Paper deep dive
PsyMemBehav: Psychology-Informed and Memory-Grounded Two-Level Identity-to-Action Reasoning for Social-Media Behavior Simulation
Wenji Zhang, Xilong Cheng, Yunxiao Qin, Mai Tu, Tongye Zhu, Yuhao Zhou, JingBo Ji, Yuting Tan, Duoyin Li, Hongjiang Xiao, Yuan Zhang
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models (LLMs) support social simulation, but faithfully reproducing real users requires grounding both structured platform decisions and language in sparse, noisy behavioral evidence. We introduce PsyMemBehav, which maps psychology-informed profiles and retrieved memories to target-first platform behavior through two-level identity-to-action reasoning. To support this approach with comprehensive behavioral traces, we construct a dataset PsyMemBehav-bench from 5.57M observed behaviors of 5,862 real users (approximately 950 per user) across Reddit, Bluesky, and Twitter. Each recorded engagement is modeled as target selection followed by Comment, Like, or Repost and optional response generation. PsyMemBehav combines 41-dimensional user descriptors, top-5 BM25 memories, reverse-CoT supervision, and rubric-guided DPO over complete reasoning-action trajectories. Experiments show that PsyMemBehav 27B ranks first in 11 of 12 reported individual-level metrics, improving average response quality over GPT-5.4 from 54.28 to 59.68 (9.95% relative), and obtains the lowest group-level MAE (0.081) and MSE (0.009) among simulated systems. These results support evidence-grounded user modeling for observed-thread continuation; the benchmark remains conditional on recorded engagement and does not model no-action decisions.
Tags
Links
- Source: https://arxiv.org/abs/2610.14570v1
- Canonical: https://arxiv.org/abs/2610.14570v1
Trouble viewing inline? Open PDF directly â
Full Text
95,222 characters extracted from source content.
Expand or collapse full text
PsyMemBehav: Psychology-Informed and Memory-Grounded Two-Level Identity-to-Action Reasoning for Social-Media Behavior Simulation Wenji Zhang Xilong Cheng Yunxiao Qin Mai Tu Tongye Zhu Yuhao Zhou JingBo Ji Yuting Tan Duoyin Li Hongjiang Xiao Yuan Zhang Abstract Large language models (LLMs) support social simulation, but faithfully reproducing real users requires grounding both structured platform decisions and language in sparse, noisy behavioral evidence. We introduce PsyMemBehav, which maps psychology-informed profiles and retrieved memories to target-first platform behavior through two-level identity-to-action reasoning. To support this approach with comprehensive behavioral traces, we construct a dataset PsyMemBehav-bench from 5.57M observed behaviors of 5,862 real users (approximately 950 per user) across Reddit, Bluesky, and Twitter. Each recorded engagement is modeled as target selection followed by Comment, Like, or Repost and optional response generation. PsyMemBehav combines 41-dimensional user descriptors, top-5 BM25 memories, reverse-CoT supervision, and rubric-guided DPO over complete reasoning-action trajectories. Experiments show that PsyMemBehav 27B ranks first in 11 of 12 reported individual-level metrics, improving average response quality over GPT-5.4 from 54.28 to 59.68 (9.95% relative), and obtains the lowest group-level MAE (0.081) and MSE (0.009) among simulated systems. These results support evidence-grounded user modeling for observed-thread continuation; the benchmark remains conditional on recorded engagement and does not model no-action decisions. 11 1 All code, checkpoints, and data will be made publicly available immediately upon acceptance. 1Communication University of China, China 2State Key Laboratory of Media Convergence and Communication, China 202311153001@mails.cuc.edu.cn, chengzhengyu330@cuc.edu.cn, qinyunxiao@cuc.edu.cn *Equal contribution. â Corresponding author. 1 Introduction Online public opinion develops through repeated user interactions: people comment, like, repost, choose whom to address, and express stances within evolving discussion threads. When aggregated, these micro-level decisions shape observable stance distributions, interaction networks, attention concentration, and information diffusion. Simulating this process offers a controlled setting for studying how public opinion may evolve under different discussion contexts (Hu et al. 2024; Liu et al. 2025; Jeon et al. 2025; Cheng et al. 2025). LLM agents offer a flexible mechanism for such simulations, but the credibility of aggregate results depends in part on whether individual behaviors are grounded in heterogeneous real-user evidence (Argyle et al. 2023; Gao et al. 2024). Fluent yet generic or uniformly active agents can distort action frequencies, target selection, and stance composition when their outputs are aggregated. We therefore study user-level targetâactionâresponse prediction as a behavioral foundation for evaluating thread-level public-opinion simulation in observed social-media contexts. Figure 1: Motivation example of identity-driven social-media user simulation. Given Renâs profile, retrieved memory, and current thread, the simulator first activates profile and memory evidence relevant to the thread, then selects the target before predicting the action and response. Conventional character role-play usually starts from an explicit persona and evaluates whether generated dialogue remains persona-consistent (Xu et al. 2026b; Xiao et al. 2026; Cheng et al. 2026). Social-media user simulation poses a different task: it infers observable platform behavior from a userâs partial and noisy behavioral history together with the current thread context. The prediction covers both textual responses and structured decisions about target selection and the subsequent action type. We formulate the task as predicting a userâs observed engagement in a given postâcomment context, including the target, subsequent action type, and optional textual response. We normalize observed interactions into Comment, Like, and Repost22 2 Textual interactions such as replies and comments are mapped to Comment; resharing behaviors are mapped to Repost; and reaction-only behaviors are mapped to Like.. The logs record positive interactions but not impression-level exposure. Accordingly, the benchmark is conditional on a recorded engagement event, contains no explicit no-action label, and cannot distinguish unseen threads from threads that a user viewed but ignored. Within this scope, the task predicts the target, subsequent action, and optional text. Existing LLM-based social simulation systems have explored information diffusion, online conversation, and social media event modeling (Park et al. 2023; Gao et al. 2024; Hu et al. 2024; Liu et al. 2025; Jeon et al. 2025; Ji et al. 2026). These studies show that LLM agents can generate natural-language interactions and support controlled simulation of social processes. However, they often rely on coarse personas, demographic attributes, or manually specified agent descriptions, and are less focused on predicting a concrete userâs platform behavior from long-horizon behavioral histories and situated thread context. To address this, we introduce PsyMemBehav, a structured framework that predicts social media behavior by integrating user profiles, longitudinal memories, and thread context. The framework maps this evidence to behavior through two sequential steps: identity activation, which selects relevant traits and memories for the current thread, and action decision, which identifies the target and predicts the subsequent action and response. Furthermore, we align the complete reasoning-action trajectories via rubric-guided preference optimization. This ensures the simulator optimizes not just for response fluency, but also for action fidelity, memory grounding, and behavioral consistency. To evaluate this, we construct a large-scale dataset PsyMemBehav-bench, containing 5.57M behaviors from 5,862 users across Reddit, Bluesky, and Twitter. Each sample of PsyMemBehav-bench integrates 41 psychology-informed descriptors, retrieved history, context, and observed behavior. This enables the joint evaluation of structured decision-making and text generation, moving beyond traditional text-only role-playing. We evaluate PsyMemBehav at the individual level by matching held-out user interactions and at the thread level by comparing aggregate comment-section statistics in observed post contexts. In summary, our contributions are as follows: 1) We propose PsyMemBehav, a novel framework that models conditional social-media behavior through a two-stage structured decision process: identity activation and action decision. By dynamically integrating 41-dimensional psychology-informed user profiles and retrieved longitudinal memories with thread context, our approach moves beyond flat text generation to produce grounded, context-aware behavioral predictions. 2) We introduce a tailored alignment strategy combining Reverse-CoT supervision and rubric-guided DPO to optimize complete reasoning-action trajectories. This ensures the simulator is optimized not merely for response fluency, but critically for action fidelity (e.g., predicting Comment, Like, or Repost with conditional text generation), robust memory grounding, and strict persona consistency. 3) On the large-scale dataset PsyMemBehav-bench, we conduct comprehensive evaluations on both held-out user behavior and observed-thread continuation. The resulting PsyMemBehav-27B ranks first in 11 of 12 individual metrics, achieving a 9.95% relative improvement in average response quality over GPT-5.4, while yielding the lowest aggregate group-level deviation. Extensive ablation studies further validate the critical contributions of both training stages, profile dimensions, and decision components. 2 Related Work 2.1 Social-Media Account Simulation LLM-based agents have been used to simulate human-like behavior in role-playing, interactive environments, and social settings (Park et al. 2023; Argyle et al. 2023; Gao et al. 2024). Prior work shows that LLMs can maintain personas, imitate conversational styles, and produce role-consistent responses (Shao et al. 2023; Wang et al. 2024; Tu et al. 2024). In parallel, user modeling studies show that digital traces reveal stable attributes such as personality, preferences, demographics, and behavioral tendencies (Kosinski et al. 2013; Youyou et al. 2015; Hovy and Søgaard 2015), motivating personalized behavior generation from historical traces. Social-media account simulation, however, goes beyond generic persona imitation. Users do not only generate text, but also select interaction targets within a thread and choose platform-specific actions such as replying, quoting, reposting, liking, or ignoring. Although recent work explores LLM agents for interactive social behavior and social simulation (Ren et al. 2024; Zhou et al. 2024), existing approaches often focus on generic personas, aggregate populations, or domain-specific simulated users rather than real accounts grounded in historical traces and platform-specific action spaces. Our work addresses this gap by modeling real accounts with psychological traits, platform-specific behavioral traits, retrieved memory, and thread context. 2.2 Rubric-Guided Alignment for Social-Behavior Simulation Preference learning aligns LLMs with desired behaviors through human or AI feedback and direct preference optimization (Ouyang et al. 2022; Bai et al. 2022; Lee et al. 2023; Rafailov et al. 2023), typically targeting general objectives such as helpfulness, harmlessness, or instruction following. Social-media simulation instead requires faithfulness to user profiles, grounding in memory, platform-appropriate actions, thread-context coherence, and social plausibility. LLM-as-a-judge methods enable evaluation with explicit criteria rather than a single overall score (Zheng et al. 2023; Liu et al. 2023; Kim et al. 2024). Building on this idea, we construct rubric-guided preference pairs over complete reasoning-action trajectories, aligning both intermediate reasoning and resulting platform behavior rather than only final textual responses. 2.3 Group-Level Thread Simulation Agent-based modeling studies how individual behaviors produce collective phenomena such as diffusion, coordination, polarization, and network formation (Epstein 1999; Macy and Willer 2002; Lazer et al. 2009). LLM agents extend this tradition with language-based reasoning and context-sensitive interaction policies (Park et al. 2023; Gao et al. 2024), while introducing challenges such as unstable identities, unrealistic information access, and weak behavioral grounding. Recent work argues that social simulation should be evaluated beyond single-turn plausibility, including interaction dynamics, stance distributions, network structures, conversational diversity, and thread-level realism (Zhou et al. 2024; Zhang et al. 2025a). We therefore first improve account-level identity-to-action modeling through psychology-informed reasoning and rubric-guided DPO, and then evaluate whether these improvements transfer to group-level thread simulations. 3 Methodology 3.1 Problem Formulation and Overview We formulate social-media user simulation as a time-anchored multi-action prediction problem. For a user u at time t, the simulator observes xt=(Pu,Mu,t,Tt,Ct),x_t=(P_u,M_u,t,T_t,C_t), -1.00006pt where PuP_u is the user profile, Mu,tM_u,t is retrieved behavioral memory before time t, TtT_t is the current thread context, and CtC_t is the set of candidate targets, i.e. the visible post or comments the user may interact with at t. The simulator predicts a complete platform behavior yt=(ct,at,rt),y_t=(c_t,a_t,r_t), -1.00006pt where ctâCtc_tâ C_t is the selected target, atâa_t the subsequent action, and rtr_t the generated comment when required. We normalize the heterogeneous behaviors in raw logs into three visible actions, =Comment,Like,RepostA=\ Comment, Like, Repost\: textual engagement (comments, replies) maps to Comment, reaction-only behavior to Like, and resharing (including quote reposts) to Repost; original posts and past interactions serve as evidence rather than prediction labels. This requires more than fluent response generation: conditional on a recorded engagement event, the simulator must first select a target, then choose Comment, Like, or Repost and generate text when required. Because behavior is context-dependent, a plausible comment can still be wrong if the user would target a different item or act differently. We address this with two design choices. First, two-level identity-to-action reasoning situates the user in the current thread before deciding the platform behavior (§3.2). Second, two-stage trajectory training optimizes complete reasoning-action trajectories rather than only final responses, through reverse-CoT supervision and rubric-guided preference optimization (§3.3). Figure 2 gives an overview. 3.2 Two-Level Identity-to-Action Reasoning Figure 2: Overview of the proposed two-stage training framework. Stage 1 reconstructs reverse-CoT traces from observed behavior and trains the simulator by supervised fine-tuning on accepted trajectories. Stage 2 applies rubric-guided DPO to align complete reasoning-action trajectories with social-behavior criteria. The right panel summarizes the individual- and group-level evaluation. Each simulation instance is modeled as two ordered decisions: the evidence relevant to the user in the current thread is identified first, and the corresponding platform behavior is predicted second. This separation reflects the difference between a broad, history-derived profile and a context-specific action, while making the intermediate evidence selection explicit. A direct mapping from xtx_t to yty_t remains possible, but it does not expose separate supervision for evidence selection and action decision. We therefore insert an intermediate reasoning trace xtâztâyt,zt=(ztid,ztact),x_tâ z_tâ y_t, z_t=(z_t^id,z_t^act), -3.00003pt with identity activation ztidz_t^id (which profile and memory evidence is relevant here) and action decision ztactz_t^act (which target, action, and response follow from that evidence). Psychology-informed representation. We organize user evidence into two complementary groups: Pu=(Pugen,Puplat),P_u=(P_u^gen,P_u^plat), -3.00003pt where PugenP_u^gen contains 22 dimensions spanning Big Five personality, Schwartz values, moral foundations, and social dominance orientation (Goldberg 1990; Schwartz 1992; Graham et al. 2013; Pratto et al. 1994), while PuplatP_u^plat contains 19 dimensions spanning emotional tendencies, network and social capital, media literacy, and online behavior (Katz and Lazarsfeld 1955; Ellison et al. 2007; Guess et al. 2020; Wulczyn et al. 2017). These 41 variables are LLM-inferred, history-derived descriptors for behavior prediction, not validated psychometric or clinical measurements. General traits encode cross-context dispositions, whereas platform traits encode online interaction tendencies, allowing identity activation to select the evidence relevant to a specific thread. Profiles are built once from each userâs pre-split history, Pu=gprofâ(âutrain)P_u=g_prof(H_u^train), and reused for held-out instances to prevent post-split leakage. The complete taxonomy and annotation prompt are in the appendix. Memory retrieval. We use the current visible thread as a lexical query, rank each strictly pre-time historical behavior with BM25 (Robertson and Zaragoza 2009), and retain the top five as Mu,tM_u,t. The score uses neither dense similarity nor an additional recency prior. In the retrieval-size analysis, both 9B and 27B models peak at k=5k=5; fewer memories omit relevant history, whereas more introduce redundancy or noise (see the appendix). Identity activation and target-first action decision. Given the profile and retrieved memory, identity activation selects and summarizes the profile traits, historical behaviors, stance evidence, and behavioral constraints relevant to the current thread. Conditioned on this evidence summary, action decision selects the target before predicting the action and response plan. The decomposition provides explicit supervision for distinguishing Comment, Like, and Repost instead of treating every interaction as text generation. Together they yield the complete trajectory Ďt=(zt,yt)=(ztid,ztact,ct,at,rt). _t=(z_t,y_t)=(z_t^id,z_t^act,c_t,a_t,r_t). -3.00003pt The full schema and a worked example are in the appendix. 3.3 Two-Stage Training We train over the trajectory Ďt _t in two stages (Figure 2): supervised fine-tuning on reconstructed traces, then rubric-guided preference optimization over complete trajectories. Stage 1: Reverse-CoT supervision. Behavior logs provide observed actions and comments but no explicit annotation of the underlying decision process. Direct output supervision therefore leaves the intermediate mapping from user evidence to platform behavior implicit. We reconstruct the missing reasoning after the fact: given (xt,ytâ)(x_t,y_t^*), a teacher produces a posterior rationale constrained by the observed behavior, ztâ=fteacherâ(xt,ytâ)z_t^*=f_teacher(x_t,y_t^*). Since ztâz_t^* is generated after the behavior, we treat it not as ground-truth cognition but as structured supervision for learning the mapping from user evidence to observed behavior. Following behavior-conditioned user modeling paradigms (Zhang et al. 2025b), the reconstructed trace connects the user profile, retrieved memory, and thread context with the observed action. It is therefore intended to provide a plausible and behavior-consistent reasoning structure, rather than to recover a unique latent cognitive process. The observed behavior is used only during trace construction and is unavailable at inference time. A validation judge JtraceJ_trace (checking profile consistency, memory grounding, contextual relevance, identity-activation coherence, and action-decision coherence; Figure 2) keeps only consistent traces, SFT=(xt,ztâ,ytâ)âŁJtraceâ(xt,ztâ,ytâ)=1t=1N,D_SFT=\(x_t,z_t^*,y_t^*) J_trace(x_t,z_t^*,y_t^*)=1\_t=1^N, -3.00003pt and we minimize the negative log-likelihood of the full structured output, âSFT=â(x,zâ,yâ)âźSFTâlogâĄĎθâ(zâ,yââŁx),L_SFT=-E_(x,z^*,y^*) _SFT _θ(z^*,y^* x), -3.00003pt initializing the simulator to emit identity activation, action decision, and behavior in one trajectory. Judge criteria, the form of JtraceJ_trace, and prompts are in the appendix. Stage 2: Rubric-guided trajectory DPO. SFT yields well-formed trajectories but does not say which of several fluent ones is the more faithful simulation. We therefore optimize preferences over complete trajectories. The rubric itself is induced rather than hand-specified. We sample trajectory pairs and ask a judge to write free-form critiques explaining why one trajectory is a less faithful simulation than the other; an LLM then consolidates the full set of critiques into ten criteria that together account for the observed weaknesses, which the human annotators verify for coverage and non-redundancy. Grounding the evaluation axes in the weaknesses simulators actually exhibit, rather than in a priori intuitions, makes the rubric an explanatory cover of the failures rather than an arbitrary checklist. Once induced, the rubric is frozen and applied uniformly during preference optimization. For each input we compare a trajectory rolled out by the SFT model against a reference synthesized from the observed behavior, scoring both with a fixed ten-criterion social-behavior rubric (action fidelity, interaction-intensity calibration, context grounding, persona/memory/stance consistency, comment naturalness, specificity, format validity, and evidence-to-action coherence; full definitions in the appendix). Two choices matter. First, since action selection is central, action fidelity is up-weighted, SâĄ(Ď)=2âs1â(Ď)+âj=210sjâ(Ď)11,S(Ď)= 2s_1(Ď)+ _j=2^10s_j(Ď)11, -6.00006pt so the action is never outweighed by surface fluency. Second, the judge sees only the ground-truth action label aiâa_i^*, not the gold comment, forcing it to assess whether each trajectory chooses an appropriate action and uses evidence faithfully rather than matching reference wording. To avoid spurious supervision, we retain a pair only when the judge is confident, the score margin exceeds a threshold, the trajectories are not near-duplicates, and the preference is not explained by length. Algorithm 1 summarizes the full construction. Given the resulting DPO=(xi,Ďi+,Ďiâ)i=1MD_DPO=\(x_i, _i^+, _i^-)\_i=1^M, we optimize a DPO objective (Rafailov et al. 2023) over complete trajectories: the preference margin scores the full (ztid,ztact,ct,at,rt)(z^id_t,z^act_t,c_t,a_t,r_t) under Ďθ _θ relative to the SFT reference Ďref _ref with strength β, aligning both the intermediate reasoning and the platform behavior with the rubric. The complete rubric, trace-validation criteria, and judging prompts are in the appendix. Algorithm 1 Trajectory Preference Construction 0: input xix_i; SFT policy ĎSFT _SFT; observed behavior yiây_i^* with action label aiâa_i^*; rubric score SâĄ(â )S(¡) 1: ĎigenâźĎSFT(â âŁxi) _i^gen _SFT(¡ x_i) simulator rollout 2: Ďirefâ(ziref,yiâ) _i^refâ(z_i^ref,y_i^*) synthesize reference from observed behavior 3: score Ďigen,Ďiref _i^gen, _i^ref with SâĄ(â )S(¡), exposing aiâa_i^* to the judge but not the gold comment 4: (Ďi+,Ďiâ)â( _i^+, _i^-)â higher- / lower-scoring trajectory 5: if pair passes confidence, margin, duplication, and length filters then 6: add (xi,Ďi+,Ďiâ)(x_i, _i^+, _i^-) to DPOD_DPO 7: end if Models Behavior & Response Quality Profile Fidelity Act. Tgt. Sem. Intent Emo. Stance Style Memory Natural. Avg. Psych Trait GPT-oss-120B 54.63 37.35 34.49 41.50 43.39 51.39 37.70 52.14 45.95 42.92 73.12 73.78 GPT-5 61.80 58.80 46.99 54.44 54.22 65.12 46.36 60.79 58.57 54.12 76.52 75.23 Gemini-2.5-Pro 63.43 49.28 41.43 44.45 45.97 58.07 42.83 54.91 61.28 49.84 72.86 71.85 GPT-5.4 66.12 59.46 45.05 54.21 56.36 69.25 45.49 58.19 62.94 54.28 76.59 75.11 Claude-Sonnet-4.6 63.67 57.23 41.57 49.83 51.84 62.62 43.61 55.58 60.40 50.75 77.09 71.39 Qwen3.7-Max Preview 62.02 54.01 40.85 49.47 52.16 63.76 45.18 58.98 56.60 51.16 74.73 72.87 Gemini-3.1-Pro-Preview 62.81 51.61 37.53 46.26 50.24 60.77 48.12 57.66 60.02 50.26 75.91 73.72 DeepSeek-V4-Pro 64.32 47.01 36.62 45.15 47.36 59.94 40.70 52.96 55.19 46.86 75.63 71.35 Qwen3.5-9B 51.61 38.79 26.89 33.42 38.95 48.28 33.09 39.55 36.12 35.60 70.77 70.25 PsyMemBehav 9B 59.52 48.22 36.87 46.58 48.71 58.54 40.25 51.64 58.90 48.78 75.33 74.43 Qwen3.5-27B 56.57 47.90 33.87 39.84 42.23 49.43 40.30 54.52 49.62 44.26 72.23 71.02 PsyMemBehav 27B 71.40 65.34 50.35 55.49 58.89 68.89 53.57 64.19 66.39 59.68 80.38 78.25 Table 1: Individual-level user simulation results. Blue rows denote our proposed PsyMemBehav models. Act. denotes action accuracy over Comment, Like, and Repost; Tgt. denotes target accuracy. The seven response-quality dimensions are judged by Gemini-3-Flash. Psych and Trait are profile-fidelity scores computed by re-annotating generated behaviors and comparing them with the original user profile using 100Ă(1âNED)100Ă(1-NED). Higher is better for all metrics. Best results are bolded. Dimension Human mean Gemini Humanâ-Gemini Semantic 44.22 41.98 +2.24+2.24 Intent 48.46 50.84 â2.38-2.38 Emotion 50.94 54.54 â3.60-3.60 Stance 52.31 56.51 â4.20-4.20 Style 55.79 54.38 +1.41+1.41 Profile/memory 61.29 65.02 â3.73-3.73 Naturalness 67.77 64.92 +2.85+2.85 Overall 54.40 55.46 â1.06-1.06 Table 2: Human and Gemini mean scores on 150 instances from 30 anonymized user IDs (five per ID); differences are human minus Gemini. 4 Experimental Setup To evaluate the effectiveness of the proposed PsyMemBehav framework, we conduct experiments on the social-media user simulation task defined in Section 3. The model first selects the target, then predicts Comment, Like, or Repost and any required comment text. 4.1 Base Models We conduct experiments on two backbone models: Qwen3.5-9B and Qwen3.5-27B. The 9B model is used to examine whether the proposed framework remains effective under a smaller simulator, while the 27B model serves as the stronger backbone for our main experiments. Both models are trained under the same two-stage pipeline, resulting in two models: PsyMemBehav 9B and PsyMemBehav 27B. 4.2 Baselines We compare PsyMemBehav 9B and PsyMemBehav 27B against the following baselines: GPT-5.4 (OpenAI 2026), Gemini-2.5-Pro (Comanici and others 2025), Claude-Sonnet-4.6 (Anthropic 2025), Qwen3.7-Max (Yang and others 2025), and DeepSeek-V4-Pro (Xu et al. 2026a). Additionally, Qwen3.5-9B and Qwen3.5-27B (Yang and others 2025) are included as open-weight baselines. All models receive the same full user profile, BM25-retrieved pre-event memories, thread context, and candidate targets. They use the unified end-to-end prompt in the appendix, which selects a target before predicting the action and branch-specific text; gold target and action labels are withheld. 4.3 Evaluation Metrics We evaluate the models at both individual and group levels. Individual-level evaluation. We report two behavior-level metrics. Target Accuracy measures whether the model selects the correct interaction target among the candidate posts and comments. Action Accuracy measures whether the model chooses the correct platform actionâComment, Like, or Repostâfor the selected target. For generated comments, we use Gemini-3-Flash as an automated judge to score seven quality dimensions: semantic alignment, intent alignment, emotional consistency, stance consistency, style fidelity, persona memory, and naturalness. We report both individual dimension scores and their average. The complete scoring prompt appears in the appendix. Group-level simulation. To evaluate multi-user thread continuation, we place profiled simulators in the same observed thread seed and roll out their target-first actions. Following prior work that evaluates collective simulations through interaction, stance, and influence structure (Zhou et al. 2024; Liu et al. 2025), we report four complementary metrics. Reply Reciprocity is the share of interacting dyads with mutual directed replies, a standard property of directed networks (Garlaschelli and Loffredo 2004). Stance deviation is the â1 _1 distance between simulated and real Pro, Anti, and Neutral distributions, using the conventional favor/against/neither stance formulation (Mohammad et al. 2016). Top1 KOL is the reply share received by the most-addressed user, operationalizing key-opinion-leader concentration (Katz and Lazarsfeld 1955); Top5 PageRank is the normalized PageRank mass of the five highest-ranked users (Brin and Page 1998). MAE and MSE summarize absolute and squared deviations across thread-level statistics. Detailed formulas are in the appendix. 4.4 Dataset and Data Split We construct PsyMemBehav-bench from Reddit, Bluesky, and Twitter through official APIs. It contains 5.57M behaviors from 5,862 real users (approximately 950 per user), with usernames replaced by randomized hash identifiers. Platform events are normalized as Comment, Like, or Repost. Each instance combines a two-group profile, five BM25-retrieved pre-event memories, thread context, and candidate targets with the observed target, action, and response. A user-disjoint chronological split assigns 5,000 users to training and 862 to validation and testing; held-out users are unseen, and all profile and memory evidence predates prediction. Collection favors accounts with rich observable histories. Thus, the scale denotes longitudinal density rather than platform-wide representativeness and may overrepresent active users while underrepresenting low-activity, private, or new users. Among closely related studies reporting real-user counts, FineRob uses 1,866 users and 78.6K records (approximately 126 histories per user), whereas MOSAIC uses 204 real users and 30 posts and GA-S3 models group agents over 30 events. PsyMemBehav therefore provides larger and denser longitudinal evidence for user-level targetâactionâresponse modeling, not a claim of platform-wide coverage (see the appendix). Variant Act. Tgt. Avg. Psych Trait Training ablation Base 51.61 38.79 35.60 70.77 70.25 CoT-P 47.31 35.90 33.92 69.69 68.81 Out-SFT 49.23 40.30 39.70 71.73 70.34 R-CoT 53.61 46.64 43.37 74.58 75.79 R-CoT+R-DPO 59.52 48.22 48.78 75.33 74.43 Component ablation under R-CoT Full 53.61 46.64 43.37 74.58 75.79 w/o Gen. 52.48 43.77 39.66 70.71 74.69 w/o Plat. 52.39 42.32 41.13 73.11 71.45 w/o IdAct 51.27 41.03 38.77 69.43 72.50 w/o ActDec 50.47 40.62 39.37 69.89 71.83 Table 3: Ablation results on Qwen3.5-9B. Act. is action accuracy; Tgt. is target accuracy; Avg. is the average score over seven response-quality dimensions. Psych and Trait are profile-fidelity scores for general psychological traits and platform-specific traits. Model Reciprocity Stance Dev. â Top1 KOL Top5 PR MAE â MSE â Real 0.043 0.000 0.213 0.596 0.000 0.000 PsyMemBehav 27B 0.121 0.155 0.232 0.524 0.081 0.009 GPT-5.4 0.182 0.175 0.197 0.529 0.099 0.014 Claude-Sonnet-4.6 0.313 0.322 0.191 0.472 0.185 0.048 GPT-5 0.375 0.496 0.226 0.507 0.233 0.091 Qwen3.7-Max-Preview 0.458 0.452 0.131 0.483 0.266 0.099 DeepSeek-V4-Pro 0.444 0.649 0.136 0.448 0.319 0.152 Table 4: Thread-level continuation results. We use PsyMemBehav 27B as our simulator and compare its generated interaction threads with those produced by strong general models. Real denotes statistics computed from real comment sections. Reciprocity measures bidirectional reply interactions. Stance Dev. measures deviation from the real stance distribution. Top1 KOL and Top5 PR measure the concentration of influence in the simulated interaction graph. MAE and MSE summarize the overall deviation from real group-level statistics. Lower is better for metrics marked with â ; for other metrics, closer to the real distribution indicates better simulation fidelity. 5 Experimental Results and Analyses 5.1 Individual-Level Results Table 1 reports target and action accuracy, seven response-quality dimensions, and profile fidelity, computed by re-annotating generated behavior and applying 100Ă(1âNED)100Ă(1-NED) to the original profile (see the appendix). PsyMemBehav 27B ranks first in 11 of 12 metrics, with target accuracy of 65.34, action accuracy of 71.40, average response quality of 59.68, and profile-fidelity scores of 80.38 (Psych) and 78.25 (Trait). Relative to GPT-5.4, its average response quality rises from 54.28 to 59.68 (9.95% relative), alongside higher target and action accuracy. Training improves both backbones consistently: PsyMemBehav 9B raises average response quality from 35.60 to 48.78 over Qwen3.5-9B, and PsyMemBehav 27B raises it from 44.26 to 59.68 over Qwen3.5-27B, with concurrent gains in target accuracy, action accuracy, and profile fidelity. At 27B, the complete training pipeline improves action accuracy by 14.83 points, target accuracy by 17.44, average response quality by 15.42, Psych fidelity by 8.15, and Trait fidelity by 7.23 over the matched Qwen3.5 backbone. At 9B, the corresponding gains are 7.91, 9.43, 13.18, 4.56, and 4.18 points. The only metric not led by PsyMemBehav 27B is stance consistency: its 68.89 is 0.36 points below GPT-5.4âs 69.25. Gains therefore span decisions, responses, and profile preservation. Evaluation robustness. We score the same anonymized outputs with Qwen3.5-27B, GPT-5.4, Gemini-2.5-Pro, and Claude-Sonnet-4.6 as independent judges. PsyMemBehav 27B ranks first under all four, exceeding the runner-up by 7.73, 3.03, 4.90, and 2.83 points, respectively. Agreement across judge families reduces the likelihood that the main ranking is driven by one evaluatorâs self-preference, although it does not by itself establish human validity. Full cross-judge scores are reported in the appendix. Table 2 calibrates the automated evaluator against human judgments on 150 instances from 30 anonymized user IDs (five per ID). The overall means differ by 1.06 points (54.40 vs. 55.46), and the signed dimension gaps are mixed rather than uniformly favoring one evaluator. This close aggregate alignment supports model-level comparison, while the stance (4.20) and emotion (3.60) gaps motivate dimension-specific caution; full analysis appears in Appendix. 5.2 Ablation Study Table 3 separates training ablations from component ablations on Qwen3.5-9B. R-CoT/CoT-SFT is supervised fine-tuning on accepted reverse-CoT traces without DPO, whereas R-DPO is rubric-guided trajectory optimization. Naive CoT prompting degrades performance, and output-only SFT improves response quality without improving action prediction. R-CoT instead improves target accuracy, action accuracy, and profile fidelity; adding R-DPO further improves most behavior and response metrics. Relative to Base, the complete R-CoT+R-DPO pipeline raises target accuracy from 38.79 to 48.22, action accuracy from 51.61 to 59.52, and average response quality from 35.60 to 48.78. The two stages contribute differently. Relative to Base, R-CoT adds 7.85 points in target accuracy, 2.00 in action accuracy, 7.77 in average response quality, 3.81 in Psych fidelity, and 5.54 in Trait fidelity. Adding R-DPO contributes a further 1.58, 5.91, 5.41, and 0.75 points to target accuracy, action accuracy, response quality, and Psych fidelity, while Trait fidelity decreases by 1.36. Thus, reverse-CoT accounts for most target and trait gains; DPO mainly strengthens action selection and response quality despite a small trait-preservation trade-off. Removing either profile group degrades performance, confirming that general psychological and platform-specific evidence are complementary. Removing identity activation mainly reduces response quality, whereas removing action decision produces the lowest target and action accuracy, supporting their respective roles in evidence selection and target-first behavior prediction. The component losses also reveal specialization. Removing identity activation causes the largest drop in average response quality (â4.60-4.60) and Psych fidelity (â5.15-5.15), whereas removing action decision causes the largest losses in target accuracy (â6.02-6.02) and action accuracy (â3.14-3.14). Platform-specific traits contribute more to target accuracy (â4.32-4.32) and Trait fidelity (â4.34-4.34) than general traits (â2.87-2.87 and â1.10-1.10), while general traits contribute more to response quality (â3.71-3.71 versus â2.24-2.24). This pattern matches the intended division of labor among evidence and reasoning components. 5.3 Group-Level Thread Simulation Table 4 compares PsyMemBehav 27B with strong general models on continuations of the same observed thread contexts. PsyMemBehav 27B obtains the lowest stance deviation (0.155), MAE (0.081), and MSE (0.009), while its reciprocity (0.121) is closest to the real value (0.043). Its lower reciprocity error also avoids the over-interactive group-chat pattern produced by general-purpose agents. Relative to GPT-5.4, PsyMemBehav reduces stance deviation by 11.4%, MAE by 18.2%, and MSE by 35.7%. GPT-5 and GPT-5.4 remain closer on Top1 KOL and Top5 PageRank, respectively, so PsyMemBehav does not lead every structural statistic. In direct pairwise evaluation, PsyMemBehav is preferred over GPT-5.4 in 64.0% of thread continuations versus 34.0%, with 2.0% ties (Appendix). These results concern continuation within observed threads, not platform-wide social dynamics. PsyMemBehavâs absolute reciprocity gap from the real threads is 0.078, compared with 0.139 for GPT-5.4, a 43.9% reduction. For influence concentration, however, its Top1 KOL gap is 0.019 versus 0.013 for GPT-5, and its Top5 PageRank gap is 0.072 versus 0.067 for GPT-5.4. The strongest improvements therefore concern interaction reciprocity, stance composition, and aggregate fidelity, while centralization remains model-dependent. 6 Conclusion and Limitations We present PsyMemBehav, combining 41-dimensional user descriptors, BM25 memories, target-first reasoning, reverse-CoT supervision, and rubric-guided DPO. PsyMemBehav 27B achieves the strongest individual results and lowest aggregate deviation in observed-thread continuation, supporting evidence-grounded user models for micro-level public-opinion simulation. Limitations include missing exposure/no-action labels, LLM-inferred rather than self-reported profiles, answer-conditioned reasoning traces, and aggregate rather than per-instance human validation. Future work should address these issues across broader settings. Acknowledgments This work is supported by the National Natural Science Foundation of China (No. 62206259). References Anthropic (2025) Anthropic Claude sonnet 4.6. External Links: Link Cited by: §4.2. Argyle et al. (2023) L. P. Argyle, E. C. Busby, N. Fulda, J. R. Gubler, C. Rytting, and D. Wingate Out of one, many: using language models to simulate human samples. Political Analysis 31 (3), p. 337â351. External Links: Document Cited by: §1, §2.1. Bai et al. (2022) Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnon, et al. Constitutional AI: harmlessness from AI feedback. arXiv preprint arXiv:2212.08073. External Links: Document Cited by: §2.2. Brin and Page (1998) S. Brin and L. Page The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN Systems 30 (1â7), p. 107â117. External Links: Document Cited by: §E.2, §4.3. Cheng et al. (2026) X. Cheng, Y. Qin, Y. Tan, Z. Li, Y. Wang, H. Xiao, and Y. Zhang PsyMem: fine-grained psychological alignment and explicit memory control for advanced role-playing LLMs. Transactions of the Association for Computational Linguistics 14, p. 510â529. External Links: Link, Document Cited by: §1. Cheng et al. (2025) Z. Cheng, Z. Lin, Y. Yang, Z. Wei, and S. Chen Interactive simulation and visual analysis of social media event dynamics with LLM-based multi-agent modeling. Visual Informatics 9 (3), p. 100260. External Links: Document, Link Cited by: §1. Comanici et al. (2025) G. Comanici et al. Gemini 2.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261. Cited by: §4.2. Ellison et al. (2007) N. B. Ellison, C. Steinfield, and C. Lampe The benefits of facebook âfriends:â social capital and college studentsâ use of online social network sites. Journal of Computer-Mediated Communication 12 (4), p. 1143â1168. External Links: Document Cited by: §A.1, §3.2. Epstein (1999) J. M. Epstein Agent-based computational models and generative social science. Complexity 4 (5), p. 41â60. External Links: Document Cited by: §2.3. Gao et al. (2024) C. Gao, X. Lan, N. Li, Y. Yuan, J. Ding, Z. Zhou, F. Xu, and Y. Li Large language models empowered agent-based modeling and simulation: a survey and perspectives. Humanities and Social Sciences Communications 11 (1), p. 1â24. External Links: Document Cited by: §1, §1, §2.1, §2.3. Garlaschelli and Loffredo (2004) D. Garlaschelli and M. I. Loffredo Patterns of link reciprocity in directed networks. Physical Review Letters 93 (26), p. 268701. External Links: Document Cited by: §E.2, §4.3. Goldberg (1990) L. R. Goldberg An alternative âdescription of personalityâ: the big-five factor structure. Journal of Personality and Social Psychology 59 (6), p. 1216â1229. External Links: Document Cited by: §A.1, §3.2. Graham et al. (2013) J. Graham, J. Haidt, S. Koleva, M. Motyl, R. Iyer, S. P. Wojcik, and P. H. Ditto Moral foundations theory: the pragmatic validity of moral pluralism. In Advances in Experimental Social Psychology, Vol. 47, p. 55â130. External Links: Document Cited by: §A.1, §3.2. Guess et al. (2020) A. M. Guess, M. Lerner, B. Lyons, J. M. Montgomery, B. Nyhan, J. Reifler, and N. Sircar A digital media literacy intervention increases discernment between mainstream and false news in the united states and india. Proceedings of the National Academy of Sciences 117 (27), p. 15536â15545. External Links: Document Cited by: §A.1, §3.2. Hovy and Søgaard (2015) D. Hovy and A. Søgaard Tagging performance correlates with author age. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), Beijing, China, p. 483â488. External Links: Document, Link Cited by: §2.1. Hu et al. (2024) Y. Hu, G. Sherpa, L. Zhang, W. Li, Q. Bai, Y. Wang, and X. Wang An LLM-enhanced agent-based simulation tool for information propagation. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, p. 8679â8682. Note: Demo Track External Links: Document, Link Cited by: §1, §1. Jeon et al. (2025) M. S. Jeon, M. Mendoza, M. FernĂĄndez, E. Providel, F. RodrĂguez, N. Espina, A. Carvallo, and A. Abeliuk Simulating conversations on social media with generative agent-based models. EPJ Data Science 14 (79). External Links: Document, Link Cited by: §1, §1. Ji et al. (2026) J. Ji, L. Li, X. Cheng, Y. Zhou, W. Zhang, Y. Tan, and Y. Qin RippleMem: from isolated retrieval to associative recollection for long-term agent memory. arXiv preprint arXiv:2608.13334. Cited by: §1. Katz and Lazarsfeld (1955) E. Katz and P. F. Lazarsfeld Personal influence: the part played by people in the flow of mass communications. Free Press, Glencoe, IL. Cited by: §A.1, §E.2, §3.2, §4.3. Kim et al. (2024) S. Kim, J. Shin, Y. Cho, J. Jang, S. Longpre, H. Lee, S. Yun, S. Shin, S. Kim, J. Thorne, and M. Seo Prometheus: inducing fine-grained evaluation capability in language models. In The Twelfth International Conference on Learning Representations, Cited by: §2.2. Kosinski et al. (2013) M. Kosinski, D. Stillwell, and T. Graepel Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences 110 (15), p. 5802â5805. External Links: Document Cited by: §2.1. Lazer et al. (2009) D. Lazer, A. Pentland, L. Adamic, S. Aral, A. BarabĂĄsi, D. Brewer, N. Christakis, N. Contractor, J. Fowler, M. Gutmann, T. Jebara, G. King, M. Macy, D. Roy, and M. Van Alstyne Computational social science. Science 323 (5915), p. 721â723. External Links: Document Cited by: §2.3. Lee et al. (2023) H. Lee, S. Phatale, H. Mansoor, T. Mesnard, J. Ferret, K. Lu, C. Bishop, E. Hall, V. Carbune, A. Rastogi, and S. Prakash RLAIF vs. RLHF: scaling reinforcement learning from human feedback with AI feedback. arXiv preprint arXiv:2309.00267. Cited by: §2.2. Li et al. (2025) K. Li, C. Dai, W. Zhou, and S. Hu LLM-based behavior prediction for social media users with continuous memory. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, Mumbai, India, p. 459â474. External Links: Document, Link Cited by: Table 8. Liu et al. (2025) G. Liu, V. T. Le, S. Rahman, E. Kreiss, M. Ghassemi, and S. Gabriel MOSAIC: modeling social AI for content dissemination and regulation in multi-agent simulations. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Suzhou, China, p. 6390â6417. External Links: Document, Link Cited by: Table 8, §1, §1, §4.3. Liu et al. (2023) Y. Liu, D. Iter, Y. Xu, S. Wang, R. Xu, and C. Zhu G-eval: NLG evaluation using GPT-4 with better human alignment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Singapore, p. 2511â2522. External Links: Document, Link Cited by: §2.2. Macy and Willer (2002) M. W. Macy and R. Willer From factors to actors: computational sociology and agent-based modeling. Annual Review of Sociology 28, p. 143â166. External Links: Document Cited by: §2.3. Mohammad et al. (2016) S. Mohammad, S. Kiritchenko, P. Sobhani, X. Zhu, and C. Cherry SemEval-2016 task 6: detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), San Diego, California, p. 31â41. External Links: Document, Link Cited by: §E.2, §4.3. OpenAI (2026) OpenAI Introducing GPT-5.4. External Links: Link Cited by: §4.2. Ouyang et al. (2022) L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schulman, J. Hilton, F. Kelton, L. Miller, M. Simens, A. Askell, P. Welinder, P. Christiano, J. Leike, and R. Lowe Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, Vol. 35, p. 27730â27744. Cited by: §2.2. Park et al. (2023) J. S. Park, J. C. OâBrien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein Generative agents: interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST â23, New York, NY, USA. External Links: Document, Link Cited by: §1, §2.1, §2.3. Pratto et al. (1994) F. Pratto, J. Sidanius, L. M. Stallworth, and B. F. Malle Social dominance orientation: a personality variable predicting social and political attitudes. Journal of Personality and Social Psychology 67 (4), p. 741â763. External Links: Document Cited by: §A.1, §3.2. Qiu et al. (2025) Z. Qiu, H. Lyu, W. Xiong, and J. Luo Can LLMs simulate social media engagement? a study on action-guided response generation. arXiv preprint arXiv:2502.12073. External Links: Document, Link Cited by: Table 8. Rafailov et al. (2023) R. Rafailov, A. Sharma, E. Mitchell, S. Ermon, C. D. Manning, and C. Finn Direct preference optimization: your language model is secretly a reward model. In Advances in Neural Information Processing Systems, Vol. 36, p. 53728â53741. Cited by: §2.2, §3.3. Ren et al. (2024) R. Ren, P. Qiu, Y. Qu, J. Liu, W. X. Zhao, H. Wu, J. Wen, and H. Wang BASES: large-scale web search user simulation with large language model based agents. In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, p. 902â917. External Links: Document, Link Cited by: §2.1. Robertson and Zaragoza (2009) S. Robertson and H. Zaragoza The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval. External Links: Document, Link Cited by: §A.2, §3.2. Schwartz (1992) S. H. Schwartz Universals in the content and structure of values: theoretical advances and empirical tests in 20 countries. In Advances in Experimental Social Psychology, Vol. 25, p. 1â65. External Links: Document Cited by: §A.1, §3.2. Shao et al. (2023) Y. Shao, L. Li, J. Dai, and X. Qiu Character-LLM: a trainable agent for role-playing. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Singapore, p. 13153â13187. External Links: Link Cited by: §2.1. Tu et al. (2024) Q. Tu, S. Fan, Z. Tian, T. Shen, S. Shang, X. Gao, and R. Yan CharacterEval: a Chinese benchmark for role-playing conversational agent evaluation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Bangkok, Thailand, p. 11836â11850. External Links: Document, Link Cited by: §2.1. Wang et al. (2024) N. Wang, Z. Y. Peng, H. Que, J. Liu, W. Zhou, Y. Wu, H. Guo, R. Gan, Z. Ni, J. Yang, M. Zhang, Z. Zhang, W. Ouyang, K. Xu, W. Huang, J. Fu, and J. Peng RoleLLM: benchmarking, eliciting, and enhancing role-playing abilities of large language models. In Findings of the Association for Computational Linguistics: ACL 2024, Bangkok, Thailand, p. 14743â14777. External Links: Document, Link Cited by: §2.1. Wulczyn et al. (2017) E. Wulczyn, N. Thain, and L. Dixon Ex machina: personal attacks seen at scale. In Proceedings of the 26th International Conference on World Wide Web, p. 1391â1399. External Links: Document Cited by: §A.1, §3.2. Xiao et al. (2026) Z. Xiao, M. Li, X. Wang, L. Li, L. Shui, M. Ji, and B. Cai DREAM: llm-based dynamic role-playing via event-aware memory graph. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2, KDD â26, New York, NY, USA, p. 5686â5697. External Links: ISBN 9798400722592, Link, Document Cited by: §1. Xu et al. (2026a) A. Xu, B. Lin, B. Xue, B. Wang, B. Xu, B. Wu, B. Zhang, C. Lin, C. Dong, C. Ling, et al. Deepseek-v4: towards highly efficient million-token context intelligence. arXiv preprint arXiv:2606.19348. Cited by: §4.2. Xu et al. (2026b) Z. Xu, D. Chen, S. Wang, J. Li, C. Wang, M. Han, and Y. Wang AdaMARP: an adaptive multi-agent interaction framework for general immersive role-playing. In Findings of the Association for Computational Linguistics: ACL 2026, M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, p. 31230â31283. External Links: Link, Document, ISBN 979-8-89176-395-1 Cited by: §1. Yang et al. (2025) A. Yang et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388. Cited by: §4.2. Youyou et al. (2015) W. Youyou, M. Kosinski, and D. Stillwell Computer-based personality judgments are more accurate than those made by humans. Proceedings of the National Academy of Sciences 112 (4), p. 1036â1040. External Links: Document Cited by: §2.1. Zhang et al. (2025a) Y. Zhang, Z. Song, H. Zhou, W. Ren, Y. P. Chen, J. Yu, and W. Yang GA-S3: comprehensive social network simulation with group agents. In Findings of the Association for Computational Linguistics: ACL 2025, Vienna, Austria, p. 8950â8970. External Links: Document, Link Cited by: Table 8, §2.3. Zhang et al. (2025b) Z. Zhang, S. Liu, Z. Liu, R. Zhong, Q. Cai, X. Zhao, C. Zhang, Q. Liu, and P. Jiang LLM-powered user simulator for recommender system. Proceedings of the AAAI Conference on Artificial Intelligence 39 (12), p. 13339â13347. External Links: Document Cited by: §3.3. Zheng et al. (2023) 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 Judging LLM-as-a-judge with MT-bench and chatbot arena. In Advances in Neural Information Processing Systems, Vol. 36, p. 46595â46623. Cited by: §2.2. Zhou et al. (2024) X. Zhou, Z. Su, T. Eisape, H. Kim, and M. Sap Is this the real life? is this just fantasy? the misleading success of simulating social interactions with LLMs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Miami, Florida, USA, p. 21692â21714. External Links: Document, Link Cited by: §2.1, §2.3, §4.3. Appendix A User Representation and Memory Retrieval This technical supplement provides additional details on the user representation, memory retrieval, reasoning traces, evaluation rubric, qualitative examples, and prompt templates used in PsyMemBehav. We organize the supplement around the main components of the system: user profile construction, memory retrieval, trace construction, DPO preference construction, additional analyses, case studies, and prompts. A.1 Trait-based User Representation We represent each user with two complementary profile groups: Pu=(Pugen,Puplat).P_u=(P_u^gen,P_u^plat). The general psychological profile PugenP_u^gen captures relatively stable dispositions, including personality traits, values, moral preferences, and broad social orientation. These dimensions describe how a user tends to interpret information, express disagreement, evaluate social issues, and maintain persistent preferences across topics. The platform-specific profile PuplatP_u^plat captures online behavioral tendencies. These dimensions describe how the same user tends to behave in social-media environments, including opinion leadership, social capital, critical media literacy, and toxicity-related interaction patterns. This separation is important because two users may share similar psychological dispositions while expressing them differently on a platform: one may engage through careful correction, another through short ironic replies, and another through passive endorsement. The general psychological profile PugenP_u^gen comprises 22 dimensions across four categories. Big Five personality (Goldberg 1990) captures openness, conscientiousness, extraversion, agreeableness, and neuroticism. Schwartz values (Schwartz 1992) cover ten motivational value types: self-direction, stimulation, hedonism, achievement, power, security, conformity, tradition, benevolence, and universalism. Moral foundations (Graham et al. 2013) include care/harm, fairness/cheating, loyalty/betrayal, authority/subversion, and sanctity/degradation. Social dominance orientation (Pratto et al. 1994) provides two dimensions: SDO dominance and SDO anti-egalitarianism. The platform-specific profile PuplatP_u^plat comprises 19 dimensions across four categories. Emotional profile covers expressiveness, stability, negativity bias, empathy, anxiety tendency, and anger tendency. Network and social capital (Katz and Lazarsfeld 1955; Ellison et al. 2007) captures opinion leadership, network centrality, and social capital. Media literacy (Guess et al. 2020) includes critical media literacy, source evaluation, and misinformation susceptibility. Online behavior (Wulczyn et al. 2017) covers posting frequency, reply tendency, repost tendency, like tendency, toxicity, trolling tendency, and bot-likeness. The profile is produced from each userâs pre-split history âutrainH_u^train by an annotation model gprofg_prof. To avoid leakage, only behaviors before the prediction split are used for profile construction. Each stored profile contains numerical trait scores and evidence-grounded summaries. The numerical scores provide compact control variables, while the summaries preserve interpretable evidence for generation and analysis. A.2 Memory Retrieval For each prediction instance at time t, PsyMemBehav retrieves a compact memory set Mu,tM_u,t from the userâs previous behaviors: âu<t=hu,1,hu,2,âŚ,hu,n.H_u^<t=\h_u,1,h_u,2,âŚ,h_u,n\. We use BM25 lexical retrieval (Robertson and Zaragoza 2009). The current post together with its visible comment context forms the query qtq_t, and each historical behavior hu,jh_u,j is represented as a candidate document du,jd_u,j using its available textual content. The BM25 relevance score is sBM25â(du,j,qt) s_BM25(d_u,j,q_t) =âwâqtIDFâĄ(w)âfâĄ(w,du,j)â(k1+1)DâĄ(w,du,j), = _wâ q_tIDF(w) f(w,d_u,j)(k_1+1)D(w,d_u,j), DâĄ(w,du,j) D(w,d_u,j) =fâĄ(w,du,j)+k1â(1âb+bâ|du,j|avgdl). =f(w,d_u,j)+k_1 (1-b+b |d_u,j|avgdl ). where fâĄ(w,du,j)f(w,d_u,j) is the frequency of query term w in document du,jd_u,j, |du,j||d_u,j| is the document length, avgdlavgdl is the average candidate-document length, and k1k_1 and b are the standard BM25 term-saturation and length-normalization parameters. The final memory set keeps the top-ranked pre-time behaviors: Mu,t=TopKhu,jââu<tâĄsBM25â(du,j,qt).M_u,t=TopK_h_u,j _u^<ts_BM25(d_u,j,q_t). We do not use dense embedding similarity or an additional recency-prior term in the retrieval score. Temporal validity is enforced by restricting the candidate pool to behaviors observed strictly before t. Unless otherwise specified, we use k=5k=5 as the default retrieval size, balancing relevant historical evidence against redundant or noisy memories. A.3 User Profile Distribution To examine whether the constructed user profiles collapse into a narrow region or preserve meaningful diversity, we project the high-dimensional profile vectors into a two-dimensional space using principal component analysis. Figure 3 shows that the profile distribution is compact but non-collapsed. Most users are concentrated around a dense center, indicating shared behavioral regularities across the population. At the same time, peripheral low-density regions indicate long-tail users with more distinctive profile combinations. This supports the use of profile-conditioned simulation, since the population contains both central tendencies and individual heterogeneity. Figure 3: PCA density visualization of user profile representations. Each hexagonal bin represents the local density of users in the projected two-dimensional profile space. Appendix B Reasoning Schema and Trace Construction PsyMemBehav uses a two-level reasoning schema to connect user identity with concrete platform behavior. The intermediate trace is represented as zt=(ztid,ztact),z_t=(z_t^id,z_t^act), where ztidz_t^id summarizes the situated user identity and ztactz_t^act explains the selected target, the action conditioned on that target, and the resulting response when required. The identity-activation component identifies the salient user role in the current context. It includes activated psychological traits, activated platform-specific traits, retrieved memory evidence, likely stance, and behavioral constraints. For example, a user may act as a skeptical fact-checker in one thread, a supportive community member in another, and an ironic commentator in a third. The identity-activation stage is designed to capture this context-dependent but profile-consistent identity. The action-decision component maps the situated evidence summary into a target-conditioned platform behavior: ctâCt,atâComment,Like,Repost.c_tâ C_t, a_tâ\ Comment, Like, Repost\. It first selects the interaction target ctc_t and then predicts how to interact with that target through ata_t. The visible response rtr_t is planned only when the selected action is Comment. For Like and Repost, the model returns the selected target and action without generating unnecessary visible text. We construct reverse CoT traces by asking a teacher model to explain why the observed ground-truth action is plausible given the user profile, retrieved memory, and current thread. The teacher is not asked to predict a new action; instead, it produces an inverse reasoning trace aligned with the known target and action. This makes the reasoning trace a supervision signal for behavior explanation rather than a free-form alternative response. A trace validator JtraceJ_trace accepts a reconstructed trace only when it satisfies profile consistency, memory grounding, contextual relevance, role-reasoning coherence, and action-reasoning coherence: Jtraceâ(xt,zt,ytâ)=âk=15â[ckâ(xt,zt,ytâ)]â0,1.J_trace(x_t,z_t,y_t^*)= _k=1^51 [c_k(x_t,z_t,y_t^*) ]â\0,1\. Only accepted traces are used as training supervision. Appendix C Evaluation Rubric and DPO Preference Construction We use a ten-criterion rubric to evaluate simulated user behaviors and construct DPO preference pairs. The rubric covers four aspects: action-level fidelity, user-profile alignment, content quality, and format/reasoning validity. This design reflects the fact that social-media simulation is not only a text generation problem: the model must first select a valid interaction target, then choose how to interact with that target, and only then produce a response if the selected action requires visible text. Action-level fidelity includes action accuracy, interaction-intensity calibration, and target/thread grounding. User-profile alignment includes persona consistency, memory activation discipline, and stance consistency. Content quality includes response naturalness and specificity. Format/reasoning validity includes output format validity and CoT plausibility. Category Criterion Definition Action-level Fidelity R1. Action Fidelity Whether the candidate selects the correct high-level behavior type, such as Comment, Like, or Repost. LikeThread and LikeBasic are both treated as Like. R2. Interaction Intensity Whether the amount of engagement is appropriate. Overly strong interaction, such as writing a long comment when the user would only like, and under-interaction, such as producing a weak or incomplete response when a reply is expected, are penalized. R3. Target Grounding Whether the candidate understands the current post, comment tree, reply target, and local social context. User-profile Alignment R4. Persona Consistency Whether the behavior matches the userâs profile, long-term interests, values, directness, social habits, and emotional style. R5. Memory Discipline Whether historical comments are used only when helpful. The model is penalized both for ignoring clearly relevant memories and for forcing irrelevant memories into the reasoning. R6. Stance Consistency Whether the stance is compatible with the userâs prior positions and the current thread, without unsupported ideological invention or stance reversal. Content Quality R7. Response Naturalness Whether the visible response sounds like a real platform user rather than an assistant, summary, template, or overly polished explanation. R8. Specificity Whether the response addresses concrete contextual details rather than giving a generic or vague reply. Format and Reasoning R9. Format Validity Whether target, action, CoT, and visible response fields are complete, legal, parseable, and consistent with the candidate target list. R10. CoT Plausibility Whether the reasoning plausibly explains the action and response without over-rationalization, fabricated private psychology, or unsupported causal claims. Table 5: Evaluation rubric for simulated user behaviors. The rubric contains ten criteria grouped into four categories. The DPO judge receives the correct high-level action label but not the gold visible text. This prevents the judge from rewarding exact surface matching and encourages evaluation of target grounding, action fidelity, persona consistency, memory discipline, stance, naturalness, specificity, format validity, and reasoning plausibility. The weighted score is computed as SâĄ(Ď)=2âR1+âi=210Ri11,S(Ď)= 2R_1+ _i=2^10R_i11, where action fidelity is double-weighted because choosing the correct action family remains a central fidelity requirement after target selection. C.1 Rubric Activation Analysis Figure 4 reports how frequently each rubric dimension is activated in the annotated evaluation pairs. Persona consistency, CoT plausibility, and context grounding are the most frequently activated dimensions, while response naturalness, specificity, and memory consistency are also highly active. This shows that the evaluation is not merely action-based, but also depends on profile alignment, social-media realism, and reasoning plausibility. Figure 4: Activation counts of rubric dimensions over the evaluation set. The numbers above each bar show the absolute number of activated pairs and their percentage in the full annotated set. Appendix D Additional Analyses This section reports additional analyses that complement the main experimental results. We include platform-level action statistics, thread-level structural properties, the effect of memory retrieval size, judge robustness, and pairwise group-simulation preference results. D.1 Platform-level Simulation Statistics Figure 5 shows the distribution of simulated agent actions. Comments account for the largest proportion of actions, followed by reposts and likes. This indicates that the simulated platform is primarily driven by active conversational behavior rather than passive engagement. Figure 5: Distribution of simulated agent actions. Figure 6 presents thread-level interaction statistics, including comments per thread, maximum thread depth, and unique participants. Most discussions remain moderate in scale, while a small subset develops into deeper and more active interactions. This suggests that the simulator does not simply generate uniformly dense discussion trees, but instead produces a mixture of shallow and deeper interaction patterns. Figure 6: Statistical properties of simulated discussion threads. D.2 Effect of Memory Retrieval Size We vary the number of retrieved past comments k and evaluate the average generation quality. As shown in Figure 7, both the 9B and 27B models achieve the best performance when k=5k=5. Smaller values provide insufficient historical context, while larger values may introduce redundant or noisy memories. Therefore, we set k=5k=5 as the default retrieval size. Figure 7: Effect of memory retrieval size k on average generation quality. D.3 Judge Robustness LLM-as-a-judge evaluation may introduce evaluator-specific preferences. To examine whether our conclusion depends on a particular evaluator, we evaluate the same anonymized model outputs using multiple judge models from different model families. As shown in Table 6, PsyMemBehav 27B receives the highest score under all reported judges, suggesting that its advantage is not caused by a single evaluator or judge-family preference. Judge PsyMemBehav GPT-5.4 Gemini Claude Qwen3.5-27B 61.96 54.23 52.02 47.60 GPT-5.4 63.09 60.06 48.94 57.12 Gemini-2.5-Pro 72.50 67.60 62.75 61.33 Claude-Sonnet-4.6 47.27 40.68 33.88 44.44 Table 6: Judge robustness analysis. Each row corresponds to one judge model; each column corresponds to an evaluated model. PsyMemBehav 27B achieves the highest score under all four judges. We also conduct a direct pairwise group-simulation preference test between PsyMemBehav 27B and GPT-5.4. For each instance, both models are given the same real thread seed and asked to generate a group-level continuation. The better simulation is selected according to its fidelity to the real discussion dynamics. Model Win Rate GPT-5.4 34.0% PsyMemBehav 27B 64.0% Tie 2.0% Table 7: Pairwise group simulation preference test between GPT-5.4 and PsyMemBehav 27B. D.4 HumanâGemini Evaluation Agreement For the human evaluation, we selected 30 anonymized user IDs and evaluated five behavior instances per ID, yielding 150 annotated instances. Human evaluators and Gemini scored the same outputs under the same seven-dimensional rubric. Table 2 in the main paper compares their dimension-level means. The human and Gemini overall means are 54.40 and 55.46, respectively, a human-minus-Gemini difference of â1.06-1.06 points. Humans score semantic alignment, style fidelity, and naturalness slightly higher, whereas Gemini scores intent, emotion, stance, and profile/memory consistency higher. This dimension-mean comparison assesses aggregate calibration; it neither establishes per-instance agreement nor replaces a formal human reliability analysis. The largest absolute gaps occur for stance (4.20 points) and emotion (3.60 points), while all other dimensions differ by at most 3.73 points. Because the signs vary, Gemini does not show a simple global tendency to be uniformly more lenient or stringent. The dimension-specific gaps therefore motivate reporting human evaluation together with LLM-judge results and cross-judge robustness. D.5 Dataset Scale and Longitudinal Coverage Our use of large-scale does not imply platform-wide coverage. Instead, it refers to the joint scale of real-user coverage, total behavioral volume, and the density of longitudinal histories available for user modeling. We therefore characterize the resource more precisely as a large-scale, longitudinal, user-centered dataset. Table 8 compares PsyMemBehav-bench with closely related work along three dimensions: the number of real users represented, the behavioral evidence available for modeling each user, and the granularity of the prediction or simulation target. Study Real-user coverage Behavioral evidence Prediction/simulation target FineRob: LLM-Based Behavior Prediction for Social Media Users with Continuous Memory (Li et al. 2025) 1,866 users 78.6K fine-grained prediction records; 126 historical records per user on average Interaction object, behavior type, and content Can LLMs Simulate Social Media Engagement? A Study on Action-Guided Response Generation (Qiu et al. 2025) User count not reported 3,990 instances; 1â31 historical examples per user Engagement-action selection and response generation MOSAIC (Liu et al. 2025) 204 real users and 204 simulated personas 30 posts plus survey-derived personas Group-level content dissemination and regulation GA-S3 (Zhang et al. 2025a) Group agents rather than identifiable real-user histories 30 events over seven days Aggregate traffic and collective behavior PsyMemBehav-bench 5,862 real users 5.57M observed behaviors; approximately 950 behaviors per user User-level interaction target, target-conditioned action, and response Table 8: Scale comparison with closely related social-media behavior prediction and simulation studies. Here, real-user coverage denotes the number of distinct real users whose observed behavioral evidence is used to construct or evaluate the simulator. Behavioral evidence captures both the total amount of data and the density of the history available for each user. Prediction/simulation target distinguishes user-level targetâactionâresponse prediction from aggregate diffusion or traffic simulation. Among the closely related studies that report real-user counts, PsyMemBehav-bench contains more real users and substantially denser longitudinal behavioral histories. Its scale characterization is therefore based not on the number of users alone, but on the combination of 5,862 real users, 5.57 million observed behaviors, and approximately 950 behaviors per user. At the same time, these statistics should not be interpreted as platform-wide coverage or as a representative sample of all social-media users. The collection procedure favors observable users with sufficiently rich histories and may therefore overrepresent highly active accounts while underrepresenting low-activity, private, or newly created accounts. Platform access constraints, the selected observation period, and activity-based sampling can further shape the resulting population. These limitations should be considered when interpreting generalization beyond the studied users and contexts. Appendix E Evaluation Metric Definitions E.1 Profile Fidelity: Normalized Edit Distance Profile fidelity measures how faithfully a simulator reproduces a userâs behavioral profile. For each user, we re-annotate the simulatorâs generated behaviors using the same annotation model gprofg_prof to obtain a predicted profile vector P^u P_u, and compare it against the original profile PuP_u derived from ground-truth history. We report fidelity using 100Ă(1âNED)100Ă(1-NED), where NED is the Normalized Edit Distance between the discretized profile sequences: NEDâĄ(Pu,P^u)=EDâĄ(Pu,P^u)maxâĄ(|Pu|,|P^u|),NED(P_u, P_u)= ED(P_u, P_u) (|P_u|,| P_u|), where EDâĄ(â ,â )ED(¡,¡) denotes the Levenshtein edit distance and |â ||¡| denotes sequence length after discretizing continuous trait scores into ordinal bins. Because both reference and generated behaviors are interpreted through the same annotator, this score measures relative cross-system consistency under a fixed profiling lens; it should not be interpreted as absolute psychometric validity. A score of 100Ă(1âNED)=100100Ă(1-NED)=100 indicates a perfect match; higher values reflect closer alignment between simulated and real user profiles. We report two profile-fidelity scores separately: one for general psychological traits (PugenP_u^gen) and one for platform-specific behavioral traits (PuplatP_u^plat). E.2 Group-Level Simulation Metrics The group-level evaluation places multiple profiled simulators into real thread seeds and measures how closely the generated interaction structure matches real comment-section continuations. Reply Reciprocity (Rec.). Reciprocity is a standard directed-network measure of mutual linkage (Garlaschelli and Loffredo 2004). We compute the fraction of interacting user pairs (u,v)(u,v) for which both reply directions occur: Rec.=|(u,v):uâvâ§vâu||(u,v):uâvâ¨vâu|.Rec.= |\(u,v):uâ v vâ u\||\(u,v):uâ v vâ u\|. Overly high reciprocity indicates an unrealistic group-chat pattern rather than broadcast discussion. Stance Deviation (Stance Dev.). Following the conventional favor/against/neither formulation of stance (Mohammad et al. 2016), each comment is annotated as Pro, Anti, or Neutral. We measure the â1 _1 distance between simulated and real stance distributions: StanceDev=âsâPro, Anti, Neutral|psimâ(s)âprealâ(s)|.StanceDev= _sâ\Pro, Anti, Neutral\ |p_sim(s)-p_real(s) |. Lower values indicate closer reproduction of collective opinion composition. Attention Concentration (Top1 KOL / Top5 PR). Top1 KOL is the fraction of all received replies captured by the single most-addressed user, reflecting key-opinion-leader concentration (Katz and Lazarsfeld 1955). Top5 PR is the sum of normalized PageRank scores (Brin and Page 1998) for the five highest-ranked users in the directed reply graph. Together they test whether simulated attention and structural influence concentrate as in the real thread. Overall Deviation (MAE / MSE). MAE and MSE aggregate, respectively, absolute and squared per-thread deviations between simulated and real statistics. Lower values indicate closer overall group-level fidelity; MSE penalizes large mismatches more strongly. Appendix F Case Studies This section provides qualitative and group-level examples that complement the aggregate results in the main paper. We include both a single-reply comparison and a replay-style group simulation comparison. The single-reply case highlights how different models vary in stance commitment and response style under the same thread context. The group-level case shows that a model can produce plausible individual replies while still failing to reproduce the collective dynamics of a real discussion. F.1 Single-reply Comparison Table 9 compares different models under the same thread context and target comment. The real user reply is cautious and low-commitment. Several strong general-purpose models produce more socially reactive or stance-bearing responses, while PsyMemBehav 27B remains more hedged and balancing. Model Generated Reply Stance Observation Real user âI can see why people read the timing as suspicious, but it may also just be production timing.â Neutral Conservative and low-commitment stance. Gemini-3.1-Pro âThey could still be reworking it. This week gave them too much new material to ignore.â Neutral Cautious and relatively close to the real style. GPT-5.4 âThat timing feels awfully convenient. People are going to read into it no matter what.â Neutral / mildly charged More socially reactive and discussion-provoking. Claude-Sonnet-4.6 âThat explanation feels way too convenient given the pressure around this.â Pro / supportive More likely to clearly agree with a suspicious reading. Qwen3.7-Max âThat excuse is transparent. They probably got spooked by legal threats or the network.â Pro / charged Stronger stance and more speculative. PsyMemBehav 27B âThe reality is probably more mixed than this thread is making it.â Neutral More hedged and balancing, closer to the real userâs conservative stance. DeepSeek-V4-Pro âThere is a narrower point here, but the broad claim feels overcooked.â Anti / neutral Often gives a cooling or corrective reply. Table 9: Single-reply comparison under the same thread context. F.2 Group Simulation Comparison The group simulation setting starts from the same real thread seed and asks each model to generate a continuation involving multiple agents. Table 4 in the main paper reports aggregate interaction and network-level results over the evaluation set. Table 10 complements that aggregate evaluation with the stance distribution of this representative replay case. Each comment is assigned to one of three mutually exclusive stance classes: Pro explicitly supports the focal claim, Anti opposes it, and Neutral (Neu.) expresses neither clear support nor opposition; their proportions therefore sum to one. Charged (Chg.) is the share of stance-bearing comments, Chg.=pâĄ(Pro)+pâĄ(Anti)=1âpâĄ(Neutral)Chg.=p(Pro)+p(Anti)=1-p(Neutral). Distance (Dist.) is the â1 _1 distance between a modelâs stance vector and the real continuation, âs|psimâ(s)âprealâ(s)| _s|p_sim(s)-p_real(s)|; it ranges from 0 to 2, and lower is better. Entropy (Ent.) is normalized Shannon entropy, ââsp(s)logp(s)/log3- _sp(s) p(s)/ 3; it ranges from 0 (all comments in one stance) to 1 (a uniform three-way distribution). Because entropy measures stance diversity rather than accuracy, neither a larger nor a smaller raw value is intrinsically better; closeness to the real entropy indicates better fidelity. Model Pro Anti Neu. Chg. Dist. Ent. Real 0.085 0.085 0.830 0.170 0.000 0.523 Gemini-3.1-Pro 0.312 0.125 0.563 0.437 0.534 0.862 GPT-5.4 0.479 0.083 0.438 0.562 0.788 0.838 Claude-Sonnet-4.6 0.322 0.212 0.466 0.534 0.728 0.955 Qwen3.7-Max 0.417 0.062 0.521 0.479 0.664 0.798 PsyMemBehav 27B 0.182 0.121 0.697 0.303 0.266 0.744 DeepSeek-V4-Pro 0.423 0.135 0.442 0.558 0.776 0.906 Table 10: Stance-distribution comparison for a representative continuation from the same real thread seed. Neu. = Neutral; Chg. = Charged; Dist. = stance distance; Ent. = entropy. Aggregate interaction and network-level results are reported in Table 4 in the main paper; metric definitions are provided in the Group-Level Simulation Metrics subsection. Discussion. This representative case shows that strong general-purpose models generate substantially more stance-bearing responses than the real continuation. The real thread is predominantly neutral (0.830), with a charged share of 0.170. Among the simulated systems, PsyMemBehav 27B has the highest neutral share (0.697), the lowest charged share (0.303), and the smallest stance distance (0.266), making its stance distribution the closest to the real continuation. Its entropy (0.744) is also the closest to the real value (0.523). This case-level analysis complements, rather than duplicates, the aggregate thread-level comparison in Table 4 in the main paper. Appendix G Prompt Templates This section lists the main prompt templates used in PsyMemBehav. We include the prompts for CoT validation, action generation, reference CoT construction, pairwise DPO judging, and seven-dimension semantic evaluation. All prompt tables are shown as double-column tables for readability. A Reference Prompt for CoT Validation. ⏠1 You are a validator for generated hidden CoT in a real social media user memory activation dataset. 2 3 You will receive the system prompt, dialogue prefix, original assistant target, and a generated assistant value containing <think> plus the original target. 4 5 Your job is to decide whether the <think> is valid supervision for deciding when to use the scene memory pool. 6 7 === REFERENCE CONTEXT === 8 reference_content_with_memories 9 10 === ORIGINAL ASSISTANT TARGET === 11 target_assistant_content 12 13 === GENERATED ASSISTANT VALUE TO VALIDATE === 14 generated_reasoning 15 16 Check the following: 17 18 1. Structure: 19 - There must be exactly one <think> block before the assistant target. 20 - The <think> should contain exactly three natural paragraphs. 21 - It should not use explicit awareness/activation/decision tags, headings, bullets, or numbered step labels. 22 - The assistant target after </think> must exactly match the original assistant target. 23 24 2. Source grounding: 25 - Every fact in the <think> must be supported by the system profile, current scenario, dialogue prefix, Relevant Past Memories, Other Characters Relationships, or original assistant target. 26 - No fabricated scene details, motives, actions, or emotions. 27 28 3. Current-turn boundary: 29 - The <think> must explain only the current assistant turn. 30 - It must not use later dialogue or future plot knowledge unavailable to the character at this story point. 31 - If the system profile contains future events, the CoT should not use them unless the current character could know them. 32 33 4. Information leakage: 34 - The <think> must not leak information unavailable to the character at the current turn. 35 - The <think> must not reveal information that the visible assistant target withholds. 36 - The <think> may use the original assistant target to check stance alignment and response policy, but it must not quote, closely paraphrase, or explicitly pre-announce the targetâs distinctive surface wording, actions, or inner thoughts. 37 - Treat future-biographical details in the profile as unsafe unless the character could know or act from them at this story point. 38 39 5. Memory decision: 40 - Independently decide whether the scene memory pool is needed. 41 - Scene memory pool means only the listed Relevant Past Memories, not the character profile. 42 - If the current dialogue, current scenario, and profile are sufficient, memory should be marked not needed. 43 - If the target requires information absent from current context/profile but present in a relevant memory, memory should be marked needed. 44 - Check that the generated CoT names the correct memory ids when memory is needed. 45 - When memory is not needed, the CoT should explain why the current dialogue, current scenario, and/or profile are sufficient. 46 - If memory is needed, check that the generated CoT explains how each activated memory affects behavior in the current turn. 47 - If multiple memories are activated, check that the generated CoT compares their directions: aligned, opposed, or mixed, and explains the final net effect. 48 49 6. Profile vs memory: 50 - If the required information is in the profile, the CoT should not claim that a scene memory is needed. 51 - It may say profile knowledge is needed, but scene memory pool is not. 52 53 7. Answer leakage: 54 - The <think> must not quote or closely paraphrase the final assistant wording. 55 - It may describe response function, tone, and constraints, but not the final surface form. 56 57 8. Training usefulness: 58 - The <think> should clearly analyze the current information need, memory relevance, and response policy. 59 - It should not be a generic template or a repetition of the final answer. Table 11: A reference prompt for CoT validation. Unified End-to-End Behavior-Prediction Prompt. ⏠1 You are roleplaying as social-media user username. Given the user profile, retrieved pre-event history, current thread, and candidate targets, first select the interaction target. Conditioned on that target, predict the userâs recorded engagement as exactly one of Comment, Like, or Repost. Generate textual content only for Comment. Base both decisions on evidence in the input and do not sound like an assistant. 2 3 [Persona] 4 rendered_profile 5 6 [Retrieved Past Comments] 7 rendered_history_block 8 9 [Thread] 10 [post] @post_author -- "post_title" 11 post_text 12 13 [Comments] 14 - [cand_id_1] @author_1 (reply to [post]): "full_text_1" 15 - [cand_id_2] @author_2 (reply to [parent_id]): "full_text_2" 16 17 [Candidate Targets] 18 [post], [cand_id_1], [cand_id_2] 19 20 Rules: 21 - Select a valid target before choosing the action; neither the gold target nor the gold action is provided. 22 - Conditioned on the selected target, infer exactly one action from Comment, Like, Repost and keep the target fixed. 23 - Emit valid JSON for exactly one action and omit all fields belonging to the other actions. 24 - Comment: output action="Comment" and a reply object containing reply_to_id and natural, thread-specific text. 25 - Like: output action="LikeThread" and a target object containing id and kind; LikeThread and LikeBasic are evaluated as high-level Like. A like represents light endorsement, acknowledgment, or agreement. 26 - Repost: output action="Repost" and the same target object; a repost represents stronger public amplification than a like. 27 - Do not generate visible text for Like or Repost. Use history only as preference evidence and keep private reasoning concise and grounded. 28 29 Output exactly: 30 <think> 31 - Selected target: ... 32 - Predicted action/intensity for this target: Comment, Like, or Repost / ... 33 - Context grounding: ... 34 - Persona/memory/stance cue: ... 35 - Naturalness cue: ... 36 </think> 37 38 Then output exactly one valid <answer> branch; do not output branch labels. 39 Comment: <answer>"activity_type":"BehaviorPrediction","action":"Comment","reply":"reply_to_id":"<candidate id>","text":"..."</answer> 40 Like: <answer>"activity_type":"BehaviorPrediction","action":"LikeThread","target":"id":"<candidate id>","kind":"post | comment"</answer> 41 Repost: <answer>"activity_type":"BehaviorPrediction","action":"Repost","target":"id":"<candidate id>","kind":"post | comment"</answer> Table 12: Unified end-to-end behavior-prediction template. All inputs and target-first reasoning steps are shared; only the selected action and its required output fields differ. A Reference Prompt for Reference CoT Construction. ⏠1 # Action-DPO Reference CoT Prompt 2 3 You are writing inverse reasoning for a persona-simulation training example. 4 5 You are given: 6 - the userâs profile and retrieved history, 7 - the current post/thread context, 8 - the ground-truth action and answer block. 9 10 Your job is not to predict a new action. Your job is to explain, briefly and plausibly, why this specific user would take the provided ground-truth action in this situation. 11 12 ## What To Cover 13 14 Write a compact natural-language <think> block. It should cover: 15 - Action fit: why this action type is natural for the user here. 16 - Interaction intensity: why the amount of engagement is calibrated. 17 - Context grounding: what in the current post/thread creates the opening. 18 - Persona and memory: which profile/history signals matter. 19 - Stance and tone: the concrete stance or affective register implied by the action. 20 - Format/target: keep the selected target fixed and include visible target references only when the chosen action format requires them. 21 22 ## Constraints 23 24 - Do not invent private motives, fake memories, or unsupported ideology. 25 - Do not claim that the exact hidden ground-truth text is the only possible wording. 26 - For like/repost actions, do not over-explain as if the user wrote a comment. 27 - Keep the reasoning concise; this should be a training CoT, not an essay. 28 - Output only one <think>...</think> block. Do not output an <answer> block. 29 30 ## Output Format 31 32 <think> 33 Action fit: ... 34 Interaction intensity: ... 35 Context grounding: ... 36 Persona and memory: ... 37 Stance and tone: ... 38 Format/target: ... 39 </think> Table 13: A reference prompt for constructing inverse CoT supervision from a known ground-truth action. A Reference Prompt for Pairwise DPO Judging. ⏠1 You are a strict pairwise preference judge for persona-simulation DPO data. 2 3 You will receive: 4 - the target userâs profile and retrieved past comments; 5 - the current social-media thread and candidate targets; 6 - the correct high-level action type only: comment/reply, like, or repost; 7 - Candidate A and Candidate B, each containing CoT, target, action, and optional visible reply text. 8 9 Important: 10 - The ground-truth visible content is hidden. Do not assume exact wording. 11 - LikeThread and LikeBasic are both treated as high-level like. 12 - Judge target grounding first, then the action conditioned on that target, persona fit, memory use, stance, naturalness, specificity, format, and CoT plausibility. 13 - For comment/reply, judge whether the visible text is plausible for this user and context. 14 - For like/repost, penalize unnecessary visible comments. 15 - Do not reward length or polished assistant-like writing by itself. 16 17 Score each candidate from 1 to 5: 18 R1. Action Fidelity 19 R2. Interaction Intensity Calibration 20 R3. Target and Thread Grounding 21 R4. Persona Consistency 22 R5. Memory Activation Discipline 23 R6. Stance Consistency 24 R7. Response Naturalness 25 R8. Specificity 26 R9. Format Validity 27 R10. CoT Plausibility 28 29 Use the weighted score: 30 weighted_score = (2*R1 + R2 + R3 + R4 + R5 + R6 + R7 + R8 + R9 + R10) / 11 31 32 Choose A, B, tie, or drop. 33 Drop weak comparisons if: 34 - confidence < 4; 35 - score gap is too small; 36 - candidates are near-duplicates; 37 - the winner wins mainly by being longer; 38 - either candidate has invalid target, action, format, or malformed JSON; 39 - the CoT fabricates private psychology or forced memory. 40 41 Return JSON only: 42 43 "sample_id": "...", 44 "correct_action_type": "comment|like|repost|other", 45 "candidate_a_scores": 46 "R1": 1, "R2": 1, "R3": 1, "R4": 1, "R5": 1, 47 "R6": 1, "R7": 1, "R8": 1, "R9": 1, "R10": 1, 48 "weighted_score": 0.0 49 , 50 "candidate_b_scores": 51 "R1": 1, "R2": 1, "R3": 1, "R4": 1, "R5": 1, 52 "R6": 1, "R7": 1, "R8": 1, "R9": 1, "R10": 1, 53 "weighted_score": 0.0 54 , 55 "chosen": "A|B|tie|drop", 56 "confidence": 1, 57 "score_gap": 0.0, 58 "winner_reason": "brief evidence-bound reason", 59 "loser_weakness": "main concrete weakness", 60 "keep_for_dpo": true 61 Table 14: A reference prompt for pairwise DPO preference judging. A Reference Prompt for Seven-Dimension Semantic Evaluation. ⏠1 You are a strict, consistent, and objective evaluator for social media user simulation. 2 Your task is to assess how well the model-generated comment simulates the target userâs real response in the same social media context. 3 You will be given: 4 1. The original social media post 5 2. The target userâs history, profile, or memory information 6 3. The target userâs real comment, used as the ground truth 7 4. The model-generated comment 8 Important evaluation principles: 9 - The ground-truth comment is not the only possible correct response, but it represents the userâs actual reaction in this context. 10 - Do not judge only by lexical overlap. Evaluate semantic meaning, interaction intent, emotion, stance, style, persona/memory consistency, and naturalness. 11 - If the generated comment uses different wording but expresses a similar meaning, attitude, emotion, and interaction goal, it can receive a high score. 12 - If the generated comment is semantically relevant but does not match the userâs style, stance, or interaction intent, reduce the corresponding scores. 13 - If the generated comment fabricates the userâs personal history, identity, preferences, relationships, or experiences, significantly reduce the persona_memory_consistency score. 14 - If the generated comment sounds like AI writing, is overly formal, overly explanatory, or resembles a summary, reduce naturalness and style_fidelity scores. 15 - Do not give a high score simply because the generated comment is more complete or grammatically fluent. The goal is whether it sounds like the target user. 16 - Be strict. A score above 90 should only be given when the generated comment is extremely close to the target userâs real response. 17 - Avoid giving all dimensions similar scores. Evaluate each dimension independently. 18 [Input] 19 Original social media post: source_post 20 Target user history/profile/memory: user_profile_or_history 21 Ground-truth comment: ground_truth_comment 22 Model-generated comment: generated_comment 23 [Scoring Scale] Assign an integer score from 0 to 100 per dimension. 24 90-100: Nearly perfect match. 25 75-89: Strong match, only minor differences. 26 60-74: Moderate match, noticeable differences. 27 40-59: Weak match, substantial deviation. 28 20-39: Poor match. 29 0-19: Very poor; irrelevant, contradictory, or fabricated. 30 [Evaluation Dimensions] 31 1. semantic_alignment 32 Assess whether the generated comment expresses the same core meaning, topic, and information points as the ground-truth comment. 33 2. intent_alignment 34 Assess whether the generated comment fulfills the same social interaction purpose (e.g., supporting, opposing, joking, asking, answering, expressing emotion, sharing experience, giving advice). 35 3. emotion_consistency 36 Assess whether the generated comment matches the ground-truth in emotion type, direction, and intensity. 37 4. stance_consistency 38 Assess whether the generated comment expresses the same stance (supportive, opposed, skeptical, neutral, ironic, etc.) toward the post, event, or opinion. 39 5. style_fidelity 40 Assess whether the generated comment resembles the target userâs language style: length, tone, punctuation, emoji usage, slang, sentence patterns, formality, and humor style. 41 6. persona_memory_consistency 42 Assess whether the generated comment is consistent with the target userâs historical preferences, interests, values, and long-term behavioral tendencies, and avoids fabricating user facts. 43 7. naturalness 44 Assess whether the generated comment sounds like a real social media comment rather than AI-generated text, a summary, or an essay-like response. 45 [Overall Score] 46 overall_score = unweighted average of the seven dimension scores, rounded to two decimal places. 47 [Output] JSON only. No markdown. No explanatory preface. All dimension scores must be integers from 0 to 100. 48 49 "semantic_alignment": "score": 0, "reason": "", 50 "intent_alignment": "score": 0, "reason": "", 51 "emotion_consistency": "score": 0, "reason": "", 52 "stance_consistency": "score": 0, "reason": "", 53 "style_fidelity": "score": 0, "reason": "", 54 "persona_memory_consistency": "score": 0, "reason": "", 55 "naturalness": "score": 0, "reason": "", 56 "overall_score": 0.00, 57 "summary": "" 58 Table 15: A reference prompt for seven-dimension semantic evaluation.