcome-correlated behavioral diff… view at source ↗
Figure 2
Figure 2. Figure 2: The disengagement deficit. (a) Conversation-level mean depth bias by real-outcome stratum with bootstrap 95% CIs. Both persona simulators track eventual buyers (≈0) while inflating eventual non-buyers (≈ +0.4). Explicit disengagement instructions (right) repair the non-buyer stratum but push buyers negative—the error is relocated, not removed, and the stratum gap (the deficit) persists. (b) The distributio… view at source ↗
read the original abstract

LLM-as-user-simulation has become core infrastructure for conversational AI: agent benchmarks (tau-bench), training pipelines, and a growing body of fidelity studies all rely on LLMs role-playing the human side of dialogue. Existing frameworks measure communicative fidelity -- whether simulators talk like humans -- against ground truth from paid participants role-playing assigned goals. We argue this has a structural blind spot: when the goal is assigned, the user's willingness is exogenous, so no framework can test whether simulators make decisions like real users whose motivation is endogenous, latent, and decaying. We introduce decision fidelity -- whether a simulated population reproduces the decision-state dynamics of real users facing real, consequential choices -- and measure it on a unique testbed: 2,790 production conversations between an LLM sales agent and real customers, including 793 with verified payment outcomes. Using a teacher-forced probe protocol that holds context and instrument fixed, we find a systematic, outcome-correlated failure we call the disengagement deficit: simulators reproduce eventual buyers almost exactly (depth bias +0.09) but inflate eventual non-buyers toward the purchase frame (depth bias +0.40; d=0.38, p<0.001), halving expressed resistance (25.1% to 13.5%) and nearly doubling deliberation (21.9% to 40.1%) while fabricating no purchases. The deficit replicates across model families (DeepSeek: d=0.41, p=0.002) and resists the obvious fix: instructing the simulator that it may disengage cuts marginal bias five-fold but barely moves the outcome-conditioned contrast (d=0.34, p=0.008). Real non-buyers say "not now" and stop; simulated non-buyers ask about price. Evaluating or training sales and persuasion agents against such simulators overstates funnel progress exactly where it matters most -- the customers who walk away.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces 'decision fidelity' as a measure for LLM user simulators, arguing that existing communicative fidelity tests cannot capture endogenous decision dynamics because they assign goals exogenously. Using 2,790 real production sales conversations (793 with verified payment outcomes) and a teacher-forced probe protocol that holds context and instrument fixed, the authors report a systematic 'disengagement deficit': simulators reproduce eventual buyers closely (depth bias +0.09) but inflate eventual non-buyers toward purchase (depth bias +0.40; d=0.38, p<0.001), halving expressed resistance and nearly doubling deliberation while fabricating no purchases. The deficit replicates across model families and persists after an explicit disengagement instruction.

Significance. If the result holds, it identifies a load-bearing limitation in LLM simulators for sales-agent benchmarks, training pipelines, and persuasion studies: they systematically overstate funnel progress precisely for the customers who walk away. The empirical grounding in verified real-world outcomes, statistical contrasts, and cross-model replication constitutes a concrete advance over prior fidelity work that relies on paid role-players with assigned goals.

major comments (2)
  1. [Methods / probe protocol] §4 (teacher-forced probe protocol): the central contrast depends on the assumption that fixing context and instrument isolates endogenous decision-state dynamics. The skeptic concern that this protocol may itself suppress natural termination in simulators (real non-buyers stop after resistance; simulators continue) is not fully dispelled by the disengagement-instruction ablation, which reduces marginal bias but leaves the outcome-conditioned d=0.34 contrast intact. A direct comparison of probe vs. free-generation trajectories on the same real contexts would test whether the measured deficit is protocol-dependent.
  2. [Data and sampling] §3.2 (sampling of 793 verified-payment subset from 2,790 conversations): selection into the verified non-buyer sample could skew toward higher-engagement cases, inflating the apparent deficit. The paper should report engagement statistics (turns, resistance rate) for the full 2,790 vs. the verified subset to rule out selection effects as a contributor to the +0.40 depth bias.
minor comments (2)
  1. [Results] The depth-bias metric definition and its exact formula should be stated explicitly in the main text (not only supplementary) to permit independent replication of the reported effect sizes.
  2. [Results] Table or figure presenting the resistance/deliberation percentages (25.1% vs 13.5%; 21.9% vs 40.1%) should include per-model breakdowns to show consistency with the aggregate d=0.38 result.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation and constructive suggestions. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Methods / probe protocol] §4 (teacher-forced probe protocol): the central contrast depends on the assumption that fixing context and instrument isolates endogenous decision-state dynamics. The skeptic concern that this protocol may itself suppress natural termination in simulators (real non-buyers stop after resistance; simulators continue) is not fully dispelled by the disengagement-instruction ablation, which reduces marginal bias but leaves the outcome-conditioned d=0.34 contrast intact. A direct comparison of probe vs. free-generation trajectories on the same real contexts would test whether the measured deficit is protocol-dependent.

    Authors: The disengagement-instruction ablation was included specifically to test whether the probe protocol suppresses termination. Even with explicit permission to disengage, the outcome-conditioned contrast remains significant (d=0.34). This supports that the deficit reflects endogenous decision dynamics rather than protocol artifact. A free-generation comparison would introduce uncontrolled variation in context and instrument, weakening the controlled contrast central to the design. We have added discussion of this limitation in the revised manuscript. revision: partial

  2. Referee: [Data and sampling] §3.2 (sampling of 793 verified-payment subset from 2,790 conversations): selection into the verified non-buyer sample could skew toward higher-engagement cases, inflating the apparent deficit. The paper should report engagement statistics (turns, resistance rate) for the full 2,790 vs. the verified subset to rule out selection effects as a contributor to the +0.40 depth bias.

    Authors: We agree that selection effects warrant explicit checking. In the revised manuscript we will report engagement statistics (average turns and resistance rates) for the full 2,790 conversations versus the 793 verified subset. revision: yes

Circularity Check

0 steps flagged

No significant circularity: central contrast uses external real purchase outcomes

full rationale

The paper's core claim compares LLM simulator behavior to verified real customer outcomes from 2,790 production conversations (793 with payment verification). No load-bearing step reduces to a fitted parameter, self-citation chain, or definitional equivalence; the disengagement deficit is quantified directly against independent ground truth rather than derived from the simulators themselves. The teacher-forced protocol is an experimental design choice, not a self-referential derivation. This is a standard empirical measurement against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on empirical comparison to external real-world outcomes with standard statistical reporting; no free parameters, invented entities, or non-standard axioms are required beyond routine assumptions of the t-test and effect-size calculations.

axioms (1)
  • standard math Assumptions underlying two-sample t-tests and Cohen's d effect sizes (sufficient sample size and approximate normality for the reported p and d values)
    Invoked when reporting d=0.38, p<0.001 on the depth bias contrast.

pith-pipeline@v0.9.1-grok · 5884 in / 1386 out tokens · 39980 ms · 2026-06-27T01:36:42.007138+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 1 canonical work pages

  1. [1]

    ResPer: Computationally modelling resisting strategies in persuasive conversations

    Ritam Dutt, Sayan Sinha, Rishabh Joshi, Surya Shekhar Chakraborty, Meredith Riggs, Xinru Yan, Haogang Bao, and Carolyn Penstein Rosé. ResPer: Computationally modelling resisting strategies in persuasive conversations. InProceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 78–

  2. [2]

    doi: 10.18653/v1/2021.eacl-main.7

    Association for Computational Linguistics, apr 2021. doi: 10.18653/v1/2021.eacl-main.7. URLhttps://aclanthology.org/2021.eacl-main.7

    work page doi:10.18653/v1/2021.eacl-main.7 2021

  3. [3]

    Using imperfect surrogates for downstream inference: Design-based supervised learning for social science applications of large language models.NeurIPS, 2023

    Naoki Egami, Musashi Hinck, Brandon M Stewart, and Hanying Wei. Using imperfect surrogates for downstream inference: Design-based supervised learning for social science applications of large language models.NeurIPS, 2023

    2023

  4. [4]

    A survey on llm-as-a-judge.arXiv preprint arXiv:2411.15594, 2024

    Jiawei Gu et al. A survey on llm-as-a-judge.arXiv preprint arXiv:2411.15594, 2024. 13

    Pith/arXiv arXiv 2024

  5. [5]

    User willingness-aware sales talk dataset

    Asahi Hentona, Jun Baba, Shiki Sato, and Reina Akama. User willingness-aware sales talk dataset. InProceedings of COLING 2025, 2025. arXiv:2412.19490

    arXiv 2025

  6. [6]

    This human study did not involve human subjects: Validating LLM simulations as behavioral evidence.arXiv preprint arXiv:2602.15785, 2026

    Jessica Hullman, David Broska, Huaman Sun, and Aaron Shaw. This human study did not involve human subjects: Validating LLM simulations as behavioral evidence.arXiv preprint arXiv:2602.15785, 2026

    arXiv 2026

  7. [7]

    A user simulator for task-completion dialogues.arXiv preprint arXiv:1612.05688, 2016

    Xiujun Li, Zachary C Lipton, Bhuwan Dhingra, Lihong Li, Jianfeng Gao, and Yun-Nung Chen. A user simulator for task-completion dialogues.arXiv preprint arXiv:1612.05688, 2016

    Pith/arXiv arXiv 2016

  8. [8]

    Learning when to quit in sales conversa- tions.arXiv preprint arXiv:2511.01181, 2025

    Emaad Manzoor, Eva Ascarza, and Oded Netzer. Learning when to quit in sales conversa- tions.arXiv preprint arXiv:2511.01181, 2025

    arXiv 2025

  9. [9]

    ConvApparel: A benchmark dataset and validation framework for user simulators in conversational recommenders.arXiv preprint arXiv:2602.16938, 2026

    Ofer Meshi, Krisztian Balog, Sally Goldman, Avi Caciularu, Guy Tennenholtz, Jihwan Jeong, Amir Globerson, and Craig Boutilier. ConvApparel: A benchmark dataset and validation framework for user simulators in conversational recommenders.arXiv preprint arXiv:2602.16938, 2026. Google

    arXiv 2026

  10. [10]

    Salesrlagent: A reinforcement learning approach for real-time sales conversion prediction and optimization.arXiv preprint arXiv:2503.23303, 2025

    M Nandakishor. Salesrlagent: A reinforcement learning approach for real-time sales conversion prediction and optimization.arXiv preprint arXiv:2503.23303, 2025

    arXiv 2025

  11. [11]

    How much does persuasion strategy matter? LLM-annotated evidence from charitable donation dialogues

    Tatiana Petrova, Stanislav Sokol, and Radu State. How much does persuasion strategy matter? LLM-annotated evidence from charitable donation dialogues. 2026

    2026

  12. [12]

    Agenda- based user simulation for bootstrapping a POMDP dialogue system

    Jost Schatzmann, Blaise Thomson, Karl Weilhammer, Hui Ye, and Steve Young. Agenda- based user simulation for bootstrapping a POMDP dialogue system. InNAACL-HLT, 2007

    2007

  13. [13]

    Ferdinand M. Schessl. The autocorrelation blind spot: Why 42% of turn-level findings in LLM conversation analysis may be spurious.arXiv preprint arXiv:2604.14414, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Lost in simulation: LLM-simulated users are unreliable proxies for human users in agentic evaluations.arXiv preprint arXiv:2601.17087, 2026

    Preethi Seshadri, Samuel Cahyawijaya, Ayomide Odumakinde, Sameer Singh, and Seraphina Goldfarb-Tarrant. Lost in simulation: LLM-simulated users are unreliable proxies for human users in agentic evaluations.arXiv preprint arXiv:2601.17087, 2026

    arXiv 2026

  15. [15]

    Towards understanding sycophancy in language models.arXiv preprint arXiv:2310.13548, 2023

    Mrinank Sharma, Meg Tong, Tomasz Korbak, et al. Towards understanding sycophancy in language models.arXiv preprint arXiv:2310.13548, 2023

    Pith/arXiv arXiv 2023

  16. [16]

    Evaluating alignment of behavioral dispositions in LLMs.arXiv preprint arXiv:2602.11328,

    Amir Taubenfeld, Zorik Gekhman, Lior Nezry, Omri Feldman, Natalie Harris, Shashir Reddy, Romina Stella, Ariel Goldstein, Marian Croak, Yossi Matias, and Amir Feder. Evaluating alignment of behavioral dispositions in LLMs.arXiv preprint arXiv:2602.11328,

    arXiv

  17. [17]

    Persuasion for good: Towards a personalized persuasive dialogue system for social good

    Xuewei Wang, Weiyan Shi, Richard Kim, Yoojung Oh, Sijia Yang, Jingwen Zhang, and Zhou Yu. Persuasion for good: Towards a personalized persuasive dialogue system for social good. InACL, 2019

    2019

  18. [18]

    Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan.τ-bench: A benchmark for tool-agent-user interaction in real-world domains.arXiv preprint arXiv:2406.12045, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Ai-salesman: Towards reliable large language model driven telemarketing.arXiv preprint arXiv:2511.12133, 2025

    Qingyu Zhang, Chunlei Xin, Xuanang Chen, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun, Qing Ye, Qianlong Xie, and Xingxing Wang. Ai-salesman: Towards reliable large language model driven telemarketing.arXiv preprint arXiv:2511.12133, 2025. 14

    arXiv 2025

  20. [20]

    Judging llm-as-a-judge with mt-bench and chatbot arena.NeurIPS, 2023

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena.NeurIPS, 2023

    2023

  21. [21]

    Mind the Sim2Real gap in user simulation for agentic tasks.arXiv preprint arXiv:2603.11245, 2026

    Xuhui Zhou, Weiwei Sun, Qianou Ma, Yiqing Xie, Jiarui Liu, Weihua Du, Sean Welleck, Yiming Yang, Graham Neubig, Sherry Tongshuang Wu, and Maarten Sap. Mind the Sim2Real gap in user simulation for agentic tasks.arXiv preprint arXiv:2603.11245, 2026

    arXiv 2026

  22. [22]

    RealUserSim: Bridging the reality gap in agent benchmarking via grounded user simulation.arXiv preprint arXiv:2605.20204, 2026

    Ming Zhu, Juntao Tan, Rithesh Murthy, Jielin Qiu, Liangwei Yang, Wenting Zhao, Silvio Savarese, Shelby Heinecke, and Huan Wang. RealUserSim: Bridging the reality gap in agent benchmarking via grounded user simulation.arXiv preprint arXiv:2605.20204, 2026. 15 A Decision-State Instrument Prompt (Φ) The instrument is an LLM perceiver that maps a user turn, i...

    Pith/arXiv arXiv 2026