TWICE: Modeling the Temporal Evolution of Personalized User Behavior via Event-Driven Agents

Baihan LI; Bingrui Jin; Kunyao Lan; Mengyue Wu

REVIEW 4 major objections 7 minor 40 references

TWICE: Modeling the Temporal Evolution of Personalized User Behavior via Event-Driven Agents

T0 review · 4 major / 7 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read TWICE simulates a user's tweets by conditioning on life events and time-aware memory, and the paper reports that this produces posts that match the user's evolving style and emotional state more closely than static-persona baselines.

desk verdict A sensible integration of profiling, event memory, and style rewriting, but the evaluation leaks the target tweet into the input and appears to use future posts, so the empirical claims don't hold. read the letter →

arxiv 2602.22222 v2 pith:3INXY2BO submitted 2025-12-15 cs.IR cs.MA

Bingrui Jin , Kunyao Lan , Baihan LI , Mengyue Wu This is my paper

classification cs.IRcs.MA

keywords usersimulationtemporaldynamicsevent-drivenmemoryLLMagentssocialmediamentalhealthstyleadaptationtweetingbehavior

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

TWICE is an attempt to make user simulation temporally real: rather than modeling a person as a fixed persona, it represents who the user is via a structured profile, organizes their history into time-windowed and event-grouped memory, and generates a tweet in two steps—deciding what to say from the event-grounded context, then rewriting it in the user's own voice. The paper claims this event-centered, time-aware design improves semantic similarity, style similarity, readability, and emotional alignment with real tweets across several LLM backbones, with the largest gains for users who have mental-health diagnoses. If correct, the result matters because it offers a concrete mechanism—retrieval with time decay and event-state weighting—that lets any LLM condition on longitudinal behavioral change, not just identity. The evaluation, however, gives the simulator the target tweet's own content as the trigger event and builds profiles without a temporal cutoff, so the reported gains are partly a measure of how well the model can exploit that supplied information.

What carries the argument

The load-bearing mechanism is the event-driven, time-aware memory module. History is aggregated into fixed-window chunks and into groups by detected life events and symptom categories; retrieval scores combine cosine similarity with the trigger event, an exponential time-decay factor, a per-tweet importance boost, and an event-state consistency coefficient. This scoring restores the causal chain from past experiences to the present moment. The second mechanism is the two-stage workflow: Stage I generates an event-grounded draft conditioned on profile, retrieved memory, and event summary; Stage II rewrites that draft using the user's Big Five traits, a natural-language style description, and

What would settle it

Re-run the headline comparison with a strict cutoff: build the profile and memory only from posts strictly before the target tweet's timestamp, and obtain the trigger event from a preceding post or a separate source. If semantic similarity and temporal-alignment improvements over the no-memory baseline largely vanish under this protocol, the central claim that event-centered memory drives the gain is falsified.

Watch

Extended reading notes

Core claim

The central claim is that separating 'what the user would say' from 'how this user would say it', and grounding both in memory organized around life events and symptoms with explicit time-decay retrieval, yields tweets that are semantically closer to the user's history, more stylistically consistent, more readable, and closer in emotional structure than generation from a raw timeline or a profile alone. The paper asserts this holds across backbone LLMs, and that it is especially clear for users with mental-health diagnoses, whose language signatures are more distinctive and therefore easier to track. The authors interpret this as evidence that event-centered memory is a promising mechanism f

Load-bearing premise

The evaluation assumes the trigger event can be extracted from the very tweet the model is supposed to predict, and that the user profile may include posts from after the simulated moment; if a strict temporal cutoff is enforced, the reported gains may shrink or disappear.

Editorial extensions

If this is right

If the claims hold, user simulators can move from static personas to temporally grounded agents that change how they express themselves as life events accumulate.
The framework is model-agnostic, so the event-memory retrieval recipe can be lifted onto any LLM backbone without retraining.
The explicit time-decay and event-state weighting gives a practical retrieval prior for any sequential user-modeling task, not just tweet generation.
The consistently stronger results for diagnosed users suggest event-triggered simulation could be a tool for mental-health research, generating plausible individual-level responses to life events.
Decoupling content from style means practitioners can dial the trade-off between factual alignment and personal voice, which is useful for assistant responses and recommendation explanations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A strict temporal holdout—building the profile and memory only from posts before the simulated timestamp and sourcing the trigger event from a preceding post rather than the target tweet—would test whether the gains reflect genuine temporal modeling or leakage of the target's content; the paper never enforces such a cutoff.
The same retrieval scoring could be adapted as a general attention prior over a user's timeline for non-generative tasks such as next-post prediction or churn modeling, where temporal coherence matters as much as semantic match.
The finding that diagnosed users are easier to simulate suggests simulatability tracks stylistic distinctiveness; a natural extension is to measure how TWICE's advantage shrinks for users with bland or inconsistent posting styles.
Because the memory module maintains event groups over years, it could support a long-running agent that updates a rolling event journal, enabling interactive evaluation over simulated months rather than one-shot generation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: TWICE is a reasonable recipe—structured profiles, event-driven memory, and a two-stage content-then-style workflow—and the writing is clear about what each piece does. The integration itself is new, even if the components are borrowed. The authors also deserve credit for using a large longitudinal dataset and for stating the limitation that no human evaluation was done. But the experimental evaluation, as designed, cannot support the paper's central claim that TWICE predicts temporally evolving user behavior.

The load-bearing problem is leakage. Appendix B.5 describes the procedure as: "We first perform event extraction on a post that contains significant events experienced by the user"—and that post is the ground-truth tweet the system is later asked to reproduce. So the event summary E_t fed into Stage I encodes the content of the target itself (e.g., "diagnosed with severe depression"). The generation is conditioned on (U, M≤t, E_t), meaning the model is essentially restating part of its input. That inflates semantic similarity and any temporal-alignment metric. Second, the user profile is built from the full timeline without a stated cutoff. The B.2 case study shows a profile containing 2020 events (father's death, COVID-19) while B.5 simulates a 2019-11-29 tweet. So U leaks post-t information as well. Even if one of these were fixed, the other would still be a serious concern; together they make the reported gains untrustworthy.

Also note the Abstract promises "strong baselines," but Table 2 only compares three backbone LLMs inside TWICE; there is no external user-simulator baseline. Table 3 is an ablation, not a baseline comparison. No significance testing is reported, and the style-similarity metric actually drops after the rewriting workflow—the authors attribute this to style reinforcement, which is honest but does not help the central claim.

That said, the framework itself is not implausible. A corrected prospective evaluation—event extraction from pre-event context only, a clear temporal cutoff for profile construction, and a real baseline such as a retrieval-augmented persona model—could make this a useful contribution. The paper is worth engaging with as a design proposal, not as evidence.

For peer review: I would send it out rather than desk-reject, because the system is clearly specified and the idea matters. But I would require major revision on the leakage issue before any acceptance. If the authors cannot fix the temporal integrity, the empirical claims should be withdrawn and the paper reframed as a system description with ablations.

Referee Report

4 major / 7 minor

Summary. TWICE is an LLM-based user-simulation framework for generating a user's tweet conditioned on a structured profile, an event-driven, time-aware memory, and a current event summary. It uses a two-stage generation pipeline: Stage I produces event-grounded content; Stage II rewrites it in the user's style. The paper evaluates on the Twitter-STMHD dataset with automated metrics (semantic similarity, style similarity, readability, emotion KL divergence) across mental-health cohorts and temporal parameter settings, and reports consistent improvements from the full framework.

Significance. If validated, TWICE would be a useful step toward temporally grounded user simulation. The architecture is modular and model-agnostic, the memory-retrieval scoring in Eq. (2) is clearly specified, and the evaluation attempts to cover authenticity, consistency, and humanlikeness. However, the current experimental protocol has a fundamental temporal-integrity problem: the event used to prompt generation is extracted from the very target tweet, and the user profile appears to include posts after the simulated timestamp. These issues undermine the quantitative claims. The paper also lacks any external user-simulator baseline despite the abstract's promise of 'strong baselines.' The framework may still be viable after a properly leakage-controlled evaluation, but that evaluation is not in the manuscript.

major comments (4)

[§2.3, Appendix B.5] Equation (1) conditions generation on E_t, but Appendix B.5's event-extraction procedure is applied to the ground-truth post the model is asked to reproduce: 'We first perform event extraction on a post that contains significant events experienced by the user,' where that post is Tweet ID 1200231490409373698. The extracted E_t therefore contains the target content ('diagnosed with severe depression'), so Stage I is not predicting y* but restating an input. The semantic-similarity and temporal-alignment gains in Table 3 and Figure 3 are partly artifacts of this leakage. To support the central claim, event extraction must be performed only on information available before the target timestamp, with the target itself held out.
[§2.1, Appendix B.2/B.5] The user profile appears to be constructed over the full timeline without a temporal cutoff before the simulated event. Appendix B.2's profile includes 'father dies from cancer/pneumonia in August' and COVID-era societal events, while Appendix B.5 simulates a tweet from 2019-11-29. If U is built from the full timeline, it leaks post-t information that can inflate personalization and coherence measures. The paper never states a cutoff in §2.1 or §3. Profile construction must be restricted to tweets before time t, and the age/style description must not use future posts.
[Abstract, §3, Tables 2–3] The abstract claims TWICE 'consistently outperforms strong baselines,' but Table 2 compares only backbone LLMs inside TWICE and Table 3 is an ablation of workflow components. No prior user-simulation baseline (e.g., static persona, recency-only memory, chronological retrieval, or an existing LLM user simulator) is compared. Consequently, the paper's central assertion that event-centered memory is a promising mechanism is unsupported. Add at least one non-event-driven baseline while keeping the same backbone LLM and evaluation protocol.
[§2.4, Eq. (3), Table 3] The style-consistency metric compares the simulated tweet with the user's historical posts, but the profile and style exemplars used in generation are derived from those same historical posts. This creates a further risk of inflation: the model is explicitly given the user's style description and representative tweets, so high style similarity may reflect direct copying of the conditioning material rather than genuine simulation ability. Report style similarity with held-out style exemplars or with a lexical-overlap control, and show that the gains survive when the profile is built only from pre-t data.

minor comments (7)

[Throughout] The header abstract and the full-text abstract differ; the header promises 'strong baselines' and 'temporal evolution of personalized user behavior,' while the full text is more modest. Please align the two and ensure claims match the actual experimental design.
[Table 3] The meaning of 'Original Workflow' versus 'Workflow' columns is not defined in the text. Clarify whether 'Original' refers to single-stage generation without rewriting, a backbone-LLM direct generation, or something else.
[Figure 3] Axes and metric definitions are not labeled in the figure caption, and the four users are not described. Without axis labels and per-panel metric names, the claimed trends cannot be verified.
[§3, RQ1 settings] 'Time-weighted event sampling' is mentioned but not fully specified. Provide the exact algorithm or formula used to compute sampling weights.
[Section 2.4, Eq. (6)] The KL divergence uses softmax(VAD) vectors, but the definition of the VAD distribution is not stated. If VAD values are continuous, clarify how they are discretized or normalized.
[Table 4] Readability differences are reported as signed values (e.g., FRE −7.4792), while the text says lower absolute differences are better. Adding absolute values or a sign convention would reduce confusion.
[Limitations] The paper acknowledges the absence of human evaluation. This is appropriate, but given the automated-metric leakage issues above, human evaluation or a leakage-free automated evaluation is necessary before the main claims can be accepted.

Circularity Check

2 steps flagged · score 7.0 of 10

Event extraction is performed on the ground-truth tweet and profiles include post-event posts, so the 'prediction' is conditioned on its target.

self definitional [Appendix B.5 (Simulation Result); Eq. (1) in §2]
"We first perform event extraction on a post that contains significant events experienced by the user, and the extracted events are summarized as sudden events, with the original post's posting time recorded as the time of the occurrence of the event. ... Original User's Tweet ... Timestamp:2019-11-29 ... Text:Update: I went to the doctor ... diagnosed with a wombo combo of anxiety and severe depression ... Extracted Event: Health - Severe Depression ... <User><was diagnosed with><severe depression>"

The 'post that contains significant events' in B.5 is the ground-truth tweet the simulator is supposed to output. The extracted event E_t restates that tweet ('diagnosed with severe depression'), and Eq. (1) conditions generation on E_t. Thus the model is given a summary of its own target and produces a style-conditioned paraphrase; the semantic-similarity and temporal-alignment gains in Table 3/Figure 3 are baked in by construction rather than measuring prediction of unseen behavior.
other [Appendix B.5 (User's Event Profile); §2.1]
"Original User's Tweet ... Timestamp:2019-11-29 ... User's Event Profile ... Death:Lose oldest friend nine months before pandemic, father dies from cancer/pneumonia in August, struggle with second birthday of friend who passed, reflect on six months since father's death ... Societal:Support Peruvian Lives Matter and protest against government actions"

The profile U is constructed over the full timeline with no cutoff stated at the simulated event time. The event to simulate is dated 2019-11-29, but the profile contains events that occurred later ('pandemic', 'Peruvian Lives Matter', 'six months since father's death'). Conditioning on U therefore leaks post-t information into the simulated tweet, inflating personalization and coherence metrics and making the 'temporal prediction' partly retrospective.

full rationale

The central circularity is direct: the event summary E_t used as conditioning in Eq. (1) is extracted from the very tweet the framework is asked to predict (Appendix B.5). This makes the reported semantic-similarity and temporal-alignment improvements partly a paraphrase effect rather than evidence of predictive simulation. A second, independent leakage path is the user profile, which is built from the full timeline without a stated cutoff before the simulated event; the B.5 case study supplies a profile containing post-2019 content while simulating a 2019-11-29 tweet. The paper's own Limitation section concedes the absence of human evaluation, and the 'strong baselines' claim is not directly supported by the tables (Table 2 compares backbone LLMs; Table 3 is an internal ablation), but these are evidence gaps rather than additional circular steps. The self-cited classifiers from Chen et al. 2024 and Zhang et al. 2022 are load-bearing for profile and event construction, but they are not the main source of circularity; the main source is the target tweet being used as input. Overall, the headline claim that TWICE predicts temporally evolving user behavior is substantially undermined by construction.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The central evaluation rests on several hand-set retrieval hyperparameters, the validity of prior classifiers reused without re-validation, and an implicit assumption that no temporal leakage occurs. These assumptions are not independently verified.

free parameters (10)

decay_lambda = 0.001
Time-decay rate in the memory retrieval score (Eq. 2); set by hand in §3.
state_coeff = 1, 1.1
Event-state consistency multiplier w_state in Eq. 2; varied as an experimental setting.
time_window = 1 year
Temporal window [t_window, t) for memory retrieval; fixed in RQ1, varied in RQ2.
node_num = 3
Number of memory nodes expanded during retrieval; set in §3.
memory_num = 10, 20
Number of memory entries collected for generation; set in §3.
importance_scale_k
Importance scaling factor in Eq. 2; no value is given in the text.
importance_boost_beta
Boost added to tweet importance after retrieval (imp ← imp + β); no value is given.
threshold_p
Hard threshold on life-event/symptom classifier scores (§2.1); not reported.
profile_sampling_count = 977
Number of representative profiles sampled from 34,330 via density-aware sampling (§2.1).
style_exemplars_per_iteration = 20
Number of representative posts selected per iteration for style description (§2.1).

assumptions (6)

domain assumption Twitter-STMHD self-reported diagnoses are accurate labels for mental-health status.
RQ3 comparisons treat POS/NEG groups as valid; no independent validation of self-reports is provided.
domain assumption Life-event/symptom classifiers from Chen et al. 2024 and Zhang et al. 2022 transfer reliably to this corpus.
These classifiers are used to build event memory and profile groups; no re-validation on the full dataset is reported.
domain assumption Sentence-embedding similarity to a user's historical tweets is a valid measure of authenticity.
Authenticity metric defined in §2.4; correlates with generic similarity rather than ground-truth prediction.
domain assumption FRE/FKGL and VAD KL divergence measure humanlikeness.
Readability and emotional-expression-complexity metrics in §2.4; not validated against human judgments.
domain assumption The event extracted from a target post equals the external stimulus that caused the post.
The pipeline assumes the event summary is the cause, but it is inferred from the post itself (Appendix B.5), creating leakage.
domain assumption Profile construction over the full timeline does not introduce future information.
Implied by §2.1; contradicted by Appendix B.2, where the profile contains events after the simulated tweet's date.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TWICE: Modeling the Temporal Evolution of Personalized User Behavior via Event-Driven Agents." pith.science (2026). https://pith.science/paper/3INXY2BO

@misc{pith2026260222222,
  author       = {Pith},
  title        = {Pith review of: TWICE: Modeling the Temporal Evolution of Personalized User Behavior via Event-Driven Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3INXY2BO}},
  note         = {Machine review of arXiv:2602.22222}
}

read the original abstract

User simulators are widely used for data generation, evaluation, and agent-based interaction, but existing approaches often model users as static personas or rely on generic historical context, making it difficult to capture how individual behavior evolves over time. To address this limitation, we propose TWICE, an LLM-based framework for temporally grounded personalized user simulation. TWICE combines structured user profiling, an event-driven memory module organized around life events and behavioral shifts, and a two-stage workflow separating event-grounded content planning from personalized style adaptation. This design enables the simulator to model not only what a user says, but also how past experiences shape later expression. We evaluate TWICE on a large-scale longitudinal Twitter dataset and introduce a comprehensive evaluation framework that jointly measures authenticity, consistency, and humanlikeness. Results show that TWICE consistently outperforms strong baselines, suggesting that event-centered memory is a promising mechanism for modeling the temporal evolution of personalized user behavior.

Figures

Figures reproduced from arXiv: 2602.22222 by the authors.

**Figure 1.** LLM Agent Tweeting Simulation Framework TWICE. This framework integrates personalized user profiling, an event-driven memory module, and a workflow for personalized style rewriting. semantic and stylistic alignment; RQ2 (Temporal Alignment)—event-driven memory improves timeline coherence; RQ3 (Impact of Mental Health Status)-simulation quality between users with and without mental health disordrs. Metrics span con… view at source ↗

**Figure 2.** The Procedure of Profile Generation.The profile adopts a two-tier structure that covers both general and personalized attributes Sampling To generate diverse and realistic user profiles, we apply a density-aware sampling strategy based on Wang (Wang et al. 2025). We embed 34,330 user profiles into a semantic space using SimCSE (Gao, Yao, and Chen 2021), incorporating both objective and subjective features. These e… view at source ↗

**Figure 3.** Performance Across Four Users’ Simulated [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 3 linked inside Pith

[1]

Provide a qualitative rating of Low, Medium, or High based on the language patterns, emotional tone, and content in the user’s tweets
[2]

Provide a detailed explanation (1-3 sentences) jus- tifying the rating for the !{dimension}! trait
[3]

Focus exclusively on the user’s characteristics, dis- regarding any information related to others, unless it directly impacts the user
[4]

!{dimension}! trait definition: !{definition}!
[5]

Infer Marital Status Prompt Infer the latest marital sta- tus(married/divorced/single/widowed/unknown) of the user until now(2021.1.1)

Please strictly output a JSON object with the fol- lowing format, no additional text: { ”score”: ”Low/Medium/High”, ”explanation”: ”Explanation referencing specific tweet evidence.” } ONLY output valid JSON, WITHOUT any extra characters or explanations. Infer Marital Status Prompt Infer the latest marital sta- tus(married/divorced/single/widowed/unknown) ...

2021
[6]

Transportation and Logistics: Focus on the efficient movement of goods and people, including transportation planning, logistics management, and supply chain operations
[7]

Manufacturing and Construction: Relate to the production of physical goods, construction of build- ings and infrastructure, and managing the process of creating tangible products or structures
[8]

yesterday

Hospitality and Tourism: Focus on providing services related to travel, accommodation, and leisure, including hotels, restaurants, tour operations, and event planning. Please strictly output a JSON object with the follow- ing format, no additional text: { ”career domain”: int(0/1/2/3/4/5/6/7/8), ”explanation”: ”your explanation” } ONLY output valid JSON, ...

Show all 40 references

[9]

Creative Arts and Media: Focus on artistic expres- sion and media production, including visual arts, film, music, design, and journalism, where creativity and communication are key
[10]

Business and Finance: Centered around managing financial assets, business operations, marketing, investment, and sales, driving economic growth and organizational success
[11]

Technology and Engineering: Involve applying scientific principles to design and develop technol- ogy, software, infrastructure, and solutions to solve complex problems
[12]

Healthcare and Social Services: Focus on pro- viding medical care, mental health support, and social welfare services to improve the well-being of individuals and communities
[13]

Education and Research: Dedicate to teaching, academic research, and fostering learning and development through formal and informal education systems
[14]

Legal and Public Policy: Concern with law en- forcement, legal counsel, and shaping public policy to ensure justice, governance, and the protection of rights
[18]

The events you identify must be clearly related to !{event}! and must be directly experienced by the user — do not include posts that only describe someone else’s experience
[19]

The tweets must not be repeated expressions of the same event on the same day — you can refer to the timestamp tweet to ensure the tweets are from different days
[20]

Your goal is to identify clusters or sequences of related events, and briefly explain the relationship between them
[21]

Your output should include three parts: the tweet id of the tweets you find; the brief conclusion of the event that these tweets related to; your expla- nation
[22]

Please strictly output a JSON object with the following format, no additional text: { ”tweet id”: [tweet id 1(int), tweet id 2(int), ..., tweet id n(int)], ”event conclusion”: ””, ”explanation”: ”” } ONLY output valid JSON, WITHOUT any extra characters or explanations
[23]

Simulated Tweet Generation Prompt You are a twitter user

If there aren’t any tweets that are highly related to each other and refer to similar events, your response should be: { ”tweet id”: None, ”event conclusion”: None, ”explanation”: None } A.3 Simulating Posting Workflow TheSimulating Posting Workflowinvolves generating tweets s...
[24]

Maintain the same general content and message as the original tweet
[25]

Modify the tone, word choice, structure, and phrasing to reflect the user’s unique personality and posting style (if you know)
[26]

Make sure the rewrite feels authentic to the user, as if it could have been written by them
[27]

The rewritten tweet should match the user’s personality traits, past posting history (if you know), and overall style summary (if you know). Please strictly output a JSON object with the follow- ing format, no additional text: { ”rewritten tweet”: ” ”, ”explanation”: ”your exp...

2010
[30]

Despite my apprehensions, I went to see my ther- apist for the first time today. User Role:Experiencer Tweet 1: First Appointment with Therapist Tweet ID:1285264241784758274 Timestamp:2020-07-20 17:24:08+00:00 Language:English Source:Twitter for iPhone Likes Count:2 Quote Coun...

2020
[31]

I attended my first therapy session today despite initial hesitation
[32]

Today, I went to my first appointment with a ther- apist
[33]

Despite my apprehensions, I went to see my ther- apist for the first time today. User Role:Experiencer Tweet 2: Taking Appointment with Therapist Tweet ID:1283722480364990465 Timestamp:2020-07-16 11:17:44+00:00 Language:English Source:Twitter for iPhone Likes Count:0 Quote Cou...

2020
[34]

I finally made an appointment with a therapist af- ter postponing for two years
[35]

After two years of procrastination, I scheduled a therapy session
[36]

I’m seeing a therapist today, something I’ve put off for two long years. User Role:Initiator Tweet 3: Planning to Find a Therapist Tweet ID:1211373980080386048 Timestamp:2019-12-29 19:50:37+00:00 Language:English Source:Twitter for iPhone Likes Count:2 Quote Count:0 Reply Coun...

2019
[37]

In 2020, I plan to find a therapist because I just can’t anymore

2020
[38]

Next year, I’m going to look for a therapist since I’ve reached my limit
[39]

I’ve decided to see a therapist at the start of 2020 due to my current struggles. User Role:Initiator Tweet 4: Struggles with Seeking Therapy Tweet ID:1050471916648185856 Timestamp:2018-10-11 19:43:16+00:00 Language:English Source:Twitter for iPhone Likes Count:0 Quote Count:0...

2020
[40]

I need to see a therapist again but I’m scared
[41]

It’s been two years since I last saw a therapist and I know I need to start again
[42]

User Role:Experiencer B.4 Memory We designed two types of memory: general memory and event memory

Seeing a therapist helped me before, but now I’m afraid to take that first step. User Role:Experiencer B.4 Memory We designed two types of memory: general memory and event memory. General memory is aggregated by time, where posts from the same month form a memory node. Event m...

2020
[2018]

Gao, C.; Lan, X.; jie Lu, Z.; Mao, J.; Piao, J.; Wang, H.; Jin, D.; and Li, Y

What about Mood Swings: Identifying Depression on Twitter with Temporal Measures of Emotions.Companion Proceedings of the The Web Conference 2018. Gao, C.; Lan, X.; jie Lu, Z.; Mao, J.; Piao, J.; Wang, H.; Jin, D.; and Li, Y . 2023. S3: Social-network Simulation System with La...

2018 arXiv
[2019]

Journal of Medical Internet Research, 21

Development and evaluation of ClientBot: A patient- like conversational agent to train basic counseling skills. Journal of Medical Internet Research, 21. Wang, K.; Li, X.; Yang, S.; Zhou, L.; Jiang, F.; and Li, H. 2025. Know You First and Be You Better: Modeling Human-Like Use...

2025 arXiv
[2023]

Louie, R.; Nandi, A.; Fang, W.; Chang, C.; Brunskill, E.; and Yang, D

Large Language Model as a User Simulator.ArXiv, abs/2308.11534. Louie, R.; Nandi, A.; Fang, W.; Chang, C.; Brunskill, E.; and Yang, D. 2024. Roleplay-doh: Enabling Domain-Experts to Create LLM-simulated Patients via Eliciting and Adhering to Principles. In Al-Onaizan, Y .; Ban...

2024 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Provide a qualitative rating of Low, Medium, or High based on the language patterns, emotional tone, and content in the user’s tweets

[2] [2]

Provide a detailed explanation (1-3 sentences) jus- tifying the rating for the !{dimension}! trait

[3] [3]

Focus exclusively on the user’s characteristics, dis- regarding any information related to others, unless it directly impacts the user

[4] [4]

!{dimension}! trait definition: !{definition}!

[5] [5]

Infer Marital Status Prompt Infer the latest marital sta- tus(married/divorced/single/widowed/unknown) of the user until now(2021.1.1)

Please strictly output a JSON object with the fol- lowing format, no additional text: { ”score”: ”Low/Medium/High”, ”explanation”: ”Explanation referencing specific tweet evidence.” } ONLY output valid JSON, WITHOUT any extra characters or explanations. Infer Marital Status Prompt Infer the latest marital sta- tus(married/divorced/single/widowed/unknown) ...

2021

[6] [6]

Transportation and Logistics: Focus on the efficient movement of goods and people, including transportation planning, logistics management, and supply chain operations

[7] [7]

Manufacturing and Construction: Relate to the production of physical goods, construction of build- ings and infrastructure, and managing the process of creating tangible products or structures

[8] [8]

yesterday

Hospitality and Tourism: Focus on providing services related to travel, accommodation, and leisure, including hotels, restaurants, tour operations, and event planning. Please strictly output a JSON object with the follow- ing format, no additional text: { ”career domain”: int(0/1/2/3/4/5/6/7/8), ”explanation”: ”your explanation” } ONLY output valid JSON, ...