Source-linked AI summary
Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States
Zixuan Wang, Yufan Zhou, Jinzhou Tang, Xinle Yu, Chengjun Wu, Lyumanshan Ye, Zhaoxiang Feng, Letian Peng, Adyasha Patra, Fan Bai, Enze Ma, Zhengding Hu, Jianyang Gu, Zhao Wang, Yufei Ding, Jingbo Shang, Tianmin Shu, Zhiting Hu, Zhen Wang
TL;DR
Human-aware assistants need supervision grounded in users’ unspoken beliefs and goals, but such evidence is difficult to obtain at scale. Mind2Dialogue simulates evolving user states, gives an Oracle privileged access to them, and distills the resulting responses into students trained without state inputs. Across model families, the full corpus improves personalization metrics, with gains extending to belief and action reasoning on Qwen and Llama.
Problem
Human-aware training lacks scalable responses grounded in users’ unspoken beliefs, goals, and circumstances because those states are not directly observable.
Method
Mind2Dialogue uses a psychology-guided shared-state simulator and privileged distillation to train students on Oracle responses while withholding evolving states from their inputs.
Results
Training on the full M2D-CORPUS improves every reported personalization metric across Qwen, Llama, and OLMo, with PrefEval generation gains of 26.6 to 40.9 points and additional belief-action gains on Qwen and Llama.
Takeaways & Limitations
The results support simulated user states as supervision for personalized assistance and, on some model families, transfer to theory-of-mind reasoning.
Takeaways & Limitations
The simulator’s stable and transient state fields are simulator-defined control variables rather than measurements of real users’ mental states.
Abstract
from arXiv · showhide
As language models become more capable, long-term collaboration in learning, reasoning, and decision-making calls for a deeper understanding of the people they serve. Yet training such human-aware language models faces a fundamental supervision gap because current datasets for LLM assistant training contain few if any well-informed responses explicitly grounded in users' unspoken beliefs and goals. Scaling such supervision is inherently constrained, as users' underlying states are not directly observable. We thus propose the Mind2Dialogue framework to mitigate this gap by simulating users' mental states and turning them into privileged supervision for human-aware training. Specifically, we first propose a psychology-guided simulator that preserves personal characteristics while updating mental states through interaction to generate coherent conversations. The key idea is to enforce a shared evolving mental state that drives user behavior and guides an Oracle assistant's responses. Our privileged distillation then trains models on the Oracle's well-informed responses to assist users without direct access to their mental states at deployment. Moreover, we propose to evaluate human-aware learning by combining personalization and theory of mind, examining how models understand people and act on that understanding. Training on the full Mind2Dialogue corpus improves every reported personalization metric over the corresponding Qwen, Llama, and OLMo instruction-tuned baselines, including gains of 26.6 to 40.9 percentage points in preference-following generation. The gains extend to belief and action reasoning on Qwen and Llama, beyond personalized assistance. Looking forward, Mind2Dialogue makes user simulation a foundation for genuine AI collaborators that understand beliefs and intentions behind people's words and support their long-term goals across education, work, and everyday life.
1 Introduction
Mind2Dialogue addresses the supervision gap in human-aware language-model training by using simulated, evolving user states to generate privileged assistant supervision. Its corpus improves personalization broadly and also transfers to belief and action reasoning on several models.
- 1 Introduction: The framework targets a supervision gap because useful responses should reflect users’ unspoken beliefs, goals, and circumstances, while private informed conversations are difficult to collect at scale.Public dialogue data offer scale but limited grounding in users’ hidden states.
- 1 Introduction: Mind2Dialogue uses shared evolving mental states to generate user behavior and Oracle responses, then distills those state-informed responses into students that lack direct state access.M2D-SIM generates interactions and M2D-CORPUS supports training while the state remains withheld from the student.
- 1 Introduction: The gains extend beyond personalized assistance to belief and action reasoning on Qwen and Llama, although OLMo shows mixed theory-of-mind results.Qwen gains 13.0 points on BigToM forward-belief accuracy, while OLMo improves on ToMi but declines on both BigToM tasks.
2 Related Work
Mind2Dialogue combines personalization, synthetic user simulation, social reasoning, and privileged-information distillation into an integrated approach to human-aware assistant training. Its distinctive choice is to let an Oracle observe the simulator-defined state while students learn from observable inputs and target responses.
- 2 Related Work: Existing personalization methods use external memory, retrieval, parameter adaptation, preference supervision, or dialogue world models to reflect individual users.These approaches address user-specific assistance but do not by themselves establish Mind2Dialogue’s shared-state Oracle supervision.
- 2 Related Work: Synthetic-dialogue research ranges from fixed persona conditioning to multi-turn simulators whose behavior reflects memory and changing circumstances.Mind2Dialogue focuses on which component observes the simulator-defined user state when responses are generated.
- 2 Related Work: Social-intelligence work studies inferring mental states and using them to predict or judge behavior, while Mind2Dialogue uses simulated state-informed responses for assistant training.Its evaluation connects personalization with theory-of-mind reasoning rather than treating them as identical objectives.
- 2 Related Work: Mind2Dialogue differs from prior privileged-information approaches by constructing supervision through an integrated simulator, corpus, and fixed-Oracle supervised-training pipeline.The Oracle observes the state generating user behavior; the student receives only observable inputs and target responses.
3 The Mind2Dialogue Framework
The Mind2Dialogue framework simulates coherent interactions from personas, scenarios, and a shared evolving state, then converts Oracle responses into student training data with the state withheld. Privileged distillation transfers state-informed behavior through observable dialogue context rather than requiring an explicit internal state representation.
- 3 The Mind2Dialogue Framework: M2D-SIM generates multi-turn interactions from a persona and shared structured state, M2D-CORPUS stores dialogues and QA examples, and M2D-CHAT trains students without state inputs.The teacher and student views distinguish state access during generation from observable inputs during training and deployment.
- 3.1 Framework Overview and Supervision Design: The framework defines student inputs as dialogue history through the user message while excluding the evolving state, creating targets whose construction uses information unavailable to the student.The student policy learns from available evidence even though the response target depends on partially unobserved user state.
- 3.1 Framework Overview and Supervision Design: A shared state connects simulated user behavior with Oracle responses, and subsequent state updates include the assistant response so later turns can reflect interaction-induced changes.The state evolves alongside the dialogue rather than remaining a static persona description.
- 3.2 Psychology-Guided Stateful User Simulation: The simulator preserves personal characteristics while updating internal states and behavior in response to situational demands and interaction history.Its state includes slowly changing values and constraints alongside transient mood, concerns, and turn-level judgments; these are simulator-defined control variables, not measurements of real users.
- 3 The Mind2Dialogue Framework: Scenario construction, behavior control, quality filtering, dialogue examples, and derived QA provide diverse yet coherent supervision across interaction types.Filtering checks schema, structure, state completeness, profile binding, persona consistency, and conflicts with fixed attributes.
- 3.4 Privileged-State Supervised Distillation: Privileged distillation trains a student on visible input-response pairs whose targets are generated using state access, optimizing response behavior without prescribing an internal state representation.The training objective is standard autoregressive cross-entropy over target response tokens, and its population optimum matches the state-marginalized Oracle response distribution.
4 Training Human-Aware Language Models
Mind2Dialogue trains human-aware models from simulated user states, scaling coherent personalized supervision without annotating every dialogue. Across model families and evaluations, the resulting training improves personalization broadly and extends to belief and action reasoning, with effects varying by backbone and supervision scale.
- 4.1 Training Data and Recipe: The 6,330-conversation analysis subset covers 47 scenarios, 10 persona-specialization clusters, and 14 user-behavior modes across six families.Behavioral analysis uses embedding-based matching, and the generation pipeline supports additional personas, scenarios, and modes without per-dialogue human annotation.
- 4.1 Training Data and Recipe: The corpus combines multi-turn Oracle dialogues with derived question-answer examples, while the student learns from visible inputs without access to evolving mental states.M2D-SIM uses scenarios, state updates, and behavior control to preserve diversity and interaction coherence without human annotation for every dialogue.
- 4.3 Simulation Quality: The simulation pilot shows larger M2D-SIM advantages at later turns, with topic-depth effect sizes reaching d=1.22 and assistant-personalization gap 0.51.Effect-size trajectories correlate positively with turn index, with r=0.90 and r=0.79.
- 4.4 Evaluation Results: Full M2D-CORPUS training improves every reported personalization metric across Qwen, Llama, and OLMo, with PrefEval-Gen gains of 26.6 to 40.9 points.The full mixture is best for each backbone; PersonaMem gains are smaller and scaling trajectories differ across models.
- 4.4 Evaluation Results: M2D-CHAT leads non-proprietary personalization baselines on all four Qwen metrics, including PrefEval generation at 56.8, 13.2 points above HumanLM.It exceeds GPT-4o-mini in three evaluations but remains below GPT-5-mini in all four.
- 4.4 Evaluation Results: Training mixture composition and scale affect tasks differently: dialogue-only training leads generation and ToM averages, whereas QA-only training leads PersonaMem-v2 MCQ but underperforms the unmodified model on generation.The full mixture reaches the highest ToM average at 57.1 while retaining generation performance close to dialogue-only training.
5 Conclusion
Mind2Dialogue addresses the supervision gap by using simulated user states to generate privileged assistant supervision. Training on M2D-CORPUS improves personalization across three model families and improves belief and action prediction for Qwen and Llama, although OLMo shows mixed reasoning results.
- Mind2Dialogue uses shared-state simulation, Oracle supervision, corpus construction, and privileged distillation with user states withheld from the student.The framework gives the Oracle direct access to evolving beliefs, goals, and circumstances while training deployable students only on visible inputs and target responses.
- Training on M2D-CORPUS improves every reported personalization metric across three model families.
- The same supervision improves belief and action prediction on Qwen and Llama, whereas OLMo gains personalization but loses on some mental-state reasoning tasks.The conclusion characterizes OLMo’s results as mixed rather than uniformly positive.
- User simulation offers a way to design training experiences through which language models learn to understand the people they serve.
A Limitations
The evidence is bounded by synthetic supervision, tested-model dependence, benchmark separation, hand-designed state representations, and the absence of validation in sustained real-user interaction.
- All data generation uses GPT-4o-mini as simulator and Oracle, so the corpus may inherit its errors, stylistic biases, and limitations in representing user states.The study does not test stronger teachers, ensembles, or alternative simulator families.
- Theory-of-mind transfer is specific to the tested models and settings, with OLMo improving on ToMi but declining on both BigToM tasks.The experiments do not distinguish whether pretraining, instruction tuning, optimization, or evaluation sensitivity explains the difference.
- The evaluation measures assistance and mental-state reasoning on separate benchmark suites, leaving their coordination within one interaction untested.The supervision-view ablation also varies supervision content and data volume together, so it does not isolate the student’s reasoning process.
- The hand-designed latent-state schema may omit relevant user aspects or encode distinctions that do not transfer to real interactions.Alternative schemas and sensitivity to individual state fields are not evaluated, and extensions require empirical validation.
- The benchmark, LLM-judgment, and human-audit evaluation does not establish performance in sustained interaction with real users.The authors call for longitudinal studies with informed consent and privacy protections before conclusions about real-user utility.
B Broader Impact
Mind2Dialogue offers a state-conditioned alternative to private interaction logs and static persona data for studying human-aware assistance. Its broader-impact discussion also emphasizes unresolved risks around profiling, stereotypes, uncertainty, and deployment safety.
- Positive impact: Mind2Dialogue provides synthetic training data in which simulated users and assistants are conditioned on a constructed state, avoiding collection of private human-assistant conversations.The corpus exposes state and generation metadata for replication and comparison, while synthetic generation does not guarantee demographic balance, factual accuracy, or privacy safety.
- Risks and mitigation: The paper identifies targeted persuasion, dependency, unwanted profiling, stereotyping, unwarranted mental-state confidence, and privacy leakage as risks that require direct evaluation and governance.It recommends communicating uncertainty, providing user control over personalization and memory, and applying domain-appropriate safety and data-governance measures.
- Privileged supervision: Unlike approaches that only infer or estimate user states, Mind2Dialogue gives its Oracle direct access to simulator-defined state when constructing response targets, then transfers those decisions through supervised distillation.The student operates without direct state access, preserving the separation between privileged supervision and deployment inputs.
- Evaluation scope: The framework connects personalization with theory of mind, using separate benchmarks to assess preference- and circumstance-aware assistance alongside reasoning about beliefs and actions.These evaluations examine whether models both use information about users and transfer simulated-interaction learning to explicit mental-state reasoning.
D Complete Evaluation Results and Simulation Analysis
The appendix reports complete student-scaling results together with simulator validation and component ablations supporting the main evaluation sections.
- Evaluation results: The complete student-scaling results underlying Section 4.4 are reported before simulator validation and component ablation results supporting Section 4.2.The passage describes the organization of the reported analyses rather than presenting outcome values.
- Simulation analysis: Simulator validation and component ablation are included as supporting analyses for the evaluation.These analyses accompany the complete scaling results in the reported appendix material.
- Analysis organization: The appendix presentation links student scaling with tests of the simulator and its components.No comparative scores or ablation outcomes are stated in the supplied passage.
D.1 Complete Data Scaling Results
Complete scaling results show that full-mixture training improves every reported personalization metric, while theory-of-mind outcomes vary across model backbones.
- Scaling results: Full-mixture training improves every reported personalization metric across the evaluated training scales.Table 6 reports personalization and mental-state reasoning panels, with changes shown relative to each backbone’s base model.
- Scaling results: Theory-of-mind outcomes vary across backbones rather than showing a uniform improvement pattern.The table distinguishes personalization from mental-state reasoning and reports results across multiple model backbones.
- Evaluation measures: Table 6 reports accuracy percentages for personalization and mental-state reasoning, with full-mixture changes measured from the base model.The table includes PM and BigToM metrics, with BigToM requiring correct answers on both true- and false-belief variants.
D.2 Simulator Validation and Component Ablations
M2D-SIM produces stronger trajectory-level judgments than Vanilla and maintains user persona specificity better than conditions lacking state maintenance. The validation uses held-out personas, while ablations isolate state and behavior controls.
- Simulator validation: M2D-SIM scores higher than Vanilla across all five judge dimensions on 20 held-out personas, with a composite Z-score gap of 1.15 (d=2.12).M2D-SIM receives the higher aggregate score for all 20 paired personas (binomial p<0.001).
- Component ablations: State maintenance supports persona specificity: removing the state document lowers turn-20 scores, while state-maintaining conditions increase over the evaluated horizon.Vanilla has d=−0.66 and Profile-only Oracle d=−0.49 relative to the full simulator; the table defines negative d as lower final specificity.
- Component ablations: The largest ablation effect is for goal specificity (d=1.03), whereas identity and communication effects are near zero.Vanilla user messages are 2.4× longer than M2D-SIM messages, so verbosity alone does not explain the gap.
- Simulation procedure: M2D-SIM combines evolving state maintenance with behavior guidance in a shared-state rollout that updates unresolved goals, affect, and trust across turns.The generation algorithm samples scenarios, updates state, selects behavior families, and generates user and Oracle turns.
F Dataset Statistics
M2D-CORPUS includes multiple corpus views built from 289 personas, with the deep-scenario subset analyzed across 47 unevenly distributed categories. The dataset also groups 153 specialization values into ten clusters to characterize conversation composition.
- Dataset scope: The deep-scenario subset contains 6,330 conversations paired with full latent-state trajectories, drawn from personas spanning five focal countries.The public release also includes broader rollout and QA-format components.
- Persona composition: The persona pool contains 153 specialization values clustered into ten K-means groups, with each cluster labeled by its most frequent specialization.Clustering uses all-MiniLM-L6-v2 embeddings, K = 10, and random seed 42.
- Scenario distribution: Scenario frequencies are uneven across 47 categories, with Figure 8 showing the 25 largest categories and cumulative coverage as categories are added by frequency.The figure reports counts for the largest categories and cumulative conversation coverage across all categories.
F.3 Alignment to profile and scenario
Assistant responses align more strongly with scenario prompts than with static personas across all analyzed sources. The analysis classifies behavior and visualizes how mode-family composition changes across dialogue turns and data sources.
- Alignment method: Alignment is measured by cosine similarity between assistant text and separate persona and scenario references, with distributions and paired medians and means shown in Figure 9.The scenario reference combines the scenario prompt with its context.
- Alignment results: Scenario alignment exceeds static-persona alignment in all three analyzed sources, with paired Wilcoxon tests reporting p < 0.001 in every case.The subsets contain 471 affective, 897 highfreq, and 2,187 lifelong conversations.
- Behavior classification: User-turn behavior is assigned to the nearest of 16 mode prototypes using normalized all-MiniLM-L6-v2 embeddings and cosine similarity.Each prototype concatenates a mode description with up to five catalog examples.
- Behavior dynamics: Figure 10 tracks six behavior-family proportions by turn position and conversation source after excluding fallback modes and renormalizing.Plots extend through turn 12 and include a point only when at least five retained turns are available.
- Training views: The training mixture combines dialogue demonstrations and three QA formats derived from the same latent-state trajectories without exposing the latent-state document to the student.The mixture contains 8,244 examples and excludes evaluation-benchmark instances and content.
F.6 Training and Evaluation Data Provenance
The training and evaluation data are generated from simulator trajectories, filtered through automated and judge-based quality checks, and audited for exact benchmark overlap. The provenance design documents both broad scenario coverage and limits of the overlap audit.
- Benchmark overlap: The released corpus contains no exact matching 13-grams with any of five evaluation suites, according to the benchmark overlap audit.Table 10 reports the benchmark, evaluation target, and number of detected exact matches.
- Audit limitations: The audit cannot rule out paraphrased overlap or benchmark exposure during teacher or student pretraining, and PersonaMem MCQ shares a response interface with one training component.Consequently, gains on PersonaMem MCQ are interpreted under a matched-format condition.
- Quality control: Each generated conversation passes six quality checks covering programmatic record properties and LLM-judge assessments of persona consistency.Programmatic checks include schema validity, turn structure and token bounds, trajectory completeness, and profile binding.
- Human audit: A human audit found that 1,216 of 1,240 sampled conversations passing automatic filtering satisfied the evaluation rubric, or 98.06%.The rubric covered persona consistency, trajectory coherence, state-response consistency, and response relevance.
- Scenario construction: Scenario generation distributes prompts across life phases and psychological threads while covering categories such as relationships, planning, career, health, identity, and family.The instructions require at least two life phases and at least one repeated psychological thread across scenarios.
I Qualitative Examples of Profile-Grounded Personalization in the Training Data
The qualitative examples show assistants grounding responses in profile-specific routines, values, relationships, and priorities. Across cases, responses translate those details into tailored conversational openings, communication strategies, and actionable advice.
- Grounding interpretation: The examples are presented as profile-grounded response spans, including paraphrases and elaborations rather than requiring verbatim copying from the persona excerpt.The reading guide explicitly distinguishes linked response content from direct quotation.
- Interpersonal advice: For a taciturn bar owner, responses suggest low-pressure ways to reconnect with his sister using familiar topics, local activities, memories, and bar-related routines.The advice recommends genuine patience and gradual movement toward deeper connection.
- Family-conflict mediation: For a tradition-anchored retiree, responses acknowledge comfort with letters and phone calls while proposing balanced, collaborative adoption of technology.Suggested strategies include a quiet setting, first-person statements, personal stories, and compromise.
- Affect regulation: For an Italian restaurant owner focused on quality and exclusivity, responses connect anxiety about competition to reputation and recommend market analysis, differentiation, premium quality, and client engagement.The proposed actions include exclusive dishes, tastings, and stronger relationships with VIP clients.
J Dataset Samples
The synthetic dialogues depict users reflecting on engineering decisions and balancing professional priorities with personal needs. Assistants respond by connecting those concerns to practical experience, quality, safety, mentoring, and focused work.
- Reflective dialogue: Reflective dialogue links an early infrastructure project and possible quality compromises to the user’s current engineering practices.The assistant invites the user to identify which aspects of that project most shaped present decision-making.
- Engineering decision-making: The dialogues connect infrastructure decision-making to trade-offs among speed, cost, quality, safety, and long-term resilience.The assistant frames hands-on experience as supporting contextual risk assessment and balanced choices in high-stakes conditions.
- Mentoring: Mentoring guidance emphasizes real-life case studies, structured lessons learned, and scenario-based training to reinforce quality and safety.These strategies use concrete experiences and simulations to make high-stakes lessons relatable and encourage collective accountability.
- Personal preferences: The assistant addresses the user’s preference for deep one-on-one conversation while recommending boundaries and scheduled breaks to protect solitude for creative work.The proposed balance reserves specific times for meaningful dialogue and focused projects.