PaperlayerView exact revision on arXiv

Source-linked AI summary

TRACER: Trajectory-Aligned Learning for Multi-Turn User Simulation

Geng Chen, Ruotong Pan, Zhirui Yang, Qiqi He, Jiawei Chen, Zhang Yunfei, Chongyuan Chen, Minxuan Lv, Zheng Yang, Win-Bin Huang, Xiangyu Wu, Wenwu Ou

arXiv:2609.28690v2cs.AI

TL;DR

Current simulators can produce plausible responses without reproducing real users’ intent evolution and outcomes across interactions. TRACER trains on real dialogues and aligns simulated trajectories with human ones through multi-turn reinforcement learning. It outperforms baselines on reported behavioral measures, while DM-Bench shows that response quality need not track conversion.

  • Problem

    Existing simulators often align linguistic style or individual responses without capturing real users’ decision-making and behavior across complete interactions.

    Source: §1 · p. 1

  • Method

    TRACER combines supervised fine-tuning on real dialogues with multi-turn reinforcement learning using outcome and trajectory rewards and deviation-aware turn-level updates.

    Source: §1 · p. 2

  • Results

    TRACER-7B exceeds the strongest baseline by 11.4 F1 points and outperforms all baselines on ACC, F1, Group-∆PCR, and OT-DTW; DM-Bench finds higher response quality does not necessarily mean higher conversion.

    Source: §1 · p. 2, §5.2 · p. 7

  • Takeaways & Limitations

    Evaluating conversational marketing agents on response quality and conversion together reveals that the two outcomes do not necessarily align.

    Source: §1 · p. 2

    Evaluation focuses on multi-turn dialogue; applying the method elsewhere requires suitable trajectory representations and deviation measures, and whether reference-based deviations help remains untested.

    Source: D LIMITATION · p. 20

Abstract

from arXiv · show

Faithful user simulation is fundamental to building, evaluating, and improving interactive AI at scale. Yet current simulators often produce plausible individual responses without reproducing the intent evolution and outcomes observed in real interactions. We propose TRACER, a multi-turn user simulator that models evolving user intent and aligns simulated trajectories with real ones. TRACER is trained in two stages: supervised fine-tuning on real user dialogues, followed by multi-turn reinforcement learning. The RL stage combines hierarchical outcome- and trajectory-level rewards with deviation-aware advantage modulation, jointly addressing reward sparsity and credit assignment challenges in long dialogues. On real customer-service sessions organized into reference cohorts, TRACER-7B surpasses the strongest baseline by 11.4 conversion F1 points, while outperforming all baselines on group-level conversion-rate error and semantic trajectory distance and generalizing to out-of-distribution scenarios. In human Turing tests, annotators identified TRACER conversations at near-chance accuracy. Building on this simulator, we further introduce the Dynamic Marketing Benchmark, which jointly evaluates persuasion and response quality via simulated interactions, revealing that higher response quality does not necessarily correspond to higher conversion rates.

1 INTRODUCTION

TRACER targets simulators that sound plausible but fail to reproduce real users’ decisions across multi-turn interactions. It combines trajectory-aligned training with cohort-based evaluation and introduces a benchmark that assesses response quality alongside conversion.

1 INTRODUCTION

  • Existing simulators align mainly with linguistic style and optimize single turns, leaving decision-making and trajectory-level credit assignment under-modeled.The paper links these limitations to unstable feedback and reduced utility for high-fidelity evaluation and training.Source: §1 · p. 1
  • TRACER uses real dialogues, inferred intent states, supervised fine-tuning, and multi-turn reinforcement learning with outcome and trajectory rewards plus deviation-aware turn updates.The turn-level updates use deviation magnitude and persistence while preserving the session-level signal.Source: §1 · p. 2
  • Evaluation groups 3,866 customer-service sessions into 762 reference groups, assessing session behavior, dialogue structure, semantic consistency, generalization, and perceived human-likeness.Annotators identified TRACER conversations at near-chance accuracy in a Turing study.Source: §1 · p. 2
  • DM-Bench evaluates conversational marketing agents on both response quality and conversion, since higher response quality does not necessarily correspond to higher conversion rates.The benchmark uses interactions with simulated users.Source: §1 · p. 2

2 RELATED WORK

Prior user simulators span persona prompting, state modeling, and multi-reward optimization, but their focus remains short of aligning complete rollouts with human trajectories. Simulators also support dialogue training and evaluation, where their reliability constrains the credibility of simulator-based evaluation.

2 RELATED WORK

  • Earlier LLM simulators use personas or static intent predictions, while newer approaches add reasoning, state modeling, or multi-reward reinforcement learning.The cited methods generally optimize individual responses or next actions rather than complete human-referenced rollouts.Source: §2 · p. 2
  • User simulators supply scalable interactive training data and feedback, and enable diverse interaction assessments beyond static benchmarks.Their reliability remains a bottleneck for the credibility of simulator-based evaluation.Source: §2 · p. 2

3 PROBLEM FORMULATION

The paper formulates multi-turn user simulation as generating annotated user behavior in response to a profile and evolving assistant dialogue. Its objective is to match human trajectory distributions, including terminal outcomes and how interactions unfold.

3 PROBLEM FORMULATION

  • At each turn, the simulator conditions on user information and dialogue context, generating a rationale, an ordinal intent score, and a user response.The assistant then responds to the evolving dialogue; sessions end at conversion, dropout, or the turn limit.Source: §3 · p. 3
  • Intent annotations pair each response with a rationale and inferred intent score, which can rise, fall, or remain unchanged across the interaction.These annotations act as an interpretable proxy for expressed engagement and commitment.Source: §3 · p. 3
  • The simulator aims to capture human trajectory distributions across terminal outcomes, intent evolution, and linguistic and temporal interaction characteristics.Evaluation aggregates trajectories from users with comparable pre-interaction contexts, initializing each rollout from its own user’s context.Source: §3 · p. 3

4 TRACER

TRACER uses supervised fine-tuning followed by interactive reinforcement learning, connecting session-level trajectory alignment to turn-level updates. Its rewards combine terminal outcomes, trajectory fit, and valid output structure, while deviation-aware modulation adjusts reinforcement across turns.

4 TRACER

  • TRACER addresses the gap between endpoint matching and coherent interaction by combining two-stage training with shared trajectory alignment for session rewards and turn-level updates.The overall alignment cost scores a completed interaction, while local discrepancies inform per-turn updates.Source: §4 · p. 3

4.1 SIMULATOR INITIALIZATION AND INTERACTIVE ROLLOUT

  • Supervised fine-tuning on real dialogues initializes user-like language generation and structured intent prediction before rollouts with a fixed assistant.Each rollout begins from the corresponding user information and initial assistant utterance, and ends at conversion, dropout, or the turn limit.Source: §4.1 · pp. 3–4, Figure 1 · p. 4

4.2 SESSION-LEVEL REWARDS

  • Session rewards combine terminal-intent agreement with DTW trajectory alignment, which preserves temporal order while accommodating different conversational pacing.Terminal agreement alone cannot constrain how a session unfolds; similar final intent can accompany substantially different trajectories.Source: §4.2 · p. 4
  • The session reward also accounts for length mismatch, discounts nearly constant intent trajectories, and checks prescribed output structure and field consistency.The resulting scalar incorporates output validity, terminal agreement, and intermediate trajectory information.Source: §4.2 · p. 5

4.3 DEVIATION-AWARE POLICY OPTIMIZATION

  • TRACER reuses the optimal DTW alignment to calculate each generated turn’s average deviation across its aligned reference turns, then aggregates deviations over a discounted trajectory suffix.The discount controls how strongly later discrepancies affect the score; smaller values emphasize nearby turns.Source: §4.3 · p. 5
  • Deviation-aware modulation adjusts turn-level reinforcement or suppression according to cumulative deviation while preserving the sign of the trajectory advantage.Larger cumulative deviations weaken positive reinforcement and increase negative suppression; the modulation is bounded and can be disabled.Source: §4.3 · p. 5

5 EXPERIMENTS

The experiments evaluate TRACER against real customer-service sessions across outcome, group-fidelity, and semantic measures, then test human-likeness, component contributions, and generalization. TRACER-7B improves on the strongest baseline in conversion F1 and reaches near-chance human identification accuracy.

5.1 EXPERIMENTAL SETUP

  • The evaluation uses 3,866 test sessions in 762 reference groups and compares simulated trajectories with real dialogues across session behavior, group fidelity, and semantic content.Simulations interact with a fixed assistant until conversion, dropout, or a turn limit; groups contain sessions with similar user conditions.Source: §5.1 · p. 6

5.2 MAIN RESULTS

  • TRACER-7B outperforms all baselines on ACC, F1, Group-∆PCR, and OT-DTW, exceeding the strongest baseline by 11.4 F1 points.Its PCR is 50.5%; W1-Turns ranks second behind HUMANLM.Source: §5.2 · p. 7, Table 1 · p. 7

5.3.1 HUMAN EVALUATIONS

  • TRACER-7B reaches 49.50% human-identification accuracy, near the 50% random-guess threshold, compared with 61.80% for SFT and over 98% for general LLMs.The Turing-style test asks annotators to identify which of two dialogues came from a human.Source: §5.3.1 · p. 7, Figure 2 · p. 7

5.3.2 ABLATION STUDY

  • The full model has the best mean across five direction-specified metrics; removing key components or replacing DTW with strict turn matching degrades performance.Removing RL entirely causes the largest overall degradation.Source: §5.3.2 · p. 8, Table 2 · p. 8

5.3.3 OUT-OF-DOMAIN GENERALIZATION

  • On unseen financial customer-service dialogues, TRACER leads the strongest baseline on terminal-state ACC, F1, and dialogue embedding similarity by 8.4, 4.8, and 0.7 points, respectively.The evaluation samples 1,000 dialogues from CSC-Conv.Source: §5.3.3 · p. 8, Figure 3 · p. 8

6 DYNAMIC MARKETING BENCHMARK

DM-Bench evaluates conversational marketing models using simulated interactions under shared user and merchant conditions. Its results show that response quality alone does not reliably indicate conversion effectiveness.

6 DYNAMIC MARKETING BENCHMARK

  • The benchmark builds on TRACER’s simulation of intent changes and terminal decisions to compare response quality with simulated conversion under shared conditions.The paper reports that quality-only benchmarks can miss gaps between good-sounding responses and decision-moving ones.Source: §6 · p. 8, §6.2 · p. 9

6.1 BENCHMARK CONSTRUCTION

  • DM-Bench contains 734 instances, with seven assistant models interacting with TRACER under the same initial user and merchant conditions; it measures simulated conversion rate and response quality.Response quality assesses accuracy, helpfulness, informativeness, and empathy.Source: §6.1 · p. 9

6.2 RESULTS AND FINDINGS

  • Across seven models, response quality and conversion rate have only moderate rank correlation (Spearman ρ = 0.57, p = 0.18, n = 7), and the highest-RQ model is surpassed in conversion by two others.Top models also fail to convert many users despite high response-quality scores.Source: §6.2 · p. 9

7 CONCLUSION

The conclusion presents TRACER as a multi-turn simulator aligned with real users in language and decision-making, and DM-Bench as an evaluation of response quality and persuasion. The authors report near-human fidelity and a mismatch between response quality and persuasion effectiveness.

7 CONCLUSION

  • TRACER combines linguistic and decision-making alignment through supervised fine-tuning and trajectory-aligned reinforcement learning.Source: §7 · p. 10
  • The authors report that TRACER outperforms existing simulators and achieves near-human fidelity in Turing tests.Source: §7 · p. 10
  • DM-Bench jointly measures persuasion effectiveness and response quality, and the reported results show these can be misaligned in mainstream LLMs.Source: §7 · p. 10

AI USE STATEMENT

The authors report using LLMs only as auxiliary tools for linguistic refinement and technical support. They state that LLMs did not contribute to the research’s conceptualization, design, or scientific interpretation.

AI USE STATEMENT

  • LLMs were used for grammar and spelling refinement and for technical support in drafting and debugging experimental code.Source: AI USE STATEMENT · p. 10
  • The authors state that LLMs played no role in conceptualizing the research problem, developing its theoretical framework, or designing experiments.Source: AI USE STATEMENT · p. 10
  • The authors report auditing, revising, and verifying all LLM suggestions, while completing the core academic contributions independently.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

The study limits its data use to authorized resources and describes de-identification and content safeguards. The appendix also specifies training, reward, and evaluation configurations.

ETHICS STATEMENT

  • All data were authorized for scientific analysis and benchmark construction, de-identified, cleaned, and audited for sensitive or harmful content.The authors state that identifying details were deleted or replaced with semantic placeholders.Source: ETHICS STATEMENT · p. 10

A.1 IMPLEMENTATION DETAILS

  • Training used Qwen2.5-7B-Instruct and Qwen3-4B-Instruct, with SFT followed by multi-turn RL; experiments ran on 8 NVIDIA H200 GPUs.SFT used two epochs and a learning rate of 1.0 × 10−5; multi-turn RL used six epochs and up to 15 dialogue turns per rollout.Source: §A.1 · p. 14
  • Group-∆PCR equally weights reference groups, while W1-Turns compares within-group user-turn distributions and OT-DTW matches semantic trajectories within groups.OT-DTW uses normalized reply embeddings and DTW costs; lower values indicate better agreement for the three group-based metrics.Source: §A.1 · p. 14
  • Table 3 reports PCR, ACC, and F1 in percent, with mean ± standard deviation across five evaluation seeds for each fixed checkpoint.The reported deviations are across evaluation seeds, not independent training runs; the reference PCR is 51.24%.Source: Table 3 · p. 15
  • Session-level rewards were propagated to valid user turns, with reward weights (λ1, λ2, λ3) = (0.2, 0.5, 0.3) and advantage modulation strength λadv = 0.2.The coefficients were selected on a held-out validation split.Source: §A.1 · p. 15

A.2 SENSITIVITY ANALYSIS OF HYPERPARAMETERS IN MULTI-TURN ADVANTAGE MODULATION

The appendix examines hyperparameter sensitivity and training dynamics, then details the data, profile, annotation, and cleaning procedures used to construct the simulator’s training and evaluation sets.

A.2 SENSITIVITY ANALYSIS OF HYPERPARAMETERS IN MULTI-TURN ADVANTAGE MODULATION

  • All tested advantage-modulation settings lower mean Group-∆PCR than the unmodulated baseline, but ACC and F1 improvements are not uniform.Across settings, ACC ranges from 78.36 to 79.80, F1 from 78.53 to 79.60, and PCR from 47.21% to 53.86%; standard deviations cover evaluation seeds, not independent training runs.Source: §A.2 · p. 15
  • ρ = 0.4 gives the best W1-Turns and OT-DTW among tested discount values, while λadv = 0.1 gives the highest ACC and F1 and lowest Group-∆PCR.The smaller λadv increases both trajectory distances relative to default; changing γ from 0.5 to 0.4 yields similar results, but does not establish robustness over a broader range.Source: §A.2 · pp. 15–16
  • The default configuration achieves the lowest OT-DTW among tested configurations while improving all five directional metrics over the unmodulated baseline.The default is (λadv, γ, ρ) = (0.2, 0.5, 0.4).Source: §A.2 · pp. 15–16

A.3 TRAINING DYNAMICS AND CONVERGENCE ANALYSIS

  • For both TRACER-7B and TRACER-4B, reward rises rapidly early in RL and then stabilizes, while entropy and response length decline gradually.The response-length trend suggests reward improvement is not achieved by increasingly long outputs; Figure 5 plots reward, entropy, and mean response length for TRACER-7B.Source: §A.3 · p. 16, Figure 5 · p. 16

B.1 DATA SOURCE

  • The source corpus comprises 36,841 real merchant–user dialogue sessions across domains including healthcare, automotive services, education consulting, and legal consulting.The data were anonymized and de-identified before entering the research pipeline.Source: §B.1 · pp. 16–17

B.2 USER PROFILE CONSTRUCTION

  • Profiles combine templated objective demographic, occupation, and consumption attributes with LLM-inferred subjective priors such as personality and initial core intent.The subjective-prior extraction masks final conversion outcomes to avoid label leakage.Source: §B.2 · pp. 17–18

B.3 DIALOGUE AUGMENTATION AND STRUCTURED TRAJECTORY CONSTRUCTION

  • Dialogue augmentation adds a context-inferred rationale proxy and a discrete 0–5 intent level to each user turn, enabling joint modeling of intent states and replies.The rationale is not treated as the user's actual private mental state; level 5 requires concrete contact information in the current reply.Source: §B.3 · p. 18

B.4 DATA CLEANING

  • Cleaning removes unusable sessions, retains 3–15 user turns, and yields 11,906 sessions split into disjoint 8,040-train and 3,866-test sets.Test sessions form 762 reference groups; the 1:1 balance applies to group-majority labels, while session-level conversion is 51.24%.Source: §B.4 · p. 19

C EXTENDED ABLATION

The extended TRACER-4B ablation reports broadly consistent component trends with the 7B model. The limitations section identifies open-ended social dialogue, transfer beyond dialogue, and potential misuse as boundaries or concerns.

C EXTENDED ABLATION

  • Removing advantage modulation, trajectory reward, or multi-turn RL generally worsens mean ACC, F1, Group-∆PCR, and W1-Turns; removing RL causes the largest deterioration.Table 5 reports TRACER-4B ablation results as mean ± standard deviation over five evaluation seeds.Source: C EXTENDED ABLATION · p. 19, Table 5 · p. 20
  • Without DTW alignment, ACC drops more at 4B than at 7B, F1 drops 1.8 points at both scales, and W1-Turns is affected more at 7B.Multi-turn RL improves behavioral accuracy and group-level fidelity at both scales.Source: C EXTENDED ABLATION · p. 20

D LIMITATION

  • Open-ended social dialogue remains untested; absent explicit task objectives and clear success criteria, its reward modeling and credit assignment are more challenging.The authors leave systematic study of this setting to future work.Source: D LIMITATION · p. 20
  • Evaluation is limited to multi-turn dialogue; applying advantage modulation to agentic settings requires suitable trajectory representations and deviation measures for heterogeneous action spaces.Whether reference-based deviation signals improve optimization in those settings remains an empirical question.Source: D LIMITATION · p. 20
  • A high-fidelity simulator aligned with real users could be misused to optimize manipulative or deceptive persuasion strategies.Source: D LIMITATION · p. 20

E.1 REASONS LLM-SIMULATED USERS ARE JUDGED NON-HUMAN

The appendix analyzes why simulated users are judged non-human across linguistic style, interaction rhythm, and persona and logic. It also specifies an unseen financial customer-service evaluation and response-quality judging setup.

E.1 REASONS LLM-SIMULATED USERS ARE JUDGED NON-HUMAN

  • 3.0% of TRACER-7B samples are flagged for lacking human imperfections, compared with 93.8%–98.8% for Gemini and Doubao; overly neat structures appear in 17.0% of TRACER-7B outputs.Overly formal or written register is 4.0% for TRACER-7B; the table’s lower percentages indicate better performance.Source: §E.1 · pp. 20–21, Table 6 · p. 21
  • TRACER-7B has 0.8% information overload and 4.4% mechanical response patterns, versus 78.2%–87.2% information overload and up to 82.4% mechanical patterns for Gemini and Doubao.Excessive cooperativeness is 27.2%, and super-human compliance occurs in 0.6% of cases.Source: §E.1 · p. 21
  • TRACER-7B shows 1.4% out-of-character behavior, zero emotional discontinuity, and 0.2% each for lack of situation sense and over-politeness.Gemini and Doubao have 38.4%–46.2% persona inconsistency, and over-politeness reaches 14.0%.Source: §E.1 · p. 21

F OUT-OF-DOMAIN GENERALIZATION EVALUATION DETAILS

  • The appendix describes CSC-Conv as an open-source financial customer-service corpus unseen during training for out-of-domain dialogue evaluation.The remaining dialogues are sampled for evaluation after ambiguous issue-resolution cases are filtered; the sample contains 1,000 sessions with a 689:311 resolved-to-unresolved ratio.Source: F OUT-OF-DOMAIN GENERALIZATION… · pp. 21–22

G.1 RESPONSE QUALITY EVALUATION

  • Response quality in multi-turn dialogues is judged using the prompt from Zhu et al. (2026), with gpt-oss-120B chosen to mitigate potential self-preference bias.The prompt is illustrated in Figure 7.Source: §G.1 · p. 22

G.2 HUMAN-REFERENCED STRATEGY-RESPONSE EVALUATION

This section presents human-referenced strategy-response agreement as descriptive evidence from sampled interactions and documents the associated human evaluations. It also explains DTW as an order-preserving, flexible alignment method for comparing intent trajectories despite differences in turn pacing.

G.2 HUMAN-REFERENCED STRATEGY-RESPONSE EVALUATION

  • Strategy-direction recalls are descriptive agreement with annotator expectations, not causal effects, real-user outcomes, per-assistant performance, or evidence of generalization.The pooled row covers 449 strategy instances from 200 conversations; urgency has only six instances, and the study claims no statistically significant differences between strategies.Source: §G.2 · p. 22

G.3 RESULTS

  • The DM-Bench results section reports evaluations of seven LLMs and describes human role-play across 100 shared contexts, totaling 700 conversations.The provided passages identify conversion rate, conversion turns, and response-quality measures, but do not provide comparative values or findings.Source: §G.3 · pp. 22–23, Table 8 · p. 23, Table 10 · p. 23

H HUMAN ANNOTATION

  • The paper describes annotation procedures, annotator qualifications, remuneration, and quality control for its human evaluations.The annotation team comprised 10 trained annotators, each with about one year of experience, and a dedicated reviewer conducted random sampling and consistency checks.Source: H HUMAN ANNOTATION · p. 23, §H.1 · pp. 23–24

H.1 ANNOTATOR DETAILS

  • Ten trained annotators with at least associate degrees were paid 50 RMB per hour, while a reviewer checked annotation quality and consistency.Annotators had approximately one year of AI-data annotation experience.Source: §H.1 · pp. 23–24

H.2 TURING-STYLE EVALUATION

  • Annotators compared real and simulated dialogue excerpts using profile-aware judgments of naturalness, information style, emotion, consistency, situational awareness, and imperfections.They also selected applicable non-human reasons after judging which dialogue was more likely human, with guidance cautioning against relying on grammar, length, or isolated utterances.Source: §H.2 · p. 24

H.3 HUMAN ANNOTATION FOR LLM-INFERRED USER INTENTS AND RATIONALE

  • A blind human evaluation of 300 stratified conversations assesses turn-level user-intent scores and the plausibility of LLM-generated rationales.Intent scores range from 0 for an ended conversation to 5 for contact information actually provided; rationale ratings consider dialogue history and user profile.Source: §H.3 · pp. 26–27

I.1 MOTIVATION

  • Because plausible simulated and real intent trajectories can progress at different local speeds, strict turn-by-turn matching may incorrectly penalize them.DTW is adopted to measure global trajectory consistency while allowing local stretching or compression across turns.Source: §I.1 · p. 27

I.2 GENERAL FORMULATION OF DYNAMIC TIME WARPING

  • Dynamic Time Warping aligns sequences through an order-preserving warping path that permits local stretching and compression, with dynamic-programming complexity O(NM).Unlike pointwise comparison, it supports sequences with different lengths, progression speeds, or local phase shifts.Source: §I.2 · p. 27, Figure 8 · p. 28
  • Because DTW seeks a globally optimal nonlinear correspondence, the sequences need not have equal lengths or align their elements by matching indices.The path elements pair positions across the two sequences, while the local discrepancy function supplies the cost for each pairing.Source: §I.2 · pp. 28–29
  • DTW aligns sequence elements along a path constrained to start and end at both sequences’ boundaries, preserve temporal order, and use local steps that permit one-to-many, many-to-one, and one-to-one matches.Together, these constraints allow flexible nonlinear alignment without violating either sequence’s internal order.Source: §I.2 · pp. 28–29
  • DTW computes each cumulative prefix-alignment cost as the local cost plus the minimum of the three neighboring cumulative costs.The cumulative matrix is initialized at the first element and along its first row and column; its minimum valid-path cost gives the alignment cost.Source: §I.2 · p. 29

I.3 DTW-BASED INTENT TRAJECTORY ALIGNMENT IN OUR METHOD

This section describes DTW-based alignment of simulated and real intent trajectories to derive trajectory-level rewards and turn-level advantage modulation, along with the annotation, dialogue-cleaning, and prompting details that guide intent-conditioned user behavior and replies.

I.3 DTW-BASED INTENT TRAJECTORY ALIGNMENT IN OUR METHOD

  • The trajectory reward combines DTW path cost, a turn-count penalty, normalization by real trajectory length, and a flattening discount for nearly constant simulated intents.The flattening discount applies when real intent varies substantially but the simulated trajectory stays at one or nearly one level for a long period.Source: §I.3 · p. 30
  • Turn-level deviations average intent differences across all real turns aligned to each simulated turn, and discounted downstream deviations modulate each turn’s advantage.Larger future deviation weakens positive reinforcement and strengthens negative updates, linking persistent drift to turn-level credit assignment.Source: §I.3 · p. 31
  • DTW permits local pacing variation: an extra simulated turn can be treated as a delay, while a one-turn intent transition can align with two real turns.This avoids treating every local turn shift as an error while preserving temporal order.Source: §I.3 · p. 32
  • DTW alignment yields both a trajectory-level plausibility signal and turn-level deviations, encouraging the simulator to match real users’ overall intent evolution rather than fixed turn positions.The cumulative alignment cost informs the trajectory reward, while local deviations inform future-deviation scores and turn-level advantage modulation.Source: §I.3 · pp. 30–32

J CASE STUDY

  • The case study describes a decisive, concise prospective car buyer seeking down-payment information, available models, and a test drive, followed by a baseline dialogue.The profile identifies an 18–23-year-old user in Chaozhou with a high-school education or below, a partner, no children, and a general-worker occupation.Source: J CASE STUDY · p. 34

TRACER-7B (Ours)

  • TRACER-7B examples include users stating car-buying needs, budgets, and test-drive requests; the accompanying failure analysis lists lack of flaws, mechanical cooperation, and out-of-character behavior.The examples include requests for consultant contact and model recommendations.Source: TRACER-7B (Ours) · p. 34

K PROMPT DETAILS

  • The prompt-details section instructs persona extraction to use a concise behavioral label and core intent to report the user’s initial objective and key constraints, excluding ending outcomes and contact details.It specifies a JSON output with persona_summary and task_background fields.Source: K PROMPT DETAILS · p. 35
  • The annotation prompt assigns each user turn a rationale, intent level, and behavioral choice, then outputs results in chronological order as JSON.The fields also include the user's inner monologue and evaluator-facing score rationale.Source: K PROMPT DETAILS · p. 36
  • Intent levels run from 0 for exit or churn to 5 for explicit contact information, with mandatory endpoint assignments for those behaviors.The rubric distinguishes low, passive, active, and strong-intent responses; providing a contact request without concrete information does not itself count as conversion.Source: K PROMPT DETAILS · pp. 36, 39
  • The cleaning prompt rejects sessions with meaningless or irrelevant replies, accidental clicks, or agent contact information in turn one, but annotation inconsistency alone does not invalidate dialogue.When derived annotations conflict with observable behavior, retain the dialogue and correct the annotations.Source: K PROMPT DETAILS · p. 37
  • The merchant-agent prompt prioritizes answering from relevant knowledge, merchant information, and dialogue policy, then guiding conversion and asking follow-up questions.It also calls for fluent, friendly, human-like, concise replies and prohibits fabricated information or definite promises.Source: K PROMPT DETAILS · p. 38
  • The user-simulation prompt maps intent level 5 to leaving contact information and requires explicit contact details in the reply.The provided examples of explicit contact details include a phone number or WeChat ID.Source: K PROMPT DETAILS · p. 39
  • At intent level 0, the behavioral choice must be Exit; at levels 1–4, it must be Continue the conversation.Source: K PROMPT DETAILS · p. 39
  • Replies should use colloquial language consistent with the user profile and adjust their detail and warmth to the evaluated intent level.Source: K PROMPT DETAILS · p. 39
Loading 2609.28690v2…