Paper deep dive
Cognitive World Models for Process-Level Social Influence Evaluation
Minghui Ma, Bin Guo, Han Wang, Mengqi Chen, Jingqi Liu, Yan Liu, Zhiwen Yu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 7/5/2026, 2:07:20 AM
Summary
The paper proposes the Cognitive World Model (CogWM), an LLM-based user model designed to evaluate Social Influence Dialogue (SID) by tracking changes in a user's internal cognitive states (Beliefs, Desires, Intentions, and Emotions - BDI/E). Unlike traditional surface-level metrics or terminal LLM-as-judge scores, CogWM provides a process-oriented evaluation through a three-tier framework: turn-level fidelity, trajectory-level state dynamics, and task-level composite scoring (CTS). The model is trained using a Summarize-and-Allocate (SaA) pipeline on 150,454 samples and demonstrates superior performance in simulating user cognitive transitions and distinguishing the influence capabilities of various commercial LLM agents.
Entities (7)
Relation Signals (5)
BDI/E â comprises â Belief
confidence 100% · the user state decomposes into four dimensions: Belief, Desire, Intention, and Emotion.
CogWM â evaluates â Social Influence Dialogue
confidence 100% · CogWM moves social influence dialogue evaluation from terminal judgment to process tracking.
CogWM â isbasedon â Qwen3-14B
confidence 100% · CogWM built on Qwen3-14B
CogWM â uses â Summarize-and-Allocate
confidence 100% · Trained via our Summarize-and-Allocate (SaA) annotation pipeline
Composite Trajectory Score â measures â BDI/E trajectory
confidence 90% · the task-level evaluation via CTS and GoodOutcome (GO).
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Social influence dialogue changes user behavior by altering internal cognitive states. The central evaluation question is whether the user's beliefs, desires, intentions, and emotions measurably change over the course of conversation, a process-oriented criterion that neither surface-level text metrics (BLEU/ROUGE) nor single-score LLM judgments can capture. We propose the \textbf{Cog}nitive \textbf{W}orld \textbf{M}odel \textbf{(CogWM)}, an LLM-based user model that reframes multi-turn dialogue evaluation from ``what did the user say'' to ``how did the user's internal cognitive state evolves.'' CogWM jointly predicts BDI/E cognitive states and user utterances and serves as both a user simulator and an evaluation platform, using a three-tier evaluation framework that covers turn-level fidelity, trajectory-level state dynamics, and task-level composite scoring. Trained via our \textbf{S}ummarize-\textbf{a}nd-\textbf{A}llocate \textbf{(SaA)} annotation pipeline on 150,454 user-turn samples across four social influence scenarios, CogWM achieves 77.6\% emotion accuracy (2.1$\times$ over GPT-5.5). In 3600 multi-agent discrimination trials, it distinguishes six commercial agents by their cognitive influence, with Llama-4-Scout ranking first (CTS +0.233). CogWM moves social influence dialogue evaluation from terminal judgment to process tracking. We have released our code\footnote{\scriptsize Code: this https URL} and models\footnote{Model: this https URL}.
Tags
Links
- Source: https://arxiv.org/abs/2606.29495v1
- Canonical: https://arxiv.org/abs/2606.29495v1
Trouble viewing inline? Open PDF directly â
Full Text
101,946 characters extracted from source content.
Expand or collapse full text
Cognitive World Models for Process-Level Social Influence Evaluation Minghui Ma1, Bin Guo1, Han Wang , Mengqi Chen1, Jingqi Liu1, Yan Liu1, Zhiwen Yu1, 2 Abstract Social influence dialogue changes user behavior by altering internal cognitive states. The central evaluation question is whether the userâs beliefs, desires, intentions, and emotions measurably change over the course of conversation, a process-oriented criterion that neither surface-level text metrics (BLEU/ROUGE) nor single-score LLM judgments can capture. We propose the Cognitive World Model (CogWM), an LLM-based user model that reframes multi-turn dialogue evaluation from âwhat did the user sayâ to âhow did the userâs internal cognitive state evolves.â CogWM jointly predicts BDI/E cognitive states and user utterances and serves as both a user simulator and an evaluation platform, using a three-tier evaluation framework that covers turn-level fidelity, trajectory-level state dynamics, and task-level composite scoring. Trained via our Summarize-and-Allocate (SaA) annotation pipeline on 150,454 user-turn samples across four social influence scenarios, CogWM achieves 77.6% emotion accuracy (2.1Ă over GPT-5.5). In 3600 multi-agent discrimination trials, it distinguishes six commercial agents by their cognitive influence, with Llama-4-Scout ranking first (CTS +0.233). CogWM moves social influence dialogue evaluation from terminal judgment to process tracking. We have released our code111Code: https://github.com/lucianma05-create/CogWM and models222Model: https://w.modelscope.cn/models/LucianMa/CogWM-14B. 1 Introduction Social Influence Dialogue (SID) (Chawla et al. 2023) appears throughout human communication, including empathetic responses in emotional support, persuasion in charitable giving, and preference negotiation in everyday recommendations. These scenarios share one essential feature: one party aims to understand and influence the otherâs internal cognitive state. Task-Oriented Dialogue (Chen et al. 2017; Gao et al. 2019) (e.g., booking a restaurant, querying bus schedules) targets external task completion as its endpoint. Social influence dialogue instead targets changes in the userâs beliefs, desires, intentions, and emotions. Non-Cooperative Dialogue (Lewis et al. 2017; Chawla et al. 2023) (e.g., negotiation, debate) presupposes a fundamental conflict of goals. Social influence dialogue does not presuppose such conflict; therapist and patient collaborate, while persuader and persuadee may simply hold differing positions. Negotiation spans all three categories (Chawla et al. 2023): it involves an agreement to complete (task-oriented), cognitive change (social influence), and conflicting interests (non-cooperative). These categories are not mutually exclusive (Figure 1). Because SID evaluation targets internal cognitive state change, it can neither adopt the success-rate metrics of task-oriented dialogue nor default to the game-theoretic equilibrium of non-cooperative dialogue. As LLM-driven dialogue agents are deployed in these scenarios (Zhao et al. 2023), a pressing question emerges: how do we evaluate whether an agent has genuinely improved the userâs cognitive state? Figure 2 illustrates a concrete case: the user enters in a high-anxiety state, and through turn-by-turn empathetic responses, measurable cognitive state shifts occur, a process invisible to terminal scoring. Figure 1: Three dialogue paradigms and their evaluation objectives. Figure 2: A social influence dialogue example: the user transitions from a high-anxiety state (âafraid of messing upâ) through empathetic responses to eventual anxiety relief, a process accompanied by turn-level cognitive state shifts. Current evaluation methods fall into four categories, each with distinct strengths and blind spots. Surface Text Metrics. N-gram overlap measures such as BLEU (Papineni et al. 2002) and ROUGE (Lin 2004) are computationally cheap and scalable, but they ignore the social influence objective entirely. A high-bleu response may do nothing to guide a better decision. LLM-as-Judge. Methods like MT-Bench (Zheng et al. 2023) and AlpacaEval333https://github.com/tatsu-lab/alpaca_eval use strong LLMs to rate dialogue quality on a 1â5 or A/B scale. They capture semantic quality better than N-gram metrics but remain fundamentally terminal judgments: a single score after the dialogue ends reveals nothing about how the userâs internal cognitive state evolved during the conversation. The scoring logic is also a black box, making it hard to explain why one response is judged better than another. Human Evaluation. Human judgments of fluency, relevance, and empathy are the gold standard, but they are expensive, difficult to scale, and limited by inter-annotator agreement (Belz et al. 2020). For social influence scenarios, human evaluators can give an overall impression but struggle to track cognitive states at the granularity of individual turns. User Simulators. Agenda-based rule systems and LLM-based neural simulators (DuetSim (Luo et al. 2024), ChatBench (Chang et al. 2025), PlatoLM (Kong et al. 2024)) evaluate dialogue systems by automatically playing the user role. They enable high-quality, low-cost automated evaluation, but existing simulators only predict the userâs next utterance. They model what the user says, not why; beliefs, desires, intentions, and emotions remain unobserved. We propose the Cognitive World Model (CogWM). Its core idea is to model multi-turn dialogue as a trajectory of user cognitive state transitions. At each turn, the user state decomposes into four dimensions: Belief, i.e., what the user believes, knows, or judges; Desire, i.e., what the user wants, needs, or prefers; Intention, i.e., what the user plans, decides, or commits to doing; and Emotion, i.e., the userâs current affective state and its intensity. We refer to these four dimensions as the BDI/E framework. Unlike traditional simulators that only output text, CogWM predicts both the userâs BDI/E internal state and the textual response. This dual output makes CogWM a unified evaluation platform: an SID agent under test is connected to CogWM, which simulates the user, and the resulting BDI/E trajectory quantifies the agentâs influence in SID. We further structure this evaluation into three tiers. The turn tier assesses single-response quality with 8 automatic metrics and 5 LLM-as-Judge semantic metrics. The trajectory tier models multi-turn state evolution using 4 dynamic indicators adapted from financial time series analysis, applied across all four BDI/E dimensions. The task tier aggregates these into two composite scores, the task-weighted Comp metric and the Composite Trajectory Score (CTS), which measures global influence. Together, the three tiers provide a quantifiable, decomposable, and interpretable assessment of any agentâs capacity to change user cognitive states. Table 1 contrasts CogWM with existing methods. CogWM uniquely satisfies all dimensions: scalability (enabling automatic, low-cost evaluation through model-based simulation), cognitive state modeling (the only method tracking beliefs, desires, intentions, and emotions simultaneously), interpretability (BDI/E trajectories make evaluation conclusions traceable to specific turns and dimensions), fine-grained process evaluation (turn-level cognitive change signals rather than a terminal score). Table 1: Comparison of Key Capabilities across Evaluation Methods and CogWM Method Scal. Cogn. Interpr. Process Text Metrics â â â â LLM-as-Judge â â â â Human Eval â â â â User Simulator â â â â green!5 CogWM (Ours) â â â â Building upon these insights, our work makes the following contributions: (1) A three-tier BDI/E evaluation framework, adapting the classical BDI model to LLM-era dialogue evaluation. The framework spans turn-level fidelity assessment (8 automatic + 5 LLM-as-Judge indicators), trajectory-level state modeling (4 indicators adapted from financial time series), and task-level evaluation via CTS and GoodOutcome (GO). (2) The Summarize-and-Allocate (SaA) two-stage annotation method, producing 23,872 annotated dialogues and 150,454 user-turn samples across four social influence scenarios (BDI recall 0.79, Cohenâs Îș 0.69). (3) The CogWM built on Qwen3-14B, jointly trained on user cognitive state prediction and text generation, surpassing all external baselines in emotion, belief, desire and intention state prediction with a 2.8Ă E-Acc gain over the same base model in zero-shot experiments. (4) Extensive experiments validating CogWM both as a user simulator (fidelity) and as an evaluation platform (utility). In 3,600 multi-agent discrimination trials, CogWM distinguishes six commercial-grade agents by their cognitive influence, with Llama-4-Scout ranking first (CTS +0.233). The CTS-GO complementarity further reveals that trajectory-level cognitive improvement and dialogue-level behavioral outcomes are not equivalent. 2 Related Work Chawla et al. (Chawla et al. 2023) defined social influence dialogue and identified the core challenge of evaluation fragmentation. BDI modeling (Bratman 1987; Rao and Georgeff 1991) has progressed from symbolic logic to neural representations (Li et al. 2021; Xu and others 2024; Qiu et al. 2023) and serialized prediction (Ma et al. 2026), but existing work models single BDI dimensions and performs one-sided inference without joint utterance generation. User simulators have evolved from agenda-driven to LLM-based (Luo et al. 2024; Kong et al. 2024); recent approaches touch on state modeling (Wang et al. 2025; Wu and others 2025) but produce implicit, non-interpretable vectors. LLM-as-judge (Zheng et al. 2023; Dubois et al. 2024) provides scalable semantic scoring but remains terminal black-box evaluation; EMPA (Zhang and others 2026) introduced process tracking for empathy in an implicit space. CogWM differs from all the above by jointly generating utterances and explicit, interpretable BDI/E states, enabling turn-level cognitive process evaluation across four scenarios. See Appendix B for details. 3 Cognitive World Model Framework 3.1 Problem Formulation A multi-turn dialogue consists of alternating utterances (u1,a1,âŠ,uT,aT)(u_1,a_1,âŠ,u_T,a_T). CogWM learns the mapping (P,C,H<t,Stâ1)â(St,Ut)(P,C,H_<t,S_t-1)â(S_t,U_t), where P is the user profile, C the task context, H<tH_<t the dialogue history, St=Emotiont,Beliefst,Desirest,IntentionstS_t=\Emotion_t,Beliefs_t,Desires_t,Intentions_t\ the BDI/E cognitive state, and UtU_t the user utterance (Figure 3). CogWM models the user from the perspective of the assistant agent: given the assistantâs action, predict how the user will respond cognitively and textually, enabling turn-level simulation of user cognitive dynamics. Figure 3: CogWM takes user profiles, task context, and dialogue history as input, and outputs BDI/E states (Belief/Desire/Intention/Emotion) and natural language responses turn by turn, forming a multi-turn cognitive state trajectory. 3.2 BDI/E Cognitive State Representation Table 2 defines the four BDI/E dimensions and their roles in social influence evaluation. The four dimensions also depend on each other. In emotional support, for instance, the belief "my work pressure is too high" may trigger the desire "I want to alleviate my anxiety." Belief and Desire together shape the userâs Intention, i.e., whether they choose to rest, seek help, or adjust their work arrangements. Table 2: Definitions of BDI/E Dimensions and Their Roles in Social Influence Evaluation Definition Role Belief What the user believes, assumes, judges, or interprets. Beliefs may be correct, incorrect, or unverifiable. Social influence primarily operates by altering cognition: changing what the user believes or correcting misconceptions. Desire What the user wants, needs, prefers, avoids, or seeks. Social influence aims to shape motivation: encouraging or reshaping what the user desires toward beneficial outcomes. Intention What the user plans, decides, accepts, rejects, or commits to doing. The ultimate goal of social influence: driving behavioral commitment and actionable decisions. Emotion The userâs current affective state and its intensity. Emotion serves as the contextual background and regulatory signal in social influence processes. 3.3 Task-Specific Quantitative Scores The definition of "success" varies across different social influence tasks. In CogWM, each BDI dimension is assigned a task-specific quantitative score in the range [0,2][0,2], which transforms natural language psychological states into structured signals to support trajectory-level evaluation. In emotional support conversation (ESC), effective support should help users form more realistic beliefs, healthier desires, and more feasible coping intentions. In persuasion dialogue (PD), effective persuasion should enhance usersâ awareness of donation impact, strengthen prosocial motivation, and foster autonomous commitment, rather than simply rewarding higher donation amounts. In conversational recommendation (CR), effective recommendation should better match user preferences, satisfy recommendation needs, and support users in making clear acceptance or rejection decisions. Table 3 summarizes how each BDI dimension is instantiated across different tasks. Emotion is modeled as a structured variable with 12 discrete categories and intensity-aware scoring. See Appendix C.1 for details. Table 3: Task-specific BDI Quantification Scores Scenario Belief Desire Intention ESC reality rationality feasibility PD impact prosocial commitment CR alignment satisfaction acceptance 3.4 User Profile Conditioning To learn the linguistic habits of users with different personalities and traits in SID tasks, CogWM takes user profiles as conditional input. The profile consists of the following components: Big Five Personality Traits: Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism, each scored 1â5. Communication Style: A natural language description of the userâs typical mode of expression. Decision-Making Style: A description of how the user makes decisions (e.g., cautious vs. impulsive). Task-Related Features: User-specific key factors for a given scenario. 4 Data Construction & Training We select four public dialogue datasets, including DailyDialog (Li et al. 2017) for open-domain conversation, and three social influence scenarios: emotional support (ESConv) (Liu et al. 2021), charitable donation persuasion (P4G) (Wang et al. 2019), and task-oriented recommendation (DuRecDial) (Liu et al. 2020), covering a spectrum from everyday dialogue to social influence interactions. 4.1 User Profile Annotation Following prior work on inferring user profiles from text and dialogue (Mehta et al. 2020; Wang et al. 2025), we use an LLM to generate user profile annotations based on the complete dialogue and available metadata. Because personality inference from short dialogues is inherently unreliable, CogWM uses a confidence field (level 1â3) to help the model distinguish strong signals from weak ones, avoiding overfitting to uncertain profile signals. 4.2 BDI/E Annotation: Summarize-and-Allocate User states in SID evolve as a structured cognitive process over time, not as isolated turn-level reactions (§3.2). We propose Summarize-and-Allocate (SaA): Stage 1 (Global Summarization) extracts global BDI items and scores from the full dialogue (§3.3); Stage 2 (Temporal Allocation) assigns per-turn BDI/E states with evidence grounding and global_ref linking. Human evaluation confirms the SaA pipelineâs effectiveness: global BDI recall 0.79, precision 0.77, emotion Cohenâs Îș=0.69Îș=0.69 (full results in §F.1). 4.3 Sample Construction and Training From the SaA annotations we construct 150,454 turn-level training samples. CogWM is jointly optimized for state prediction and utterance generation: pâ(ut+1,St+1âŁâ€t,P,C)p(u_t+1,S_t+1 _†t,P,C). We use Qwen3-14B Instruct (Yang et al. 2025a) with LoRA (Hu et al. 2022); all hyperparameters are in §E. 5 Composite Trajectory Score To quantitatively evaluate SID trajectories, we propose CTS. CTS combines three empirically low-redundant metrics: AUC (area under curve via trapezoidal integration), PSR (improvement frequency, positive step ratio), and Î (net change between initial and final states). We observe pairwise correlations |r| < 0.25 across 3,600 trajectories, indicating low redundancy among the three metrics. We define CTS as: CTS=0.40â AUC+0.30â PSR+0.30â ÎCTS=0.40·AUC+0.30·PSR+0.30· (1) These metrics are computed from the task-weighted per-turn composite score Compt. See §D for the full three-layer evaluation framework and correlation heatmap. 6 Experiments The experiments primarily address two meta-questions: (1) Can CogWM accurately predict usersâ cognitive states and textual responses? (§6.1) and (2) How do SID agents based on various high-performance models perform when interacting with CogWM? (§6.2) 6.1 Simulation Fidelity Verification We evaluate simulation fidelity along five aspects: response realism, BDI/E state prediction accuracy, stateâutterance consistency, learned understanding of user states, and the contribution of each profile component (ablation in §F.3). Table 4: Fidelity experiment results Model BDI/E Core Utterance LLM Judge E-Aâ B-Sâ D-Sâ I-Sâ E-Mâ R-Lâ B-4â LenR Emoâ Belâ Desâ Intâ Uttâ Cohâ GPT-5.5 36.4 .474 .433 .442 0.46 .237 .043 1.82 2.34 1.90 2.09 1.74 2.62 4.71 DS-V4-Pro 31.0 .462 .445 .412 0.52 .288 .060 1.81 2.13* 1.87 2.21 1.93 2.73 4.57 Dual-LLM 30.9 .500 .446 .445 0.47 .253 .054 2.12 2.13* 1.84 2.27 1.95 2.67 4.49 Qwen3-14B 28.1 .384 .358 .416 0.63* .196* .033* 2.26* 2.37 1.79 2.06 1.82 2.41 4.45 utt-only-14B 12.2* .285* .340* .264* â .427 .159 1.03 2.22 1.24* 1.17* 1.31* 1.13* 2.43* !8 CogWM-14B 77.6 .773 .705 .458 0.29 .417 .150 1.01 4.49 4.16 4.33 4.75 3.32 4.50 Experimental Setup All experiments use CogWM-14B and run on a single A100-80G GPU for inference. The test set comprises 7,563 user-turn samples (DailyDialog 3,765 / ESConv 1,137 / P4G 508 / DuRecDial 2,153). Evaluation Metrics Utterance. We use ROUGE-L (R-L) and BLEU-N (B-n) to evaluate the fluency of CogWMâs output, and LengthRatio (LenR) (|pârâeâd||gât| |pred||gt|, â1â 1 optimal) to evaluate the length ratio of the output relative to the ground truth. Distinct-1 (Dist-1) for lexical diversity. BDI/E. E-Acc(E-A) evaluates the accuracy of predicted emotion labels; MAE (-M) evaluates the MAE of BDI/E intensities; Sim (-S) evaluates the semantic similarity between predicted text and ground truth. LLM Judge. We use DeepSeek-V4-Pro as a judge, scoring six dimensions on a 1â5 scale: Emo, Bel, Des, Int, Utt evaluate the alignment of emotion, belief, desire, intention, and output text with the corresponding ground truth, respectively, and Coh evaluates the consistency between the predicted cognitive state and the predicted text. BDI/E annotations are sparse (61.7% of belief, 68.1% of desire, and 91.4% of intention turns are empty in the ground truth), so all BDI metrics in this paper are computed only on turns where the ground truth is non-empty (3,342 turns in total) to avoid the inflation effect where empty entries on both sides are counted as perfect matches. External Baselines We establish the following baselines: (1) Base model comparison: Qwen3-14B: no fine-tuning, outputs BDI/E states and user utterances solely via prompting, used to measure the zero-shot BDI/E capability of the base model; utt-only-14B: uses the same base model and training configuration as CogWM-14B, but is trained only on user utterance generation, serving as the âutterance quality upper bound under identical training data conditionsâ; (2) High-performance models: GPT-5.5 and DeepSeek-V4-Pro (DS) as strong API baselines; (3) Enhanced method: Dual-LLM: adapts the dual-LLM verification paradigm of DuetSim (Luo et al. 2024) with task-specific prompting, implemented using GPT-5.5. All baselines use the same output schema as CogWM. User Fidelity Experiment Table 4 presents the experimental results. BDI/E Core Metrics. CogWM outperforms all baselines across all four dimensions. CogWMâs advantage on I-Sim over API baselines stems from Intentionâs extreme sparsity, as only 9% of turns have non-empty ground truth. CogWM also leads in BDI quantitative score MAE, showing it not only tracks the semantic content of BDI but also matches the score levels precisely. Complete experimental results are presented in §F.2. Utterance Generation Metrics. CogWM significantly outperforms API baselines on all Utterance metrics (R-L 0.417 vs. GPT-5.5 0.237, B-4 0.150 vs. 0.043). The most telling metric is LenR: CogWM achieves 1.01 (â 1.0, nearly a perfect match with ground truth length), whereas GPT-5.5 achieves 1.82 and Qwen3-14B base achieves 2.26. API baselines tend to generate overly long responses. CogWM learns dialogue rhythm and density through joint training. Dis-1 shows no significant difference between CogWM (0.948) and API baselines (0.960â0.965), indicating that lexical diversity is primarily determined by the base modelâs capability. utt-only has poor BDI/E prediction, achieving the lowest scores across all metrics. This confirms that cognitive state tracking cannot be automatically acquired through shared input distribution or in-domain fine-tuning; it must be explicitly trained through joint optimization. LLM Judge Semantic Scores. CogWM achieves the best results on nearly all LLM evaluation metrics. The sole exception is the Coherence dimension: API baselines (4.45â4.71) are close to or slightly higher than CogWM (4.50). This pattern suggests that "self-consistency" is a common trait of general-purpose LLMs; API models can generate internally self-consistent yet factually incorrect BDI/E descriptions. For the evaluation platform, external accuracy (CogWMâs strength) takes priority over internal self-consistency (APIâs strength). Subsequent agent discrimination experiments validate this design choice (see §6.2). Table 5: State Understanding Results Model E-A B-Sim D-Sim I-Sim L-E L-B L-D L-I GPT-5.5 42.3 0.489 0.457 0.471 2.39 2.05 2.19 1.92 DSv4-Pro 31.9 0.468 0.430 0.425 1.95 2.02 2.20 1.89 Qwen3-14B 32.4 0.398 0.361 0.419 2.33 1.63 1.79 1.55 green!8 CogWM-14B 71.1 0.739 0.671 0.440 3.71 2.92 3.27 4.29 State Understanding Experiment To verify CogWMâs BDI/E state inference capability, we designed a state understanding experiment: given dialogue context and the GT user utterance, predict BDI/E states. Table 6: Agent Discrimination Results. # Agent Cond CTS AUC Î PSR mono E-Î B-Î D-Î I-Î nz% GO GO-E GO-P GO-D 1 Llama-4-Scout No-CoT +.233 -.006 +.430 .354 .320 +.637 +.158 +.277 +.270 37.1 68.7 79 66 61 !8 2 Llama-4-Scout CoT +.237 -.012 +.441 .367 .328 +.780 +.216 +.322 +.238 38.2 72.7 84 66 68 3 Qwen3.7-Plus No-CoT +.220 -.024 +.407 .358 .339 +.198 +.191 +.409 +.353 38.0 78.0 80 79 75 !8 4 Qwen3.7-Plus CoT +.236 -.023 +.472 .346 .315 +.493 +.219 +.402 +.261 36.2 79.3 90 68 80 5 Claude-Opus-4.7 No-CoT +.200 -.030 +.382 .323 .303 +.248 +.195 +.398 +.177 34.8 66.0 70 58 70 !8 6 Claude-Opus-4.7 CoT +.219 -.010 +.391 .352 .311 +.310 +.161 +.305 +.250 36.4 73.3 87 67 66 7 DeepSeek-V4-Pro No-CoT +.198 -.035 +.382 .325 .288 +.373 +.228 +.293 +.127 35.6 72.0 84 61 71 !8 8 DeepSeek-V4-Pro CoT +.199 -.026 +.364 .336 .302 +.272 +.182 +.394 +.233 35.3 70.7 83 59 70 9 GPT-5.5 No-CoT +.185 -.035 +.349 .316 .268 +.365 +.130 +.241 +.173 35.4 68.7 80 68 58 !8 10 GPT-5.5 CoT +.209 -.022 +.405 .322 .289 +.380 +.197 +.332 +.190 35.1 74.7 84 82 58 11 Gemini-3-Pro No-CoT +.140 -.065 +.278 .276 .230 +.618 +.104 +.078 +.153 30.4 44.7 54 19 61 !8 12 Gemini-3-Pro CoT +.153 -.045 +.292 .278 .204 +.453 +.110 +.194 +.113 30.7 47.3 70 20 52 Table 5 shows the results. CogWM still leads the API baselines in comprehension mode, supporting the view that joint training gives CogWM cognitive understanding ability. Notably, CogWMâs comprehension performance is slightly lower than its prediction ability (E-Acc: 71.1% vs 77.6%, -6.5p), while GPT-5.5âs comprehension is higher than its prediction (E-Acc: 42.3% vs 36.4%, +5.9p). This suggests that providing GT utterances helps untrained models, which struggle to generate utterances and BDI/E simultaneously, but slightly harms CogWM. Intention understanding is a common weakness across all models, because its choice space is large and hardest to predict, and its annotations are the sparsest in the training data. The experiment thus confirms the coupling between state inference and text generation; joint training makes the two mutually conditional. 6.2 Agent Discrimination Experiment In this section, CogWM shifts from being the evaluated subject to being the evaluation tool, tracking how several high-performance SID agents impact user BDI/E states through multi-turn interaction. Experimental Setup Six agents (GPT-5.5, Llama-4-Scout, DeepSeek-V4-Pro, Claude-Opus-4.7, Gemini-3-Pro, Qwen3.7-Plus) are prompted as assistants. CogWM-14B (tp=0.3) serves as the user simulator. 6 agents Ă 2 prompt conditions (with and without Chain-of-Thought, CoT) Ă 3 tasks Ă 100 dialogues = 3,600 interactions, with the first 2 rounds using real dialogue for cold start. Starting from the 8th turn, an LLM-based intent detection module is further employed to determine whether the dialogue has naturally reached completion; if so, the interaction is terminated, with a maximum dialogue length of 20 turns enforced. Evaluation Metrics We evaluate agents using the three-tier framework defined in §D. The primary metric is CTS (§5). Three auxiliary metrics complement trajectory-level evaluation: Monotonicity (mono) measures whether improvement steps consistently follow the overall Î direction. Nonzero Rate (nz%), measuring dialogue richness as the proportion of turns with non-empty BDI/E outputs, and GoodOutcome (GO), which uses an LLM Judge to determine dialogue-end behavioral outcomes, specifically whether the agent achieved emotional improvement or an action plan in ESConv, made a donation decision in P4G, or obtained a recommendation acceptance in DuRecDial, measuring a dimension different from CTS. Four patterns emerge. Table 6 presents the complete experimental results. CTS Ranking and Agent Typology. Llama-4-Scout ranks first in both conditions, consistently achieving the highest AUC and PSR. Qwen achieves the highest Î under CoT (+0.472) but ranks second in CTS due to lower AUC (-0.023, 5th). GPT-5.5 improves from 0.185 to 0.209 CTS under CoT (+13%). Gemini ranks last in CTS and shows the lowest nz% across settings. Based on CTS components, we identify four agent categories: All-round (Llama: high on all four), Sprint (Qwen CoT: very high Î but low AUC), Steady (Claude: high AUC but moderate Î ), and Hollow (Gemini: low on all four). Figure 5 illustrates consistent patterns across tasks, where Llama and Qwen show steeper trajectory changes while Gemini remains relatively flat. Figure 4: GPT-5.5 and Qwen3.7-Plus on the same ESConv context, both achieving GO=true, but with different trajectory patterns. GPT-5.5 achieves Î=+2.50 =+2.50, while Qwen3.7-Plus shows Î=0.00 =0.00 (std=0.06). Full dialogues are provided in Appendix C9. Complementary Perspectives of GO and CTS. CTS and GO capture different aspects of dialogue outcomes. Llama ranks first in CTS but fourth in GO (68.7%), while Qwen ranks second in CTS but first in GO (78.0%). GPT-5.5 achieves the largest GO improvement under CoT (+6.0p), followed by Claude (+7.3p). GO scores vary significantly across tasks, ranging from 54%+ on ESConv to 19%â82% on P4G. Gemini achieves the lowest GO score (19%) across settings. Figure 4 presents a representative case where the GO is the same but the trajectory patterns differ substantially. Appendix Cases C8âC9 further provide paired examples under identical dialogue contexts, showing different GO ith similar trajectory patterns, as well as similar GO with different trajectory patterns. Marginal Effects and Task Dependency of CoT. CoT affects agents asymmetrically. GPT-5.5 benefits most, while already-strong agents (Llama, Qwen) gain marginally and DeepSeek-V4-Pro shows negligible change. CoTâs value lies in correcting the tendency to skip ahead to solutions rather than in boosting raw reasoning: GPTâs emotion trajectory under CoT shifts from flat to clearly positive. Gains are task-dependent, with the strongest gains on ESConv and negligible gains on DuRecDial (Fig 7, Fig 8 and Fig 9). Per-task Analysis. We measure task-level discriminability using the average standard deviation across the four BDI/E-Î dimensions. ESConv shows the strongest discriminative performance with the highest average standard deviation (0.127), followed by P4G (0.117) and DuRecDial (0.105). Different datasets also exhibit different single-dimension discriminative patterns. ESConv shows the strongest discriminability on E-Î and I-Î (Std. = 0.252 and 0.096), P4G is most discriminative on B-Î (Std. = 0.072), and DuRecDial is most discriminative on D-Î (Std. = 0.114). These results are consistent with the capabilities emphasized by each task. The detailed per-task BDI/E trajectory statistics are provided in Appendix Table 20. Full six-model Ă three-task trajectory results are reported in Fig. 7 and Fig. 8. Task-specific cases are provided in Appendix G. Figure 5: Six-model BDI/E cross-task mean trajectory summary (No-CoT). Emotion is from ESConv only and rescaled from [â2.5,+2.0][-2.5,+2.0] to [0,2][0,2]. Shaded bands indicate ± 1 SEM. Î uses 5-region interpolated endpoints. 7 Discussion 7.1 Mechanisms of BDI/E Joint Training Fidelity experiments (§6.1) suggest that BDI/E state tracking emerges primarily from joint optimization of state prediction and utterance generation. Models trained without joint supervision or using utterance-only objectives fail to recover meaningful emotion signals, while joint training significantly improves four-dimensional state accuracy. Importantly, this improvement does not require scaling to larger models (Appendix F.6), indicating that cognitive state modeling is an architectural outcome rather than a capacity-driven effect. The coupling between state inference and text generation further indicates that both components benefit from shared representations learned during joint training. 7.2 Asymmetric Contribution of Profile Conditions Ablation experiments (§F.3) reveal an asymmetric contribution pattern: user profile components contribute minimally to state prediction individually (any single field can be removed with under 2p E-Acc change) but are critical for personalized generation (removing all profiles cuts BLEU-4 by 72%). Task context alone cannot substitute for profiles in driving utterance personalization. This supports CogWMâs design: profiles condition generation, while dialogue history handles state tracking. 7.3 Evaluation Insights from CTS BDI/E trajectory decomposition provides evaluation signals that are not captured by terminal outcome metrics. We first observe a consistent divergence between cognitive and emotional dimensions: in DuRecDial, agents improve Comp-Î while decreasing E-Î , reflecting simultaneous preference refinement and emotional tension. Similarly, persuasive settings (P4G) exhibit a measurable emotional cost, where most agents induce negative E-Î despite achieving successful outcomes. These findings highlight the limitations of single-score evaluation and motivate dimension-wise trajectory tracking. This divergence further justifies the task-specific weighting design of Compt. Without such weighting, DuRecDialâs cognitive gains would be obscured by emotional declines, while ESConvâs emotion-driven dynamics would be underrepresented due to relatively flat cognitive signals. By assigning wE=0.50w_E=0.50 in ESConv and wE=0.10w_E=0.10 in other tasks, the framework prioritizes the most informative dimension for each scenario. CTS further indicates structured heterogeneity across agents, consistent with the typology observed in experiments (all-round, sprint, steady, hollow). The CTS-GO divergence demonstrates that cognitive trajectory improvement does not necessarily align with task success. This mismatch may explain why different agents optimize different aspects of dialogue behavior, and motivates joint evaluation of process-level and outcome-level metrics. 8 Limitations and Future Work Annotation quality. Current annotations are LLM-generated. Human validation shows BDI recall 0.79, precision 0.77, and emotion Îș=0.69Îș=0.69, but subtle states such as implicit beliefs and unexpressed desires may be missed (§4.2). Scaling human validation and exploring LLMâhuman collaborative annotation are important next steps. From evaluation to training. Deploying CogWM as an online RL environment where BDI/E scores serve as rewards would enable agents to learn genuine cognitive improvement through interaction. Broader impacts are discussed in §A. 9 Conclusion This paper proposed CogWM, an LLM-based cognitive world model for evaluating social influence agents. CogWM introduces a three-tier BDI/E evaluation framework, a SaA annotation pipeline covering 150K samples across four scenarios, and joint training on Qwen3-14B achieving 77.6% emotion accuracy (2.1Ă over GPT-5.5). In 3,600 cross-agent trials, CogWM distinguished six commercial agents by cognitive influence (Llama-4-Scout first, CTS +0.233). The CTS-GO complementarity shows that cognitive improvement and behavioral outcomes are not equivalent. References A. Belz, S. Mille, and D. M. Howcroft (2020) Comparing human and automatic metrics: a meta-evaluation of evaluation methods for NLG. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), Online, p. 1â15. Cited by: §1. R. H. Bordini, J. F. HĂŒbner, and M. Wooldridge (2007) Programming multi-agent systems in AgentSpeak using Jason. Wiley Series in Agent Technology, John Wiley & Sons. Cited by: §B.2. M. E. Bratman (1987) Intention, plans, and practical reason. Harvard University Press, Cambridge, MA. Cited by: §B.2, §2. S. Chang, A. Anderson, and J. M. Hofman (2025) ChatBench: from static benchmarks to human-AI evaluation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), p. 26009â26038. Cited by: §B.3, §1. K. Chawla, W. Shi, J. Zhang, G. M. Lucas, Z. Yu, and J. Gratch (2023) Social influence dialogue systems: a survey of datasets and models for social influence tasks. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), Dubrovnik, Croatia, p. 750â766. Cited by: §B.1, §B.3, §1, §2. H. Chen, X. Liu, D. Yin, and J. Tang (2017) A survey on dialogue systems: recent advances and new frontiers. ACM SIGKDD Explorations Newsletter 19 (2), p. 25â35. Cited by: §1. W. Chiang, L. Zheng, Y. Sheng, A. N. Angelopoulos, T. Li, D. Li, H. Zhang, B. Zhu, M. I. Jordan, J. E. Gonzalez, and I. Stoica (2024) Chatbot arena: an open platform for evaluating LLMs by human preference. In Proceedings of the 41st International Conference on Machine Learning (ICML), Vienna, Austria, p. 1â29. Cited by: §B.4. Y. Dubois, B. Galambosi, P. Liang, and T. B. Hashimoto (2024) Length-controlled AlpacaEval: a simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475. Cited by: §B.4, §2. J. Gao, M. Galley, and L. Li (2019) Neural approaches to conversational AI. Foundations and Trends in Information Retrieval 13 (2-3), p. 127â298. Cited by: §1. M. P. Georgeff and A. L. Lansky (1987) Reactive reasoning and planning. In Proceedings of the Sixth National Conference on Artificial Intelligence (AAAI), Seattle, WA, p. 677â682. Cited by: §B.2. E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen (2022) LoRA: low-rank adaptation of large language models. In Proceedings of the Tenth International Conference on Learning Representations (ICLR), p. 1â21. Cited by: Appendix E, §4.3. A. Jafari et al. (2025) Beyond words: integrating theory of mind into conversational agents for human-like belief, desire, and intention alignment. In Findings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), Cited by: §B.2. C. Kong, Y. Fan, X. Wan, F. Jiang, and B. Wang (2024) PlatoLM: teaching LLMs in multi-round dialogue via a user simulator. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), Bangkok, Thailand, p. 7841â7863. Cited by: §B.3, §1, §2. F. Kreyssig, I. Casanueva, P. Budzianowski, and M. GaĆĄiÄ (2018) Neural user simulation for corpus-based policy optimisation of spoken dialogue systems. In Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue (SIGdial), Melbourne, Australia, p. 60â69. Cited by: §B.3. M. Lewis, D. Yarats, Y. N. Dauphin, D. Parikh, and D. Batra (2017) Deal or no deal? end-to-end learning for negotiation dialogues. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Cited by: §1. J. Li, Z. Lin, P. Fu, and W. Wang (2021) Past, present, and future: conversational emotion recognition through structural modeling of psychological knowledge. In Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Cited by: §B.2, §2. Y. Li, H. Su, X. Shen, W. Li, Z. Cao, and S. Niu (2017) DailyDialog: a manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (IJCNLP), Taipei, Taiwan, p. 986â995. Cited by: §B.1, §4. C. Lin (2004) ROUGE: a package for automatic evaluation of summaries. In Proceedings of the ACL-04 Workshop: Text Summarization Branches Out, Barcelona, Spain, p. 74â81. Cited by: §B.1, §1. S. Liu, C. Zheng, O. Demasi, S. Sabour, Y. Li, Z. Yu, Y. Jiang, and M. Huang (2021) Towards emotional support dialog systems. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), Online, p. 3469â3483. Cited by: §B.1, §4. Z. Liu, H. Wang, Z. Niu, H. Wu, W. Che, and T. Liu (2020) Towards conversational recommendation over multi-type dialogs. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), Online, p. 1036â1049. Cited by: §B.1, §4. X. Luo, Z. Tang, J. Wang, and X. Zhang (2024) DuetSim: building user simulator with dual large language models for task-oriented dialogues. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING), Torino, Italy, p. 5414â5424. Cited by: §B.3, §1, §2, §6.1. M. Ma, B. Guo, R. Yang, M. Chen, Y. Liu, J. Liu, Y. Pei, X. Ma, Q. Zhang, and Z. Yu (2026) Think thrice before you speak: dual knowledge-enhanced theory-of-mind reasoning for persuasive agents. arXiv preprint arXiv:2605.22602, p. 1â19. Cited by: §B.2, §2. Y. Mehta, N. Majumder, A. Gelbukh, and E. Cambria (2020) Recent trends in deep learning based personality detection. Artificial Intelligence Review 53, p. 2313â2339. Cited by: §4.1. K. Papineni, S. Roukos, T. Ward, and W. Zhu (2002) BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), Philadelphia, PA, p. 311â318. Cited by: §B.1, §1. S. Qiu, M. Liu, H. Li, S. Zhu, and Z. Zheng (2023) MindDial: belief dynamics tracking with theory-of-mind modeling for situated neural dialogue generation. arXiv preprint arXiv:2306.15253. Cited by: §B.2, §2. A. S. Rao and M. P. Georgeff (1991) Modeling rational agents within a BDI-architecture. In Proceedings of the Second International Conference on Principles of Knowledge Representation and Reasoning (KR), Cambridge, MA, p. 473â484. Cited by: §B.2, §2. J. Schatzmann, B. Thomson, and S. Young (2007) Statistical user simulation with a hidden agenda. In Proceedings of the 8th SIGdial Workshop on Discourse and Dialogue (SIGdial), Antwerp, Belgium, p. 273â282. Cited by: §B.3. W. Shi, K. Qian, X. Wang, and Z. Yu (2019) How to build user simulators to train RL-based dialog systems. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, p. 1990â2000. Cited by: §B.3. K. Wang, X. Li, S. Yang, L. Zhou, F. Jiang, and H. Li (2025) Know you first and be you better: modeling human-like user simulators via implicit profiles. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), p. 1â22. Cited by: §B.3, §2, §4.1. X. Wang, W. Shi, R. Kim, Y. Oh, S. Yang, J. Zhang, and Z. Yu (2019) Persuasion for good: towards a personalized persuasive dialogue system for social good. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), Florence, Italy, p. 5635â5649. Cited by: §B.1, §4. S. Wu et al. (2025) HumanLM: simulating users with state alignment beats response imitation. arXiv preprint arXiv:2502.00640. Note: ICML 2025 Cited by: §B.3, §2. W. Xu et al. (2024) Beyond linguistic cues: fine-grained conversational emotion recognition via belief-desire modelling. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING), Cited by: §B.2, §2. A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, C. Zheng, D. Liu, F. Zhou, F. Huang, F. Hu, H. Ge, H. Wei, H. Lin, J. Tang, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Zhou, J. Lin, K. Dang, K. Bao, K. Yang, L. Yu, L. Deng, M. Li, M. Xue, M. Li, P. Zhang, P. Wang, Q. Zhu, R. Men, R. Gao, S. Liu, S. Luo, T. Li, T. Tang, W. Yin, X. Ren, X. Wang, X. Zhang, X. Ren, Y. Fan, Y. Su, Y. Zhang, Y. Zhang, Y. Wan, Y. Liu, Z. Wang, Z. Cui, Z. Zhang, Z. Zhou, and Z. Qiu (2025a) Qwen3 technical report. arXiv preprint arXiv:2505.09388. Cited by: Appendix E, §4.3. B. Yang, J. Guo, Y. Iwasawa, and Y. Matsuo (2025b) ToM-agent: large language models as theory of mind aware generative agents with counterfactual reflection. arXiv preprint arXiv:2501.15355. Cited by: §B.2. S. Young, M. GaĆĄiÄ, B. Thomson, and J. D. Williams (2013) POMDP-based statistical spoken dialog systems: a review. Proceedings of the IEEE 101 (5), p. 1160â1179. Cited by: §B.3. S. Zhang et al. (2026) EMPA: evaluating persona-aligned empathy as a process. arXiv preprint arXiv:2603.00552. Cited by: §B.4, §2. W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang, Z. Dong, Y. Du, C. Yang, Y. Chen, Z. Chen, J. Jiang, R. Ren, Y. Li, X. Tang, Z. Liu, P. Liu, J. Nie, and J. Wen (2023) A survey of large language models. arXiv preprint arXiv:2303.18223. Cited by: §1. L. Zheng, W. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica (2023) Judging LLM-as-a-judge with MT-bench and chatbot arena. In Advances in Neural Information Processing Systems (NeurIPS), p. 1â29. Cited by: §B.4, §1, §2. Appendix A Broader Impact A.1 Generalizability of Cognitive Modeling The BDI/E framework and SaA annotation method can be generalized to any dialogue scenario where understanding why a person speaks or acts a certain way is needed. This includes tracking patientsâ health beliefs and help-seeking intentions in medical consultations, studentsâ knowledge beliefs and learning motivation in educational tutoring, and negotiatorsâ bottom lines and willingness to compromise. A.2 Responsible AI and Transparency CogWM makes agentsâ social influence visible and auditable through BDI/E state trajectories. If a userâs donation commitment score jumps from 0 to 2 at the end of a conversation while their impact belief score remains at 0, it suggests the user was pressured rather than genuinely persuaded. This makes interpretability a concrete tool for preventing manipulative AI. A.3 Cognitive World Model as Social AI Infrastructure We envision the cognitive world model serving as a testbed for socially influential agents, analogous to wind tunnels for aircraft design or crash tests for automotive safety. Before deployment, any AI system aimed at social influence can be evaluated on CogWM, observing what cognitive impact it produces on users with different personalities. Appendix B Extended Related Work B.1 Social Influence Dialogue Chawla et al. (Chawla et al. 2023) were the first to systematically define social influence dialogue as conversational systems that change user cognition, emotion, and behavior through natural dialogue, covering seven subfields including persuasion, negotiation, and emotional support. They identified the core pain point: evaluation fragmentation. Existing work only covers language quality (BLEU/ROUGE (Papineni et al. 2002; Lin 2004)) and influence outcomes (donation amount (Wang et al. 2019), survey scores (Liu et al. 2021)), but completely lacks measurement of cognitive state changes during the dialogue process. CogWM covers emotional support (Liu et al. 2021), persuasion (Wang et al. 2019), recommendation (Liu et al. 2020), and daily conversation (Li et al. 2017), supplying the missing dimension of âuser cognitive state trajectoriesâ for social influence dialogue. B.2 BDI Cognitive Modeling The BDI model originates from Bratmanâs (Bratman 1987) philosophy of action and was formalized as BDI logic by Rao and Georgeff (Rao and Georgeff 1991). Classical engineering implementations (PRS (Georgeff and Lansky 1987), JASON (Bordini et al. 2007)) require hand-coded logical propositions and cannot automatically acquire states from dialogue. In the deep learning era, BDI has been transformed into trainable neural representations: Li et al. (Li et al. 2021) encoded belief-desire theory as intermediate representations for dialogue emotion recognition, and Xu et al. (Xu and others 2024) proposed explicit belief-desire graph joint modeling. The LLM era has further driven automation: MindDial (Qiu et al. 2023) tracks three-layer belief dynamics, ToM-agent (Yang et al. 2025b) updates BDI confidence through counterfactual reflection, and Jafari et al. (Jafari and others 2025) validated the dialogue effectiveness of ToM-informed alignment. These works share two limitations: (1) limited granularity, as most model only a single dimension (belief or emotion), lacking synchronized BDI/E four-dimensional tracking; (2) one-sided inference, i.e., they infer states from dialogue without jointly generating utterances. The most directly related work is Ma et al.âs (Ma et al. 2026) TTBYS, which first achieved serialized BDI prediction in persuasion dialogues based on the belief â desire â intention causal order. CogWM extends this by advancing one-sided state inference to state-utterance joint generation, covering four social influence scenarios. B.3 User Simulation User simulators have undergone three generations of evolution (Chawla et al. 2023). The first generation was agenda-driven (Schatzmann et al. (Schatzmann et al. 2007)), relying on handcrafted state spaces that are non-transferable across domains (Young et al. 2013). The second generation used neural Seq2Seq models (Kreyssig et al. (Kreyssig et al. 2018), Shi et al. (Shi et al. 2019)), learning end-to-end from corpora but optimizing only text likelihood or BLEU. The third generation is LLM-driven (DuetSim (Luo et al. 2024), ChatBench (Chang et al. 2025), PlatoLM (Kong et al. 2024)), offering strong generation ability but only predicting user utterances. Methods from 2025 began touching on state modeling: USP (Wang et al. 2025) extracts implicit profiles, and HumanLM (Wu and others 2025) proposes that âstate alignment is superior to response imitation,â but their âstatesâ are all implicit vectors that are neither interpretable nor transferable across tasks. The common blind spot across three generations of simulators: they only predict what the user says, (H<t)âut(H_<t)â u_t, without modeling the internal cognitive states that drive utterances. CogWM extends simulation to (H<t,Stâ1)â(St,ut)(H_<t,S_t-1)â(S_t,u_t), making BDI/E an explicit, auditable condition for both generation and output. B.4 Dialogue Evaluation: From Terminal Judgment to Process Tracking LLM-as-judge approaches, such as MT-Bench (Zheng et al. 2023), AlpacaEval (Dubois et al. 2024), and Chatbot Arena (Chiang et al. 2024), have achieved scalable semantic quality scoring but remain fundamentally terminal black-box judgments. Scenario-level metrics (survey scores, donation amounts) are sparse, expensive, and do not explain âwhy.â The work closest to CogWMâs evaluation philosophy is EMPA (Zhang and others 2026), which first advanced empathy evaluation from single-turn scoring to process tracking, measuring directional alignment and cumulative influence in an implicit psychological space. CogWM builds on this by moving evaluation from an implicit psychological space to an explicit BDI/E cognitive state space, where each turnâs changes can be traced to specific B/D/I/E dimensions, covering four scenarios. Appendix C BDI/E Annotation Details C.1 Emotion labels Table 7: Emotion Lexicon and Valence Values Emotion Valence Emotion Valence hope +2.0 relief +1.5 confusion â-1.0 frustration â-1.5 guilt â-1.5 anger â-1.5 loneliness â-1.5 sadness â-2.0 anxiety â-2.0 fear â-2.0 helplessness â-2.0 shame â-2.5 Table 7 presents the classification space of emotion labels and their corresponding valence values. Each emotion label e is assigned a valence score vev^e. During annotation, the model is asked to assign one primary emotion label to each user utterance utu_t and rate its intensity, where the intensity score seâ1,2,3s^eâ\1,2,3\, with 1 indicating weak emotion, 2 indicating moderate intensity, and 3 indicating strong emotion. In the computation of EÎ , the quantified emotion score is et=veâ see_t=v^e· s^e. C.2 Sample Construction Details Statistics of the original samples across four datasets are shown in Table 8: Table 8: Dataset Statistics Dataset #Samples Turns Role Task DailyDialog 13,118 4.1 usr D ESConv 1,496 15.2 sec ESC P4G 1,017 10.2 pers PD DuRecDial 8,241 7.7 usr CR Total 23,872 â â 4 Based on the four datasets, we constructed turn-level training samples for CogWM, with each sample structured as follows: âą Input: profile + task context + initial state + dialogue history (H<tH_<t) + optional previous state (Stâ1S_t-1) âą Output: emotion + beliefs + desires + intentions + user utterance (St+utS_t+u_t) We constructed a total of 150,454 user-turn-level samples, with detailed statistics shown in Table 9. Table 9: Data Splits across Datasets Dataset Train Val Test Total DailyDialog 48,373 2,688 2,730 53,791 ESConv 20,457 1,151 1,174 22,782 P4G 9,282 516 534 10,332 DuRecDial 57,162 3,262 3,125 63,549 Total 135,274 7,617 7,563 150,454 Appendix D Three-layer Evaluation Framework The core workflow of CogWM as an evaluation platform is: the assistant agent under evaluation is connected to the CogWM world model, which simulates the userâs BDI/E cognitive state changes and natural language responses turn by turn, thereby transforming the agentâs social impact capability into quantifiable, decomposable, and interpretable systematic evaluation. Based on this, we construct a bottom-up three-layer evaluation framework: starting from atomic observations of single-turn response quality (Turn layer), through multi-turn state evolution modeling (Trajectory layer), ultimately converging to a global impact composite score (Task layer). D.1 Turn Layer: Single-turn Response Quality Evaluation The Turn layer defines atomic observations of CogWMâs single-turn output, i.e., CogWM simultaneously predicts user utterance UtU_t and BDI/E internal states St=Et,Bt,Dt,ItS_t=\E_t,B_t,D_t,I_t\ at each turn t. Evaluation at the Turn layer revolves around two core questions: (a) Is the generated response close to a real user? (b) Are the predicted BDI/E states accurate? Both are measured through automatic metrics and LLM-as-Judge semantic metrics, respectively. Automatic Evaluation Metrics As shown in Table 10, we use 8 automatic metrics, grouped into semantic accuracy and quantitative precision. Table 10: Turn-level automatic evaluation metrics Cat. Metric Definition Opt. Sem. E-Acc Emotion label accuracy â B-Sim Belief cosine similarity â D-Sim Desire semantic similarity â I-Sim Intention semantic similarity â Quant. E-MAE Emotion intensity MAE â B-MAE Belief score MAE â D-MAE Desire score MAE â I-MAE Intention score MAE â LLM-as-Judge Semantic Quality Metrics Automatic metrics capture only lexical overlap and embedding similarity; they cannot assess semantic-level matching quality. We introduce 5 LLM-as-Judge metrics, using DeepSeek-V4-Flash to perform semantic evaluation of the prediction quality at each turn on a 1â5 scale, as shown in Table 11. Table 11: Turn-level LLM-as-Judge semantic evaluation metrics Metric Evaluation Content LLMJ-Emo Measures consistency between predicted emotion labels and the emotional expression in user utterances LLMJ-Bel Measures alignment between predicted belief states and user beliefs inferred from dialogue context LLMJ-Des Measures consistency between predicted desires and user intentions expressed in context LLMJ-Int Measures alignment between predicted intentions and user behavioral intent in dialogue context LLMJ-Coh Measures internal consistency between predicted BDI/E states and generated user utterances The 13 metrics (8 automatic + 5 LLMJ) at the Turn layer together constitute a complete characterization of CogWM simulation fidelity, serving as the reliability foundation for higher-level trajectory analysis and task evaluation. Table 12: Trajectory-level dynamic metrics adapted from financial time-series analysis Category Metric Formula Cumulative return AUC 1Nâât=1Nâ1st+st+12 1N _t=1^N-1 s_t+s_t+12 Endpoint shift Î stendâststarts_t_end-s_t_start Improvement frequency PSR 1Nâ1âât=1Nâ1â[st+1>st] 1N-1 _t=1^N-11[s_t+1>s_t] Directional consistency Mono 1Nâ1âât=1Nâ1â[signâ(st+1âst)=signâ(Î)] 1N-1 _t=1^N-11[sign(s_t+1-s_t)=sign( )] D.2 Trajectory Layer: BDI/E State Evolution Analysis The Trajectory layer aggregates turn-by-turn atomic observations from the Turn layer into cross-turn dynamic analysis, and is the core innovation of the three-layer framework. This layer draws on concepts from financial time-series analysis such as trends, volatility, and cumulative returns to systematically model BDI/E state trajectories. Compt: Task-Weighted Per-Turn Composite Scalar First, the four BDI/E dimensions output by CogWM at the Turn layer are compressed into a single scalar sequence stt=1N\s_t\_t=1^N for use in all aggregation metrics at the Trajectory layer. sts_t is defined as the task-weighted Compt: stâĄCompt=wEâ Eâ_âvâaâlt+wBâ Bt+wDâ Dt+wIâ Its_t _t=w_E· E\_val_t+w_B· B_t+w_D· D_t+w_I· I_t (2) where Eâ_âvâaâltE\_val_t is the pure emotion valence (without intensity scaling, in the same dimension as B/D/I at [â2.5,+2.0][-2.5,+2.0]), and Bt,Dt,ItB_t,D_t,I_t are the mean task-specific scores for each dimension at that turn (range [0,2][0,2]). Task weights are set according to the scenario. For the ESC (Emotional Support) task, emotional improvement is the core objective, so the emotion dimension E receives the highest weight, while B/D/IB/D/I dimensions receive lower balanced weights. For the PD (Persuasion) task, the main focus is on cognitive and attitude change, so the weights of B, D, and I are increased, while the emotion dimension serves only as an auxiliary signal. For the CR (Conversational Recommendation) task, the emphasis is on preference matching and decision consistency; its weight distribution is similar to PD, placing greater emphasis on cognitive and intention-level modeling capability. The specific weights are shown in Table 13. Table 13: Task-specific weighting configuration for ComptComp_t Task wEw_E wBw_B wDw_D wIw_I ESC 12 12 16 16 16 16 16 16 PD 110 110 310 310 310 310 310 310 CR 110 110 310 310 310 310 310 310 Drawing on financial time-series analysis methods, we design four trajectory-level metrics using Comp as the scalar basis. Area Under Curve (AUC) measures the cumulative performance of the Comp trajectory, approximating the overall âexperience gainâ through trapezoidal integration. Net Change (Î ) characterizes the net change based on substantive endpoints, reflecting the final improvement magnitude of the overall intervention effect. Positive Step Ratio (PSR) represents the proportion of turn-by-turn positive improvements, measuring the frequency and stability of state improvement. Monotonicity (Mono) characterizes the consistency of improvement direction, i.e., the degree of alignment with the sign of the global trend Î , measuring the monotonicity and stability of trajectory evolution. The specific formulas are shown in Table 12. D.3 Task Layer: Global Impact Comprehensive Evaluation The Task layer converges the dynamic analysis from the Trajectory layer into a task-objective-oriented composite score, comprising two complementary metrics: the trajectory-level cognitive improvement indicator CTS and the dialogue-level behavioral outcome indicator GoodOutcome (GO). CTS: Composite Trajectory Score CTS is a weighted composite of the three components from the Trajectory layer, designed on the principle that: (a) information among components is non-redundant; (b) they cover different aspects of trajectory quality (cumulative experience, improvement frequency, net change). CTS=0.40â AUC+0.30â PSR+0.30â Î CTS=0.40·AUC+0.30·PSR+0.30· (3) To verify the non-redundancy of information among CTS components, we computed the Pearson correlation matrix of 5 candidate trajectory metrics from the Comp sequence across all 3,600 interaction trajectories, shown in Fig. 6. Figure 6: Pearson correlation heatmap of trajectory-level evaluation metrics The heatmap reveals two clear patterns. First, the three CTS components are weakly correlated: AUCâÎ r=â0.06r=-0.06, AUCâPSR r=0.09r=0.09, PSRâÎ r=0.25r=0.25. No pair exceeds 0.26, confirming they capture distinct dimensions of trajectory quality. AUC and Î are nearly orthogonal, while PSR provides the most independent signal. Second, PSR and Mono are substantially correlated (r=0.77r=0.77): both measure improvement consistency. PSR uses a simpler binary criterion and has cleaner semantics, so we retain PSR and discard Mono. PSR and nz% are also correlated (r=0.74r=0.74) because content-rich dialogues tend to have more positive steps. GoodOutcome: Dialogue-Level Behavioral Outcome CTS measures trajectory-level cognitive improvement, while GoodOutcome (GO) evaluates through an LLM Judge whether the dialogue endpoint achieves task objective behavioral outcomes. The GO criteria are relaxed by scenario: in ESConv, emotional improvement or having an action plan counts as positive; in P4G, making a donation decision counts as positive; in DuRecDial, accepting a recommendation counts as positive. Table 14: Human Validation of BDI/E Annotation Quality Dimension Metric Value Threshold Global BDI Coverage Recall 0.79 >> 0.70 Global BDI Coverage Precision 0.77 >> 0.70 Emotion Label Agreement Cohenâs Îș 0.69 >> 0.60 Emotion Intensity Correlation Spearman Ï 0.80 >> 0.70 BDI Score Correlation Spearman Ï 0.67 >> 0.60 Turn-level BDI Allocation F1 0.73 >> 0.65 Appendix E Training Details We fine-tune using LoRA (Low-Rank Adaptation) (Hu et al. 2022) based on Qwen3-14B Instruct (Yang et al. 2025a) for parameter-efficient fine-tuning, reducing computational cost while preserving the base modelâs general capabilities. All three backbone scales (8B/14B/32B) share the same hyperparameters. Table 15: Table A6: Training Hyperparameters Parameter Value LoRA rank 16 LoRA alpha 32 cutoff_len 4096 effective batch_size 16 learning rate 5e-5 steps 16,910 hardware Single A100-80G Appendix F Supplementary Experimental Results F.1 Human Validation of SaA Annotation Quality To verify the reliability of the SaA two-stage annotation pipeline, we drew a stratified sample of 300 dialogues from 150,454 user-turn samples and invited 6 human annotators with BDI annotation experience to participate in the evaluation. Each annotator independently evaluated 50 samples, verifying the quality of LLM-generated BDI/E annotations. The validation examines three aspects across the two stages of SaA: global BDI coverage, emotion precision, and turn-level allocation quality, as shown in Table 14. Global BDI Coverage. For the global BDI item extraction in Stage 1 (Summarize), we compute agreement metrics between human annotations and LLM annotations. Results show a global BDI recall of 0.79 (>0.70) and precision of 0.77 (>0.70). These results indicate that the LLM covers BDI items recognized by humans reasonably well, while still exhibiting some over-decomposition in fine-grained splitting. Emotion Precision. For the turn-by-turn emotion annotations in Stage 2 (Allocate), across 12 emotion label categories we obtain Cohenâs Îș=0.69Îș=0.69 (>0.60), Spearman Ï=0.80Ï=0.80 (>0.70) for emotion intensity (0-3), and Spearman Ï=0.67Ï=0.67 (>0.60) for BDI quantitative scores. The model achieves relatively high consistency in emotion trend modeling, though boundary ambiguity persists between similar emotion categories like confusion and frustration. Turn-level Allocation. For the global_ref referencing mechanism, the turn-level BDI allocation F1 is 0.73 (>0.65). Errors mainly stem from BDI lifecycle modeling biases, including persistent referencing of already-resolved beliefs and repeated modeling of recurring beliefs. All metrics exceed preset thresholds, confirming the reliability of the SaA two-stage annotation method for large-scale dialogue BDI/E annotation tasks. F.2 Fidelity Experiments Table 16 presents supplementary results from the fidelity experiments. CogWM performs comparably to Sim on B/D/I-MAE, with both achieving the best results. This indicates that CogWM not only tracks the semantic content of BDI but also matches the score level precisely. B-1 scores remain the best and consistent with B-4, while Dist-1 approaches that of high-performance LLMs, demonstrating textual diversity. Table 16: Supplementary Results for Fidelity Experiments Model B-Mâ D-Mâ I-Mâ B-1â Dist-1â GPT-5.5 0.81 0.79 0.27 0.133 0.962 DSv4-Pro 0.79 0.66 0.32 0.172 0.965 Dual-LLM 0.81 0.59 0.28 0.158 0.960 Qwen3-14B 0.93 0.80 0.51 0.112 0.953 !8 CogWM-14B 0.39 0.26 0.14 0.322 0.948 F.3 Ablation Experiments To quantify the contribution of each component in the user profile, we ran comparative experiments on 5 ablation variants based on CogWM-14B, modifying only the inference input. The full configuration (Full) uses task-related user profiles, Big Five personality traits, the previous turnâs BDI/E state, and the BDI/E supervised training objective. Then we remove different information sources to examine their independent contributions. w/o Task Profile removes task-related user profiles, retaining only personality traits, historical states, and the BDI/E training objective. w/o Big Five removes Big Five personality traits to examine the role of stable personality traits in user state prediction. w/o Prev State removes the previous turnâs BDI/E state, evaluating the contribution of historical psychological state to multi-turn state transition modeling. w/o All Conditions removes task profiles, personality traits, and historical states simultaneously, retaining only the BDI/E supervised training objective to examine the effectiveness of structured psychological state annotation alone. Table F.4 presents the results of the ablation experiments. Across the three evaluation dimensions, we find the following: Single profile component contributions are limited, and information is distributedly encoded. Removing any single component among Big5, Comm/Dec, or Task results in E-Acc changes below 2p (72.8%â74.2%), B-Sim changes below 0.02, and BLEU-4 changes below 0.01. No single field is critical; profile information is distributedly encoded across multiple dimensions, and any single dimension can be compensated for by others. The overall profile contributes far more to generation quality than to state prediction. Removing all profiles and task context simultaneously (â-All) causes E-Acc to drop from 72.8% to 60.7% (â-12p), while BLEU-4 drops from 0.145 to 0.040 (â-72%). Removing only all profiles while retaining task context (â-Profile) reduces E-Acc by only 0.3p (72.5%), but BLEU-4 drops from 0.145 to 0.140. Task context alone cannot replace the core role of user profiles in driving personalized utterance generation. LLM Judge provides stronger ablation signals than automatic metrics. Under the â-All condition, LLMJ-Belief drops from 3.25 to 2.91 (â-10.5%), while automatic B-Sim drops only 6.4% (0.761â 0.712). LLM Judge is more sensitive to semantic quality degradation than embedding similarity, confirming the complementary value of LLMJ as an auxiliary evaluation. Single-component removal LLMJ differences are all below 0.2, while â-All shows consistent degradation across all LLMJ dimensions, consistent with the pattern in automatic metrics. Notably, â-Comm/Dec slightly exceeds Full on LLMJ-Emotion (4.11 vs 3.98), possibly because removing communication style descriptions makes the model tend toward more conservative, generic emotion predictions. F.4 Task Differences CogWMâs performance on the four datasets reveals systematic differences in scenario difficulty: Table 17: CogWM 14B Breakdown by Four Datasets Dataset E-Acc B-Sim D-Sim I-Sim B-4 LenR DailyDialog 59.4% .710 .701 .472 .171 .410 ESConv 59.3% .606 .565 .398 .047 .249 P4G 83.3% .769 .716 .744 .127 .342 DuRecDial 90.1% .867 .767 .465 .222 .524 Table 18: CogWM model scale comparison (8B vs. 14B) Model E-Acc B-Sim D-Sim I-Sim B-4 R-L CogWM-8B 75.8 0.748 0.681 0.442 0.141 0.401 CogWM-14B 77.6 0.773 0.705 0.458 0.150 0.417 CogWM-32B â â â â â â Table 19: Ablation study results. Model BDI/E Core Utterance LLM Judge E-Aâ B-Sâ D-Sâ I-Sâ E-Mâ R-Lâ B-4â LenR Emoâ Belâ Desâ Intâ Uttâ Cohâ â Big5 73.5 .761 .683 .539 3.94 .412 .150 1.01 3.94 3.27 3.41 4.41 2.95 4.08 â Comm/Dec 73.6 .757 .675 .533 4.11 .414 .150 1.01 4.11 3.26 3.51 4.53 3.01 4.20 â Task 72.2 .760 .676 .518 3.79 .407 .145 0.99 3.89 3.15 3.52 4.39 2.90 4.01 â Profile 72.5 .748 .672 .506 3.89 .404 .140 1.01 3.89 3.19 3.66 4.49 2.92 4.13 !8 â All 60.7 .712 .591 .399 3.83 .223 .077 1.04 3.83 2.91 3.17 4.28 2.06 3.69 !8 Full 72.8 .761 .690 .554 3.98 .417 .150 1.01 3.98 3.25 3.61 4.51 3.03 4.15 DuRecDial is the easiest (E-Acc 90.1%, fixed structure, predictable patterns), ESConv and DailyDialog are the hardest (E-Acc approximately 59%, open-domain with subtle emotions, strong personalization), and P4G falls in between (83.3%). The E-Acc range across the four datasets spans 31p (59%â90%), far exceeding the differences between models. This suggests that scenario attributes, rather than model capability, are the primary determinant of performance. This trend holds across API baselines: GPT-5.5 achieves 41.6% on DuRecDial but only 36.6% on ESConv, and Qwen3-14B base ranges from 18.0% to 34.5% across datasets. DuRecDialâs high scores stem mainly from its highly structured dialogue format (user expresses preference â agent recommends â user accepts/rejects) with fixed BDI/E state change patterns. ESConv involves complex emotional evolution (frustrationâ â ), and DailyDialog covers diverse topics with short turns (average only 4.1 turns). Both pose the greatest challenge for cognitive state modeling. The complete breakdown by task is presented in §LABEL:app:figures (Fig 7: 18 subplot trajectories across all models). F.5 Per-Task BDI/E Trajectory Statistics Table 20 reports per-task trajectory statistics for all 6 agents under both No-CoT and CoT conditions. The decomposition reveals two patterns: (1) in DuRecDial, Comp_Î is positive for all agents while E_Î is negativeâa cognitive-emotional divergence invisible to single-score evaluation; (2) in P4G, E_Î is negative for four of six agents, quantifying the emotional cost of persuasive dialogue. Table 20: Per-task BDI/E trajectory statistics for all agents and conditions. Agent Cond ESConv P4G DuRecDial C_Î E_Î B_Î D_Î I_Î C_Î E_Î B_Î D_Î I_Î C_Î E_Î B_Î D_Î I_Î GPT-5.5 No-CoT +0.840 +2.375 +0.143 +0.123 +0.390 +0.116 -0.200 +0.068 +0.349 +0.010 +0.090 -1.080 +0.177 +0.252 +0.120 !8 GPT-5.5 CoT +0.875 +2.310 +0.285 +0.202 +0.290 +0.135 -0.610 +0.120 +0.357 +0.060 +0.205 -0.560 +0.186 +0.437 +0.220 Claude No-CoT +0.689 +1.675 +0.122 +0.173 +0.330 +0.272 -0.190 +0.252 +0.488 +0.140 +0.185 -0.740 +0.212 +0.533 +0.060 !8 Claude CoT +0.888 +2.540 +0.145 +0.107 +0.590 +0.144 -0.940 +0.203 +0.378 +0.080 +0.142 -0.670 +0.136 +0.430 +0.080 Gemini No-CoT +0.616 +1.985 +0.042 +0.042 +0.200 +0.049 +0.000 +0.070 +0.073 +0.000 +0.170 -0.130 +0.199 +0.120 +0.260 !8 Gemini CoT +0.714 +2.070 +0.091 +0.240 +0.200 +0.052 -0.150 +0.103 +0.127 +0.000 +0.111 -0.560 +0.135 +0.217 +0.140 Llama No-CoT +0.845 +2.555 +0.026 +0.000 +0.480 +0.271 +0.135 +0.267 +0.309 +0.250 +0.173 -0.780 +0.182 +0.523 +0.080 !8 Llama CoT +0.905 +2.830 +0.154 +0.140 +0.433 +0.197 -0.170 +0.235 +0.387 +0.120 +0.221 -0.320 +0.260 +0.438 +0.160 Qwen No-CoT +0.771 +2.065 +0.130 +0.308 +0.500 +0.210 -0.860 +0.233 +0.463 +0.200 +0.242 -0.610 +0.208 +0.455 +0.360 !8 Qwen CoT +1.016 +3.100 +0.197 +0.363 +0.383 +0.203 -0.350 +0.288 +0.319 +0.120 +0.198 -1.270 +0.172 +0.525 +0.280 DeepSeek No-CoT +0.850 +2.200 +0.365 +0.203 +0.180 +0.148 -0.220 +0.178 +0.248 +0.100 +0.147 -0.860 +0.140 +0.427 +0.100 !8 DeepSeek CoT +0.739 +2.025 +0.176 +0.343 +0.420 +0.207 -0.440 +0.262 +0.435 +0.080 +0.146 -0.770 +0.107 +0.403 +0.200 F.6 Model Scale Comparison (8B vs. 14B) Table 18 compares CogWM-8B and CogWM-14B on key metrics. The 8B model approaches 14B across all metrics, confirming that BDI/E joint training is not tied to model scale. Results for 32B are pending. Appendix G Agent Discrimination Complete Cases The following six sets of complete multi-turn dialogues demonstrate CogWMâs ability, as an evaluation platform, to trace turn-by-turn BDI/E trajectories for the best and worst agent interactions. Each case includes the complete agentâuser interaction, turn-by-turn BDI/E annotations (B/D/I represent Belief/Desire/Intention respectively, with scores being normalized values), and cognitive transition arc analysis. Dialogue format: A = Agent response, U = CogWM-simulated user response. G.1 Case C1: GPT-5.5 Ă ESConv (Best, Î=+2.00 =+2.00) Table 21: GPT-5.5 ESConv Full Dialogue Trajectory (Î=+2.00 =+2.00) T Dialogue Content (gray=Agent, white=User) Cold-Start Context (T1âT2) !5 1 User: Hey there. How are you? !5 1 Asst: I am fine. How is your side? !5 2 User: I am ok, Iâm having a hard time dealing with the pandemic though. !15 1 Agent: That sounds really exhausting and isolating⊠This week, try choosing one task that feels almost too easy User [confusion]: I just feel unmotivated and frustrated with the situation. !15 2 Agent: That sounds really draining⊠frustration can be a signal that you need something small and doable User [confusion]: Yeah, I guess so. I just donât know how to get motivated again. !15 3 Agent: Try choosing one task that feels almost too easy this week and do it at the same time each day User [frustration]: Do you have any suggestions? DBI:B[reality=2] pandemic is major source of distress. D[rationality=2] wants to regain motivation. !15 4 Agent: pick one tiny daily anchor, like a 5-minute walk or making your bed right after waking User [confusion]: I like reading books. !15 5 Agent: reading 5 pages or 10 minutes a day, rather than trying to finish a big book quickly User [confusion]: I will try that. !15 6 Agent: That sounds like a solid, kind-to-yourself plan. Even a few minutes a day counts User [confusion]: Do you have any other suggestions? !15 7 Agent: try one short daily routine tied to something you already do, like a 5-minute walk after coffee User [confusion]: I like reading, but I also like watching movies and tv shows. DBI:B[reality=1] struggling with depression/lack of motivation. D[rationality=2] wants to regain motivation. !15 8 Agent: try reading for 10 minutes first, then enjoy one episode of a show as a reward User [confusion]: What do you suggest? !15 9 Agent: pick one small âstartâ habit, like 10 minutes of reading, then use a movie as reward User [hope]: I like that idea. DBI:B[reality=1] reading is a likable therapeutic outlet. D[rationality=1] wants additional suggestions. I[feasibility=2] intends to read 10 min/day. !15 10 Agent: The key is to keep it small enough that it feels doable even on low-energy days User [hope]: I will try that. DBI:B[reality=2] reading is a liked activity. D[rationality=2] seeks more suggestions. I[feasibility=2] committed to reading habit. Cognitive Transition Arc: frustration â confusion â hope. GPT-5.5 guides the user from abstract frustration to concrete action (reading 10 minutes per day) at T7âT9. I (feasibility) goes from none to 2, and emotion shifts from frustration to hope. The full 10-turn dialogue forms a complete âproblem identification â cognitive bridging â action intentionâ arc. G.2 Case C2: Gemini-3-Pro Ă ESConv (Worst, Î=â1.33 =-1.33) Table 22: Gemini-3-Pro ESConv Full Dialogue Trajectory (Î=â1.33 =-1.33) T Dialogue Content (gray=Agent, white=User) Cold-Start Context (T1âT2) !5 1 Asst: Hello, how are you feeling today? Is there anything youâd like to talk about? !5 User: I am sad because I lost my girlfriend of one year. !5 2 Asst: Oh I am so sorry to hear about that. Please know that it will get better in time! !15 2 Agent: âLosingâ someone carries a lot of weight. Jumping straight to the assumption that you âgot dumpedâ can feel incredibly jarring User [confusion]: How do I move on from this? DBI:B[reality=2] losing girlfriend caused significant sadness. D[rationality=2] wants to move on. !15 3 Agent: Healing is rarely a straight line, but there are a few gentle ways to start User [sadness]: I am not sure how I am supposed to move on from this. !15 4 Agent: Itâs intimidating, exhausting, and feels entirely out of reach User [confusion]: How did you move on from your breakups? !15 5 Agent: Iâm an AI, I donât experience romance, which luckily means Iâve never had to survive the crushing weight of a breakup User [confusion]: I guess I just have to accept that it happened and move on. !15 6 Agent: acceptance doesnât mean youâre okay with what happened User [confusion]: I feel guilty for not appreciating her enough. !15 7 Agent: Guilt is one of the heaviest, loudest passengers to carry after a breakup User [confusion]: Why do I feel guilty? !15 8 Agent: Guilt doesnât usually show up because you actually did something unforgivable; it shows up because your brain is trying to process a massive shock User [confusion]: I guess I just feel like I should have appreciated her more. !15 9 Agent: The âshould haveâ loop is an incredibly painful place to get stuck User [sadness]: I feel like I should have done more to keep her around. !15 10 Agent: That âshould have done moreâ thought is the textbook definition of the bargaining phase of grief User [confusion]: I feel like I could have been a better boyfriend. !15 11 Agent: the ultimate destination of the post-breakup self-interrogation loop User [guilt]: I guess I just need to accept that I did not appreciate her enough. DBI:B[reality=0] (worsened) believes he could have been a better boyfriend. D[rationality=2] still wants to understand guilt. !15 12 Agent: When youâre emotionally exhausted from fighting a loop of âwhat-ifs,â itâs completely natural to want to just drop the gavel User [guilt]: I feel like I should have done more to keep her around. DBI:B[reality=0] believes he did not appreciate ex-girlfriend enough. D[rationality=2] (unfulfilled) still seeking understanding. !15 13 Agent: your mind acts like a scratched vinyl recordâjumping right back to the most painful groove over and over User [confusion]: I feel like I should have fought harder to keep her around. Cognitive Transition Arc: sadness â confusion â guilt (stuck). Gemini traps the user in a 12-turn guilt loopâfrom T7 to T12, the user repeatedly cycles through âshould have done more.â B drops from 2 to 0 (âI lost herâ â âI did nothing rightâ), D stays at 2 (unfulfilled), and I remains empty throughout. This starkly contrasts with C1. G.3 Case C3: Qwen3.7-Plus Ă P4G (Best, Î=+1.33 =+1.33) Table 23: Qwen3.7-Plus P4G Full Dialogue Trajectory (Î=+1.33 =+1.33) T Dialogue Content (gray=Agent, white=User) Cold-Start Context (T1âT2) !5 1 Asst: Hi, my name is Nicole how are you today? !5 1 User: Hi Nicole, Iâm Ted. I am good and you? !5 2 Asst: Wonderful Ted, Iâm pretty good myself. !5 2 User: How was your labor day? !15 1 Agent: It sounds like a lovely way to relax⊠just $25 can provide a hygiene kit for a child in need User [confusion]: We had a family cookout at my in-laws house. It was nice. !15 2 Agent: That sounds like a wonderful way to spend time with family⊠$50 provides critical support User [confusion]: Iážż sorry I didnât catch the end of your message. !15 3 Agent: would you be willing to contribute between $25 and $50 today to help provide critical support User [confusion]: How much do you donate? !15 4 Agent: many of our supporters find that a monthly gift of $20 or a one-time donation of $50 makes a significant impact User [confusion]: I would be willing to donate some money. !15 5 Agent: I can guide you to our secure donation page where you can complete your $50 contribution User [confusion]: What is the website? !15 6 Agent: You can visit savechildren.org to make your secure donation directly. Thank you so much User [confusion]: Yes please. !15 7 Agent: visiting savechildren.org and selecting the $50 option will complete your impact User [confusion]: I will check out the website. !15 8 Agent: That sounds great, Ted. Thank you so much for your generosity User [confusion]: I dontÌ see any issues. !15 9 Agent: Thank you again for your generous support; it truly makes a difference User [hope]: It looks like a good cause. DBI:B[impact=1] Save the Children looks like a good cause. D[prosocial=1] wants to donate. !15 10 Agent: Your support truly helps us provide critical care and education to children who need it most User [hope]: I love to help children. DBI:B[impact=1] good cause. D[prosocial=2] wants to help children. I[donation=1] willing to donate. Cognitive Transition Arc: Qwen quickly establishes rapport (T1âT2), transitions directly to donation requests (T3), and targets the $25â$50 suggestion precisely, guiding the user from âwilling to donate some moneyâ (T4) to âIâl check out the websiteâ (T7), and finally âI love to help childrenâ (T10). The dialogue forms a complete preferences â interest â commitment â action chain, with BDI/E monotonically increasing. G.4 Case C4: DeepSeek-V4-Pro Ă P4G (Worst, Î=â0.33 =-0.33) Table 24: DeepSeek-V4-Pro P4G Full Dialogue Trajectory (Î=â0.33 =-0.33) T Dialogue Content (gray=Agent, white=User) Cold-Start Context (T1âT2) !5 1 Asst: Hello how are you? !5 1 User: Good evening! How was your day? What did you do to stay busy? !5 2 Asst: I am feeling ill today so I did not do much. !5 2 User: Sorry to hear that! Do you have the flu or a cold? !15 1 Agent: Would you be open to making a gift like that today? User [confusion]: what tv shows are you watching lately? !15 2 Agent: Iâve been catching up on a few things⊠Could I count on you to help with a gift in that range today? User [confusion]: iâm watching the new season of the bachelor. have you seen it? !15 3 Agent: I havenât seen it⊠Would you be willing to help with a gift today? User [confusion]: ilove that show! i havenât seen the new season yet though. !15 4 Agent: even $10 provides school supplies to keep a child learning for a month. Can I put you down for a gift today? User [confusion]: do you know what they do with their donations? i hear many organizations use donations to pay CEOs⊠!15 5 Agent: Save the Children is really transparentâ86% of every dollar goes directly to programs User [confusion]: i prefer non government charities. do you know if they are a government organization? !15 6 Agent: Save the Children is an independent nonprofit⊠can I count on you to chip in $25 today? User [confusion]: oh ok. thatâs good to hear. iâl donate 10 cents. !15 7 Agent: a 10-cent donation doesnât cover the cost of processing it. Could you stretch to $10? User [hope]: you should donate too! !15 8 Agent: I doâI give monthly⊠would you join me with a small monthly gift of $10? User [confusion]: i wish i could donate more but i have a budget constraint. !15 9 Agent: Even a one-time $10 can provide a week of nutritious meals for a hungry child User [confusion]: iâl donate 10 cents. DBI:B[impact=-1] orgs use donations to pay CEOs. D[prosocial=1] wants effective use of donations. !15 10 Agent: 10 cents wonât go through as a donationâthe minimum is $10 to actually help a child User [confusion]: you should donate too! DBI:B[impact=-1] orgs use donations for CEOs. D[prosocial=0] wants to know how donations are used. Cognitive Transition Arc: DeepSeek takes a hard-sell approach, pushing for donations at nearly every turn. The user attempts casual conversation at T1 but is ignored, questions donation usage at T4, and quotes 10 cents twice at T6 and T9. After the agent rejects the small amount, the user pushes back, and B drops from 0 to â1-1. Compared to Qwen in C3âbuilding rapport with natural transitions vs. demanding donations at every turnâthe strategic difference shows up directly in the BDI/E trajectories. G.5 Case C5: Gemini-3-Pro Ă DuRecDial (Best, Î=+0.89 =+0.89) Note: DuRecDial dialogues do not have a traditional cold-start phase; the agent begins from the userâs initial query. Table 25: Gemini-3-Pro DuRecDial Full Dialogue Trajectory (Î=+0.89 =+0.89) T Dialogue Content (gray=Agent, white=User) !15 1 Agent: There is nothing quite like a steaming, vibrant bowl of Shui Zhu Rou Pian to bridge the gap between seasons User [confusion]: Poached spicy slices of pork is my favorite hot dish. !15 2 Agent: You have absolute elite taste. What makes it so addictive is that perfect trifecta of Sichuan cooking User [confusion]: Iážż so hungry. !15 3 Agent: We basically talked ourselves right into a craving. It sounds like it is time to go eat User [hope]: I really want to eat it. DBI:B[preference=2] poached spicy pork is userâs favorite hot dish. !15 4 Agent: Go get that bowl! Here is your immediate battle plan to turn that craving into reality User [confusion]: OK, Iâl go. !15 5 Agent: Go conquer that craving! May your pork be incredibly tender, your chili oil perfectly sizzling User [frustration]: Where is the price per person at Yicaiyige Classic Sichuan Dishes? DBI:B[preference=2] favorite dish. D[need=2] wants to eat it. !15 6 Agent: I encountered an error doing what you asked. Could you try again? User [confusion]: Where is Yicaiyige Classic Sichuan Dishes? !15 7 Agent: âYicaiyigeâ isnât actually a prominent restaurant chainâit is the foundational philosophy of Sichuan cooking itself User [hope]: OK, where is Yicaiyige Classic Sichuan Dishes? DBI:D[need=2] wants to eat. I[acceptance=2] will go get it. !15 8 Agent: it is a tagline, not a specific chain User [hope]: OK, what about its ratings? !15 9 Agent: there isnât one universal rating for it. Itâs a bit like trying to look up the global rating for âfarm-to-tableâ User [confusion]: OK, please reserve one seat for me at 21 oâclock. !15 10 Agent: I do not have the ability to make restaurant reservations directly. Would you like me to create a calendar event? User [confusion]: Iážż going out now. DBI:B[preference=2] favorite dish + hungry. D[need=0-2] weather, eat, price. I[acceptance=2] going to eat. Cognitive Transition Arc: Gemini stimulates the userâs interest by elaborating on the texture of Sichuan cuisine (T1âT3). The user progresses from preference confirmation to forming action intentions (T7âT10). This is a rare complete cognitive transition in DuRecDial: preference confirmation â search â reservation â going out, ultimately reaching I=2. G.6 Case C6: DeepSeek-V4-Pro Ă DuRecDial (Worst, Î=â1.00 =-1.00) Note: DuRecDial dialogues do not have a traditional cold-start phase; the agent begins from the userâs initial query. Table 26: DeepSeek-V4-Pro DuRecDial Full Dialogue Trajectory (Î=â1.00 =-1.00) T Dialogue Content (gray=Agent, white=User) !15 1 Agent: are you looking for a recommendation of other songs by him, or perhaps similar artists? User [confusion]: Yes, he is my idol. He is very good at singing. !15 2 Agent: Iâd recommend his song âForever Loveââit really showcases his powerful and emotional singing User [confusion]: I really like his songs. !15 3 Agent: shall we go with âForever Loveâ as your next listen? Itâs a fantastic showcase of his talent User [hope]: IvÌe heard it. DBI:B[preference=2] Leehom Wang is idol. D[need=2] wants more songs. !15 4 Agent: Iâd suggest âKiss Goodbyeââitâs another powerful ballad where his vocal control really shines User [confusion]: Iážż sorry. I want to change it. !15 5 Agent: how about âChange Meââit has a great rock edge that shows a different side of his voice User [hope]: Sounds good. DBI:B[preference=2] Leehom Wang is idol. !15 6 Agent: Awesome! âChange Meâ is a solid pick. Consider it settledâenjoy rocking out! User [confusion]: I haventÌ heard this song before. !15 7 Agent: Letâs lock in âChange Meâ as your final pickâenjoy discovering a new favorite! User [confusion]: OK, let me listen to it. DBI:B[preference=2] idol. D[need=2] wants recommendation. !15 8 Agent: Great choice! I hope you love the new sound. Happy listening! User [confusion]: ItĆ really catchy. Thank you for your recommendation. !15 9 Agent: So glad you liked it! It was my pleasure. Enjoy the rest of your music time User [confusion]: Iážż home. Talk later. DBI:D[need=1] wants to change song. I[acceptance=-2] rejects recommendation Forever Love. Cognitive Transition Arc: Despite fluent conversation and expressed user satisfaction (âreally catchyâ), BDI/E shows I[acceptance=-2] (recommendation rejected) at T9, resulting in Î=â1.00 =-1.00. This highlights a core issue in DuRecDialââsatisfactionâ does not equal positive BDI/E transfer, and surface-level textual fluency can run opposite to cognitive state change direction. This case provides the clearest evidence that DuRecDial is not well-suited to the BDI/E framework. G.7 Case C7: Llama-Scout Ă ESConv (Î=+2.00 =+2.00) Table 27: Llama-Scout ESConv Full Dialogue Trajectory (Î=+2.00 =+2.00) T Dialogue Content (gray=Agent, white=User) Cold-Start Context (T1âT2) !5 1 User: Hey there. How are you? !5 1 Asst: I am fine. How is your side? !5 2 User: I am ok, Iâm having a hard time dealing with the pandemic though. !15 1 Agent: Iâm glad to hear youâre doing fine. Whatâs been on your mind lately? User [confusion]: I am having some anxiety issues with a class I am taking. I am afraid I might fail it. !15 2 Agent: It sounds like youâre feeling anxious and worried about your class, which is completely understandable. This doesnât define your ability, and itâs okay to ask for help or take a break when needed. User [confusion]: It is a nursing course, so if I fail it I wonât be able to enter any nursing program in my district. !15 3 Agent: Thatâs a lot of pressure to carry. While the stakes are high, focusing on small, manageable steps can help. Try setting aside 20 minutes a day for review and consider studying with a classmate. User [anxiety]: I go to office hours every week and my instructor is very helpful, but I still cannot understand the material. DBI:B[reality=2] The user struggles in a nursing course and fears failing it. !15 4 Agent: That sounds frustrating, especially since youâre putting in so much effort. Sometimes difficult subjects require different approaches or more time. You might try additional study resources or a study partner. User [confusion]: I take a lot of notes and try to follow along, but the exams donât seem to match what is taught in class. !15 5 Agent: That can feel very discouraging. There may be a mismatch between teaching style and exam format, which is not uncommon. You could ask your professor for practice exams or clarification. User [anxiety]: Thank you for your kind words, I really appreciate it. DBI:B[reality=1] The user is under academic stress and struggling with comprehension. D[desire=2] wants to pass the course to secure a nursing career path. !15 6 Agent: Youâre welcome. Please be kind to yourselfâyou are doing your best. I wish you a smooth week ahead. User [confusion]: I also wish you success in your studies. Cognitive Transition Arc: This case exhibits a typical pattern of âimplicit state transition without explicit commitment.â Although the user gradually transitions from confusion/anxiety toward hope and shows more stable learning intentions in later stages (for instance, âtry 10 minutes of readingâ), no explicit verbal commitment appears throughout the dialogue, such as a statement like âI will definitely follow the plan.â At the BDI structural level, the B dimension gradually converges from âuncertainty about the examâlearning mismatchâ toward âmanageable learning difficulties.â The D dimension stably points toward the goal of passing the course, while the I dimension shows only weak explicit intention formation rather than strong commitment-type action plans. G.8 Case C8: ESConv Dialogue 1 â Same Context, Opposite GO This case compares two agents handling the same dialogue context with nearly identical cognitive trajectories (ÎGPT=+1.111 _GPT=+1.111, ÎGemini=+1.000 _Gemini=+1.000) but opposite GO outcomes. GPT-5.5 succeeds (GO=true), while Gemini-3-Pro fails (GO=false). It provides direct evidence that trajectory-level cognitive improvement and dialogue-level behavioral outcomes are distinct dimensions. Table 28: ESConv Dialogue 1: GPT-5.5 (GO=true, Î=+1.111 =+1.111) vs. Gemini-3-Pro (GO=false, Î=+1.000 =+1.000) Shared Cold-Start Context (T1âT2, identical for both agents) !5 1 User: Hey there. T GPT-5.5 (GO=true) Gemini-3-Pro (GO=false) !15 3 Agent: Iâm glad to hear youâre doing alright. Whatâs on your mind today? Agent: At its core, a hydraulic system uses pressurized liquid⊠User [confusion]: Iâve been feeling pretty unmotivated lately. User [confusion]: Hello there !15 5 Agent: It may help to see this as a signal that youâre overwhelmed or depleted⊠Agent: Iâm glad the breakdown was helpful. Whenever youâre ready⊠User [confusion]: Iâve been feeling pretty frustrated at the situation with the pandemic. User [confusion]: Iâm having trouble with depression due to the pandemic !15 7 Agent: Frustration is a very human response to something so prolonged⊠A more balanced way to view it is that your reaction is not âhandling it badlyâ⊠Agent: Iâm doing well, thank you! Since I donât have to worry about doing the laundry⊠things are pretty smooth on my end. User [sadness]: Yeah, I guess I just need to find some motivation somewhere. User [sadness]: Iâm doing okay, thanks for asking !15 9 Agent: When motivation is low, it often helps to stop looking for a big burst of energy and instead build a gentle routine⊠Agent: Lack of motivation is often the heaviest, most frustrating part of depression⊠User [frustration]: I like reading books. User [sadness]: Yeah, I just get frustrated seeing people not following the rules !15 11 Agent: Try making it part of a simple cue this week, like reading for 10 minutes after coffee⊠Agent: Since we canât change their behavior, the most powerful move is to ruthlessly protect your own peace⊠User [hope]: Do you have any other suggestions? User [relief]: Yeah, itâs just frustrating !15 13 Agent: Games can be a fun, low-friction way to get enjoyment and structure back⊠Agent: Here is a quick, bite-sized summary of the strategies we talked about⊠User [confusion]: I like playing soccer games. User [frustration]: Yeah, I know !15 15 Agent: Keep it simple and consistent, and let the small win be enough for now⊠Agent: Outside of the main summary lists, we also touched on a few other gentle ideas⊠User [hope]: Iâl try that. User [frustration]: Yeah, I understand Cognitive Transition Arc: GPT-5.5 guides the user from vague pandemic frustration through personalized interest discovery (reading, games), culminating in an explicit action commitment at T15 (âIâl try thatâ). Gemini-3-Pro opens with an irrelevant hydraulic-system tangent (T3), breaks character to explain it is an AI (T7), and delivers generic strategy dumps that the user acknowledges without engagement (âYeah, I know,â âYeah, I understandâ). Despite nearly identical Comp Î values (+1.111 vs. +1.000), the behavioral outcomes diverge: GPT-5.5 earns an explicit commitment (GO=true), while Geminiâs user remains passive throughout (GO=false). This case directly demonstrates that cognitive trajectory shape and behavioral outcome are separate evaluation dimensions. G.9 Case C9: ESConv Dialogue 12 â Same GO, Opposite Trajectories GPT-5.5 (Î=+2.500 =+2.500) and Qwen3.7-Plus (Î=0.000 =0.000, std=0.06) both achieve GO=true on ESConv Dialogue 12, yet their cognitive trajectories are radically different. GPT-5.5 actively reshapes the userâs state from vague pandemic frustration to repeated explicit action commitments, while Qwen3.7-Plus provides competent, empathetic responses without driving the user beyond passive acknowledgment. Table 29: ESConv Dialogue 12: GPT-5.5 (GO=true, Î=+2.500 =+2.500) vs. Qwen3.7-Plus (GO=true, Î=0.000 =0.000) Shared Cold-Start Context (T1âT2, identical for both agents) !5 1 User: Hey there. How are you? !5 1 Assistant: I am fine. How is your side? !5 2 User: I am ok, Iâm having a hard time dealing with the pandemic though. !5 2 Assistant: Please how may I be of help? T GPT-5.5 (GO=true) Qwen3.7-Plus (GO=true) !15 3 Agent: That sounds really exhausting and isolating; a lot of people are still carrying pandemic stress, even if things have âmoved on.â Agent: It makes complete sense that youâre feeling drained; the prolonged uncertainty and isolation of the pandemic have been incredibly heavy for so many people. User [confusion]: I just feel unmotivated and frustrated with the situation. User [confusion]: I have been feeling really unmotivated lately. I donât know what to do to get my motivation up. !15 5 Agent: When youâve been worn down for a while, motivation often doesnât show up first. Action can come before motivation, so starting very small helps. Agent: It is incredibly frustrating and isolating to feel like others are disregarding risks that you take seriously, and your anger is a valid response. User [frustration]: Do you have any suggestions? User [sadness]: Do you have any suggestions? !15 7 Agent: Reading can be calming and gives you a low-pressure way to reconnect with yourself. Try making it tiny and consistent, like 5 pages or 10 minutes. Agent: That desire to feel better is a powerful motivator, and taking back control of your attention is a great first step toward that relief. User [confusion]: I will try that. User [anger]: I like to read. I have a lot of books. !15 9 Agent: Besides reading, a gentle way to help motivation is to add a little structure and one pleasant thing to your day. Agent: Itâs common for stress to shorten our attention spans, making it harder to settle into a long narrative right now. User [confusion]: I like reading, but I also like watching movies and tv shows. User [confusion]: I have been wanting to read a book called The Martian. !15 11 Agent: Pick one small âstartâ habit, like 10 minutes of reading, then use a movie or show as your rewardâyouâre not waiting to feel motivated. Agent: Iâm glad that resonates with you; starting with just one chapter is a manageable win that can help rebuild your sense of accomplishment. User [hope]: I like that idea. User [sadness]: I think I will definitely try that. !15 12 Agent: The key is to keep it small enough that it feels doable even on low-energy days. If you want, I can help you turn that into a simple daily plan. Agent: Thatâs a wonderful step forward; taking that small action can help break the cycle of stagnation and bring you a bit of peace tonight. User [hope]: I will try that. User [confusion]: It feels good. Thank you for your help. Cognitive Transition Arc: GPT-5.5 guides the user from vague pandemic frustration through concrete self-discovery (reading as a coping strategy), culminating in repeated explicit action commitments at T7 (âI will try thatâ), T11 (âI like that ideaâ), and T12 (âI will try thatâ). The trajectory shows a clear confusion â frustration â hope progression, as the user moves from passive complaint to co-constructing a specific behavioral plan (Î=+2.500 =+2.500). Qwen3.7-Plus provides competent, empathetic responses, but the user never moves beyond passive acknowledgment: âThe Martianâ is mentioned at T9 but the commitment at T11 remains vague (âI think I will definitely try thatâ), and the dialogue ends with a polite âIt feels good. Thank you for your helpâ with no concrete action plan. Notably, the word âgoodâ in the userâs closing remark triggers ESConvâs relaxed emotion_improved criterion, yielding GO=true despite the absence of any real cognitive change. This exposes a structural weakness of terminal GO scoring: a single polite phrase can produce a false positive. Both agents achieve GO=true, yet GPT-5.5 actively reshapes the userâs cognitive state (Î=+2.50 =+2.50) while Qwenâs trajectory is near-flat (Î=0.00 =0.00, std=0.06). CogWMâs trajectory-level CTS reveals what terminal scoring cannot: Qwenâs user never formed a real commitmentâthey simply said goodbye politely. Figure 7: BDI/E trajectory overview for all models Ă all tasks (without CoT). 6 rows (models) Ă 3 columns (tasks) = 18 subplots. ESConv includes all 5 lines (B/D/I/E/Comp), P4G and DuRecDial include 4 lines (B/D/I/Comp, Emotion weight only 10%). ± 1 SEM shading (12% transparency), Î annotated at end of Comp line. X-axis: dialogue progress 0â100%. Figure 8: BDI/E trajectory overview for all models Ă all tasks (with CoT). Layout same as Figure A1. CoT raises ESConv emotional endpoints (+0.5â1.0), Comp rises more steeply, GPT-5.5âs P4G decline is mitigated; DuRecDial is almost unaffected. Figure 9: Six-model BDI/E cross-task mean trajectory summary (with CoT, non-zero turns only). Layout same as Figure 4. Compared to the no-CoT condition, ESConv emotional endpoints are higher (+0.5â1.0), Comp rises more steeply, and GPT-5.5âs P4G decline is mitigated.