Paper deep dive

Same Stories, Different Journeys: From Social Comparison to Sensemaking in AI-Mediated Peer Career Exploration

Pengping Tan, Baoquan Zhao, Zhenhui Peng

Year: 2026Venue: arXiv preprintArea: cs.HCType: PreprintEmbeddings: 82

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%

Last extracted: 7/18/2026, 1:52:36 PM

Summary

The paper introduces JobMate, an interactive system that transforms authentic social media career posts into persona-grounded conversational AI agents to shift peer career exploration from passive scrolling to active dialogue. A between-subjects study (N=24) comparing JobMate with native RedNote browsing demonstrated that JobMate redirects social comparison from detrimental upward comparison to constructive self-reframing and promotes sensemaking, while maintaining lower cognitive load. The system utilizes a multi-stage pipeline including data cleaning, LLM classification, structured extraction, and dual-track retrieval-augmented generation (RAG) to create personas grounded in real user-generated content.

Entities (9)

JobMate · system · 98%RedNote · platform · 95%Self-Determination Theory · theory · 92%Sensemaking · concept · 90%Social Comparison · concept · 90%Retrieval-Augmented Generation · technique · 88%GPT-3.5-Turbo · model · 85%NASA-TLX · metric · 85%text-embedding-3-small · model · 85%

Relation Signals (7)

JobMate → compareswith → RedNote

confidence 95% · We conducted a between-subjects study (N = 24, three disciplines) comparing JobMate with native RedNote browsing.

JobMate → isbasedon → Self-Determination Theory

confidence 92% · JobMate’s conversational framework is grounded in Self-Determination Theory (SDT), aiming to provide emotional support alongside informational support.

JobMate → promotes → Sensemaking

confidence 90% · Our study shows that JobMate’s AI-mediated dialogue... promoting sensemaking through active conversational engagement.

JobMate → usestechnique → Retrieval-Augmented Generation

confidence 90% · JobMate's architecture utilizes dual-track retrieval-augmented generation (Dual-track RAG) to augment personas with supporting materials.

JobMate → reduces → Social Comparison

confidence 88% · JobMate's AI-mediated dialogue redirected social comparison from potentially detrimental upward comparison toward constructive self-reframing.

JobMate → usesmodel → GPT-3.5-Turbo

confidence 85% · We used GPT-3.5-turbo to assess post validity and classify content.

JobMate → usesmodel → text-embedding-3-small

confidence 85% · All corpora used OpenAItext-embedding-3-smallfor vector embeddings

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Young job seekers frequently turn to social media to compare themselves with peers and make sense of career possibilities. However, passive feed browsing creates a paradox: the authentic peer content that provides emotional grounding also triggers potentially detrimental upward social comparison and cognitive overload. Previous work has either structured online user-generated content to reduce noise without changing the passive browsing modality, or built AI-powered career exploration systems that disregard authentic human experiences. To address this gap, we developed JobMate, an interactive system that transforms real social media career posts into persona-grounded conversational AI agents, shifting the interaction from passive scrolling to active, personalized dialogue. We conducted a between-subjects study ($N$ = 24, three disciplines) comparing JobMate with native RedNote browsing. Our study shows that JobMate's AI-mediated dialogue redirected social comparison from potentially detrimental upward comparison toward constructive self-reframing, while promoting sensemaking through active conversational engagement. However, users still relied on the authenticity of real peer content for emotional grounding. We discuss design implications for AI systems that augment authentic online user-generated content consumption across social comparison contexts.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

81,019 characters extracted from source content.

Expand or collapse full text

Same Stories, Different Journeys: From Social Comparison to Sensemaking in AI-Mediated Peer Career Exploration Pengping Tan Sun Yat-sen University Guangzhou, China tanpp5@mail2.sysu.edu.cn Baoquan Zhao Sun Yat-sen University Zhuhai, China zhaobaoquan@mail.sysu.edu.cn Zhenhui Peng * Sun Yat-sen University Zhuhai, China pengzhh29@mail.sysu.edu.cn Figure 1: End-to-end experience: native RedNote-style career feeds give way to structured persona fields (background, outcome, challenges, summary), surfaced as scannable cards with a chat entry point—shifting exploration from passive scrolling to persona-grounded dialogue while keeping real posts as the source. Abstract Young job seekers frequently turn to social media to compare them- selves with peers and make sense of career possibilities. However, passive feed browsing creates a paradox: the authentic peer content that provides emotional grounding also triggers potentially detri- mental upward social comparison and cognitive overload. Previous work has either structured online user-generated content to reduce noise without changing the passive browsing modality, or built AI-powered career exploration systems that disregard authentic human experiences. To address this gap, we developed JobMate, an interactive system that transforms real social media career posts into persona-grounded conversational AI agents, shifting the inter- action from passive scrolling to active, personalized dialogue. We conducted a between-subjects study (푁= 24, three disciplines) com- paring JobMate with native RedNote browsing. Our study shows that JobMate’s AI-mediated dialogue redirected social comparison Preprint, 2026. from potentially detrimental upward comparison toward construc- tive self-reframing, while promoting sensemaking through active conversational engagement. However, users still relied on the au- thenticity of real peer content for emotional grounding. We discuss design implications for AI systems that augment authentic online user-generated content consumption across social comparison con- texts. CCS Concepts • Human-centered computing→Natural language interfaces; Empirical studies in interaction design; Collaborative and social com- puting systems and tools. Keywords Career exploration, Social comparison, Sensemaking, Persona-grounded agents, Social media UGC, Retrieval-augmented generation, Self- determination theory arXiv:2607.11039v1 [cs.HC] 13 Jul 2026 Preprint, 2026,Tan et al. 1 Introduction Young job seekers increasingly turn to social media to find peers with comparable backgrounds whose career trajectories can serve as references for self-positioning [3,25]. These authentic narra- tives provide situated guidance that formal counseling often cannot deliver [17,19]. Yet the same browsing process also inflicts harm: exposure to peer achievements triggers upward social compari- son, increasing career frustration and anxiety [2,24,45], while noisy feeds produce cognitive overload and “pseudo-clarity” that dissolves because information was passively received rather than ac- tively processed [28,37]. Our formative study (64 survey responses, 8 in-depth interviews) confirmed both sides: participants valued authentic peer experiences above all other career references, yet described “saving many posts but never revisiting them” and feeling anxious when encountering others’ achievements. Previous work has either structured user-generated content to reduce noise [6,22,44] without changing the passive browsing modality, or built AI career exploration tools [10,15,20,41] that shift to active dialogue but rely on synthetic content. Large language model persona agents [14,27,47] offer engaging conversation, but existing personas are either fully synthetic or user-configured, with none grounded in real others’ experiences. No prior work has simultaneously preserved authentic peer content and transformed the interaction from passive browsing to active dialogue. To address this gap, we designed JobMate (Figure 1), an inter- active system that transforms real social media career posts into persona-grounded conversational AI agents. Based on our forma- tive findings, the system processes raw posts through a multi-stage pipeline driven by users’ information needs, generating person- centric cards that foreground challenges rather than achievements to promote lateral comparison [5,43]. Selecting a card opens a split- view interface: one side displays the original post as an authenticity anchor; the other provides persona-grounded dialogue informed by Self-Determination Theory [31], balancing informational and emotional support. A between-subjects study (푁=24, three disciplines) comparing JobMate with native RedNote browsing showed that both con- ditions reduced career decision-making difficulties, but JobMate achieved this at significantly lower cognitive cost (NASA-TLX Effort 푝=0.012). Qualitative analysis revealed that JobMate redirected social comparison from “I’m not as good as others” toward “what should I do next,” with disciplinary cognitive style as an important boundary condition. This paper contributes: (1) empirical evidence that the tension between value and harm in peer experience con- sumption is shaped by interaction modality rather than content alone; (2) the JobMate system, a complete architecture for trans- forming authentic user-generated content into persona-grounded conversational agents; and (3) design implications for balancing in- formational and emotional support, accommodating cognitive style differences, and ensuring content transparency in AI-mediated peer experience systems. 2 Related Work 2.1 Social Comparison and Peer Experience Consumption Social comparison theory [11] holds that individuals evaluate them- selves by comparing with others when objective standards are absent, with upward comparison (against superior targets) fre- quently triggering anxiety [5], downward comparison bolstering self-evaluation [42], and lateral comparison (against similar peers) providing normalization [43]. Social media amplifies these dynam- ics: Vogel et al. [40] found that exposure to high-achieving targets reduces self-esteem, Verduyn et al. [39] showed that passive con- sumption is associated with lower well-being, and a meta-analysis of 48 studies (푁=7,679) confirmed significant negative effects of upward comparison (푔=−0.24) [2]. In the career domain, Yang et al. [45] found that viewing peers’ career posts increases career frustration through comparison, and Luo et al. [24] demonstrated that social media use increases employ- ment anxiety among Chinese youth via upward comparison. Van Zandvoort et al. [38] further warned that embedding comparison features in applications risks triggering the same negative emotions they aim to address. Meanwhile, user-generated content remains an irreplaceable source of authentic peer experience: young job seekers rely on peer posts for situated career guidance [3,25], perceiving them as more trustworthy than institutional content [12], with on- line community support also benefiting psychological health [9,29]. This creates a core tension: the same authentic peer content that provides emotional grounding also fuels harmful comparison. Ex- isting work documents these effects but rarely explores how system design can proactively redirect comparison direction. 2.2 AI-Mediated Sensemaking and Information Exploration Sensemaking is a fundamentally active and iterative process [28], and cognitive load theory [37] highlights how noise consumes resources that should be allocated to deep processing [1]. The gen- eration effect [30,34] and the ICAP framework [7] further show that constructive and interactive engagement produces deeper learning than passive reception, providing theoretical grounding for shifting from browsing to dialogue. Large language models increasingly support active information exploration. Suh et al. [36] developed Sensecape for multilevel sense- making via semantic zoom, and Kim et al. [47] used multi-agent dialogue to burst filter bubbles [26], though Sharma et al. [33] cau- tioned that conversational search can also form “generative echo chambers.” In user-generated content structuring, PlanHelper [22] uses answer posts for plan construction, DesignQuizzer [6] trans- forms community content into conversational learning, and ComViewer [44] provides interactive visual search for mental health communities. These tools improve information presentation but preserve the pas- sive browsing modality. How the interaction modality itself reshapes cognitive processing and emotional experience in high-comparison peer content scenarios remains underexplored. Same Stories, Different Journeys: From Social Comparison to Sensemaking in AI-Mediated Peer Career ExplorationPreprint, 2026, 2.3LLM Persona Agents and Career Exploration Large language models enable conversational agents with rich per- sonas. Park et al. [27] demonstrated that LLM agents can simulate believable human behavior, Zhang et al. [46] and Shao et al. [32] showed that persona information improves dialogue consistency, and Ha et al. [14] found that users form emotional bonds with customized personas and engage in richer dialogue. Agent self- disclosure further promotes reciprocal trust-building [8, 18, 23]. In career exploration, Jeon et al. [20] built “future self” agents for letter-exchange exercises, Han et al. [15] developed a career chatbot grounded in Self-Determination Theory [31], and Du et al. [10] and Wang et al. [41] explored gamified and metaphorical career simulations. Retrieval-augmented generation [21] further en- ables grounding dialogue in external knowledge. However, existing persona agents are primarily fully synthetic [27], user-configured fictions [14], or projections of the user’s own future self [20]. No prior work has grounded personas in real others’ experiences to simultaneously preserve the authenticity of peer content, enable active conversational engagement, and redirect the direction of social comparison through design. 3 Formative Study To understand the pain points and needs of job seekers when using social media to gather career information, we conducted a formative study combining a survey and semi-structured interviews to inform the design of JobMate. 3.1 Participants and Procedure We first distributed an online survey and collected 64 valid re- sponses from university students who had experience using social media for career-related information seeking. The survey covered three areas: job-seeking motivation and difficulties, experiences and challenges with career-related social media posts, and expectations for AI-assisted tools. Based on the survey results, we recruited 8 participants (6 female, 2 male; ages 22–28;푀=24.1) from 8 differ- ent disciplines including Computer Science, Law, and Mathematics for follow-up semi-structured interviews. Each interview lasted 30–45 minutes and covered: job-seeking challenges and emotional experiences, social media usage habits and browsing difficulties, experiences seeking advice from peers or mentors, and expectations for AI-assisted career tools. Prior to each interview, participants completed a brief questionnaire about their job-seeking motivation, social media usage, and expectations for intelligent assistants. All interviews were audio-recorded, transcribed, and analyzed using thematic analysis [4]. 3.2 Findings Our analysis revealed a core tension: participants regarded authen- tic peer experiences as their most important career reference, yet systematically suffered cognitive and emotional costs while con- suming this content. We organized the findings into four themes: F1: Authentic Peer Experiences as Irreplaceable Guidance. All participants considered real peer experiences their most val- ued information source. They actively sought background-matched peers to calibrate expectations: P3 searched for interview outcomes to benchmark his position; P7 posted her situation to solicit per- sonalized advice; P8 discovered previously unconsidered career paths through peer posts. P8 noted: “RedNote gives you new ideas, for example, a traditional Chinese medicine student can find career options beyond hospitals, from military positions to pet acupuncture.” Participants generally believed that real human experiences were more timely and trustworthy than AI-generated content (P1). F2: Information Chaos and Pseudo-Clarity. Extracting ac- tionable information was prohibitively costly. P5 stated: “Too much information, impossible to tell real from fake. The exaggeration cre- ates catastrophic imagination.” Participants identified hidden adver- tisements (P4, P6, P8), extreme polarization (P2), and fragmented information resisting synthesis (P1). P4 described passive accumu- lation without internalization: “I save posts but never go back to use them.” P7 reported: “Conflicting evaluations of the same thing: the more I read, the more anxious I get.” Due to fragmented browsing, participants often “read and forget, retaining only vague impressions” (P4), making it difficult to translate information into actionable guidance. Finding background-matched, practically useful informa- tion required substantial time and effort for manual filtering and comparison. F3: Uncontrolled Social Comparison. P1 described a double bind: “Seeing the industry described as a ‘sunset industry,’ and then seeing others with multiple offers, both make me anxious.” P5 de- scribed “catastrophic imagination” triggered by negative narratives. P2 similarly noted: “Seeing excellent people with many offers makes me anxious; seeing information about industry decline also makes me anxious.” The same platform delivered both comfort and harm, while users lacked control over the direction of comparison. Anxi- ety pervaded the entire job-seeking process: anxiety about offers when without one, anxiety about choices when holding offers, and anxiety about advantages not yet obtained (P1). F4: Unmet Need for Dialogue. Participants expressed frustra- tion at the inability to converse with resonant posters. P5 stated: “When something in a post is unclear, you can’t get an immediate answer.” P3 noted: “Private messages go unanswered; communication with bloggers is difficult.” P7 described attempts to seek interaction through posting and commenting: “Getting information on RedNote mainly involves posting my own questions or commenting on others’ posts, describing my situation and asking for advice.” Participants also expressed needs for multiple roles: resume editing, interview coaching, industry expert advice, career-switching experiences, self- assessment, HR perspectives, and peers also searching for jobs (P5). This suggests that shifting the interaction modality from passive browsing to active conversation might help alleviate this tension. 4 JobMate Based on the formative findings, we designed and implemented JobMate, a system that transforms authentic career-experience posts from social media into conversational digital personas. This section presents the design goals, user interface, data pipeline, and conversational framework. 4.1 Design Goals The formative study revealed a core tension: users regarded au- thentic peer experiences as their most important reference (F1), Preprint, 2026,Tan et al. yet suffered from information chaos (F2), uncontrolled social com- parison (F3), and the inability to converse with posters (F4) while consuming this content. We distilled four design goals: DG1: Filter noise and structure authentic experiences (ad- dressing F2). The formative study found that users faced extensive advertisements, misinformation, and fragmented content on Red- Note, making information extraction prohibitively costly. JobMate filters invalid content through a multi-stage pipeline and structures experience posts into scannable persona cards, reducing screening effort. DG2: Transform passive browsing into active dialogue (ad- dressing F4). The formative study found that users wanted to in- teract with resonant posters but private messages often went unan- swered. JobMate enables users to converse with personas grounded in real posts, mapping external experiences to their own situations through questioning and articulation. DG3: Mitigate harmful social comparison while preserving emotional support (addressing F3). The formative study found that users were caught in a double bind, comforted by shared anxi- ety yet threatened by others’ achievements on the same platform. JobMate foregrounds challenges tags rather than accomplishments on persona cards, redirecting comparison from upward (threat- ening) toward lateral (normalizing) directions; dialogue provides emotional support through empathy and cognitive reframing. DG4: Anchor dialogue in real experiences to maintain authenticity (addressing F1). The formative study found that users considered real human experiences more trustworthy than AI-generated content. All JobMate personas are generated from real posts, with original posts displayed as verifiable sources so users can check the basis of AI responses at any time. 4.2 User Interface The interface supports person-centric exploration, helping users shift from passive scrolling to active dialogue while keeping au- thentic posts visible (Figures 2–3). Onboarding. Users complete a two-phase profile (Figure 2a). Phase 1 collects hard attributes (school, major, degree). Phase 2 captures soft psychological state (current mood, primary difficulty). This enables matching on both background and emotional position, providing personalized context for subsequent dialogue. Gallery view. Persona cards are presented in a scrollable grid (Figure 2b). Each card displays four core elements: background tag, job-seeking outcome, experience summary, and prominently displayed challenges tags (colored chips). The unit of exploration is a person, not a post; foregrounding challenges tags encourages users to first notice “what difficulties this person also faced” rather than “how accomplished this person is,” promoting lateral rather than upward comparison (DG3). Split-view detail. Selecting a persona opens a split-view inter- face (Figure 3): the left panel displays the original RedNote post as an authenticity anchor (DG4); the right panel offers AI con- versation with the persona, including three suggested questions generated from user onboarding information and RAG-retrieved related recommendations (similar experience stories, interview tips, industry insights). Users can verify agent responses against the source post while flexibly alternating between passive reading and active dialogue. 4.3 Data Pipeline JobMate transforms raw RedNote posts into conversational digi- tal personas through a four-stage pipeline (Figure 4). Each stage directly addresses formative findings. Stage 1: Data cleaning. We scraped career-related posts from RedNote using the web scraping tool biaoda.me, collecting 220 posts for each of three disciplines (Computer Science, Psychology, Chinese Literature), totaling 660 raw posts. After deduplication, text cleaning, and filtering posts with fewer than 100 characters, 414 valid posts remained. Stage 2: LLM classification and filtering. The formative study found that advertisements and misinformation were major pain points (F2). We used GPT-3.5-turbo to assess post validity and classify content. Classification categories were derived from in- formation needs mentioned by formative participants: (1) personal experience (job-seeking journeys, internship records, recruitment summaries); (2) interview tip (interview questions, written test ex- periences, process reviews); (3) industry insight (industry trends, position overviews); (4) recruitment information (referral codes, hiring announcements). Invalid posts (pure advertisements, low- information content, off-topic material) were filtered out. Manual verification of 100 posts showed 99% classification accuracy. After this stage, 364 valid posts remained. Stage 3: Structured information extraction. The formative study found that users struggled to quickly judge “whether this person is similar to me” from fragmented content (F2). We used GPT- 3.5-turbo to extract four structured fields from personal experience posts: (1) background info, core background summary (e.g., “985 Psychology,” “non-elite school, cross-disciplinary”); (2) job seeking outcome, final job-seeking result (e.g., “received ByteDance offer,” “landed after multiple interview failures”); (3) challenges tags, 1–3 difficulty/weakness tags (e.g., “no internship,” “failed courses,” “ca- reer pivot”), emphasizing pain points rather than achievements to support DG3; (4) experience summary, one-sentence summary of the most anxious phase and how it was overcome. Manual verifica- tion of 50 extraction results showed 94% accuracy. Stage 4: Routing and dual-track RAG aggregation. Per- sonal experience posts with complete narrative arcs (non-empty job_seeking_outcome) entered the Core Persona Pool as conversable main characters; other personal stories joined the Others’ Personal Story corpus; interview tips, industry knowledge, and recruitment posts formed the Knowledge Base. Across three disciplines, 118 core personas were generated (CS: 39, Psychology: 33, Chinese Literature: 46). Each core persona was augmented through dual-track retrieval- augmented generation (Dual-track RAG) with supporting ma- terials (DG4): •Track 1 (homogeneous): Based on similarity of background and challenges tags, retrieves Top-2 peer experiences from the Others’ Personal Story corpus with similar situations, providing “you are not alone” resonance support during conversation (DG3). Same Stories, Different Journeys: From Social Comparison to Sensemaking in AI-Mediated Peer Career ExplorationPreprint, 2026, Figure 2: JobMate interface. (a) Two-phase onboarding collecting demographic attributes (school, major, degree) and psycholog- ical state (mood, primary difficulty) for personalized matching. (b) Gallery of persona cards organized around people, each card displays background, job-seeking outcome, experience summary, and challenges tags (colored chips), with a button to start chat. Figure 3: Split-view detail interface. Left: original RedNote post as an authenticity anchor. Right: persona-grounded chat with three suggested questions tailored to onboarding context, and a book icon to open RAG-retrieved related recommendations (personal stories, interview tips, industry insights). • Track 2 (heterogeneous): Based on similarity of job-seeking outcome, retrieves Top-3 practical resources from the Knowl- edge Base related to the persona’s career destination (inter- view tips, industry insights), providing actionable informa- tion support. All corpora used OpenAItext-embedding-3-smallfor vector embeddings, with retrieval based on cosine similarity. The assem- bled “complete persona” comprises: core post (main character), similar experiences (emotional resonance), and practical resources (knowledge support), provided as context to the LLM during con- versation. 4.4 Conversational Framework JobMate’s conversational framework is grounded in Self-Determination Theory (SDT) [31], aiming to provide emotional support alongside informational support by satisfying users’ needs for relatedness, competence, and autonomy (DG2, DG3). Each persona agent receives a composite prompt containing: persona content (background, outcome, experience summary, full original post), user context (background and emotional state from onboarding), dual-track RAG-retrieved related posts, and conversa- tional rules. The agent’s role is “a real senior student who shared experiences on RedNote, now chatting via private message with a follower,” not a generic AI assistant. Preprint, 2026,Tan et al. Figure 4: Data pipeline overview. Raw posts are cleaned and LLM-classified; valid content splits into a knowledge base (interview tips, industry insights, recruitment info) and personal stories, from which we extract background, job-seeking outcome, challenges tags, and experience summary. Posts with clear outcomes seed the core persona pool; other stories and knowledge base items become supporting materials. The final digital persona combines a core post with dual-track RAG-retrieved related content. Relatedness. The agent uses empathic self-disclosure to con- nect its documented struggles with the user’s situation, expressing empathy without judgment. Example: “Don’t panic, back then my resume got rejected enough times to circle the Earth, and here I am still alive and kicking.” Competence. When users express self-doubt, the agent guides cognitive reframing, reinterpreting undervalued experiences as workplace strengths. For example, reframing “I just ran errands during my internship” as “full-chain resource integration and co- ordination capability,” shifting evaluation from “I’m not as good as others” toward “what value do I have that can be recognized” (DG3). Autonomy. Imperative language is prohibited (e.g., “you must,” “you should”); options are offered rather than directives; venting and pauses are allowed. The conversation maintains an open attitude, staying curious about the user without steering toward particular decisions. Constraints. Replies are limited to approximately 150 charac- ters, using colloquial plain text; no self-identification as AI; not every turn ends with a question. Sometimes accepting emotions or offering a virtual hug is sufficient. 4.5 Implementation JobMate is a Vue/Node.js web application. Conversational agents use GPT-5.2; classification and extraction use GPT-3.5-turbo. Embed- dings usetext-embedding-3-smallwith cosine similarity. The pipeline processed RedNote posts across three disciplines, yielding 118 persona agents with complete narrative arcs. 5 User Study We conducted a between-subjects experiment to compare JobMate with native RedNote browsing during career exploration. The study addressed three research questions: • RQ1: How does an AI-mediated system change cognitive processing and sensemaking? • RQ2: How does an AI-mediated system reshape social com- parison and emotional experience? •RQ3: How do other conditions (e.g., disciplinary cognitive style, interaction preference, and content preference) influ- ence AI-mediated exploration? 5.1 Participants We recruited 24 university students (17 female, 7 male; ages 18–25) actively preparing for job searches, from three disciplines: Com- puter Science (푛=8), Psychology (푛=8), and Chinese Literature (푛=8). These disciplines were chosen to capture diverse cognitive styles: Computer Science emphasizes logical and technical reason- ing, Psychology involves higher empathic sensitivity, and Chinese Literature entails proficiency in processing long-form text. Within each discipline, participants were randomly assigned to the JobMate condition (푛=4) or the native RedNote browsing condition (푛=4). A Kruskal-Wallis test confirmed no significant differences in pre-task questionnaire scores across disciplines or between the JobMate and RedNote groups (all 푝> 0.05), ensuring comparability. 5.2 Study Design We used a between-subjects design comparing JobMate with un- constrained native RedNote browsing. Sessions were conducted remotely via video conferencing and lasted 50–60 minutes. Both groups completed a 30-minute career exploration task with the same objectives: (1) Understand what people with similar backgrounds experi- enced during job searching, what difficulties they encountered, and where they ended up; (2) Attempt to obtain information that could increase confidence or clarify direction; (3) Based on the informa- tion gathered, reflect on possible next steps or career directions. JobMate condition. Participants first received a brief system walkthrough covering registration, card browsing, and dialogue features, then freely used the system to complete the task. RedNote condition. Participants used the RedNote mobile app as they normally would to search and browse career-experience posts related to their discipline. Same Stories, Different Journeys: From Social Comparison to Sensemaking in AI-Mediated Peer Career ExplorationPreprint, 2026, 5.3 Measures We used four quantitative instruments: the 34-item Career Decision- Making Difficulties Questionnaire (CDDQ; [13]; pre/post, 7-point; pre훼=0.914, post훼=0.888) to measure change in career decision- making difficulties; the NASA Task Load Index (NASA-TLX; [16]; 6 dimensions, 7-point) to assess cognitive load; two single items for perceived informational and emotional support (7-point); and the 12-item Self-Determination Scale (SDS; [35];훼=0.843) to evalu- ate autonomy, competence, and relatedness. JobMate participants additionally rated seven system-specific items (e.g., reuse inten- tion, interaction naturalness, card comprehension). After the task, we conducted 10–15 minute semi-structured interviews covering cognitive experience, emotional responses, and comparisons with everyday browsing; all interviews were audio-recorded and tran- scribed for thematic analysis. For the JobMate condition, we also logged conversation turns and the number of personas engaged. 5.4 Procedure Each session comprised five steps: (1) Introduction and consent (5 min): The experimenter explained the study background, proce- dure, and instructions via video conferencing. (2) Pre-task ques- tionnaire (5 min): Participants completed the CDDQ to establish baseline status. (3) Exploration task (30 min): Participants used JobMate or RedNote according to their assigned condition. (4) Post- task questionnaires (10 min): Participants completed the CDDQ post-test, NASA-TLX, perceived support items, SDS, and (JobMate only) the seven system-experience items. (5) Semi-structured in- terview (10–15 min): In-depth discussion of usage experience, in- cluding sources of informational and emotional support, cognitive load, and comparisons with daily browsing habits. 5.5 Data Analysis Quantitative analysis. For between-group comparisons (푁=24), Shapiro–Wilk tests indicated normal distribution; we therefore re- port independent-samples푡-tests. Pre–post changes were analyzed with paired푡-tests. Discipline-specific subgroup comparisons (푛=4 per cell) used Mann–Whitney푈tests for exploratory analysis given the small sample sizes. Qualitative analysis. Interview transcripts were analyzed using reflexive thematic analysis [4]. Two researchers first conducted open coding independently, then reconciled themes through axial coding. Core themes related to cognitive processing, emotional experience, and social comparison were extracted. Interaction log analysis. Conversation turns and number of personas engaged were summarized descriptively to identify differ- ent exploration strategies in the JobMate condition. 6 Results We answered the research questions by integrating quantitative and qualitative findings from the user study. Shapiro–Wilk tests showed that data from both JobMate and RedNote conditions were normally distributed; therefore, we report independent-samples 푡 -tests for between-group comparisons. Figure 5: Pre–post CDDQ subscales for JobMate vs. RedNote (plots label the control as Baseline; this denotes native Red- Note browsing). Response options were coded 0 (strongly disagree) to 6 (strongly agree) with the difficulty-indicating statements; higher values indicate stronger agreement with experiencing difficulty. 6.1 Cognitive Processing and Sensemaking Comparable reduction in career decision difficulty. Both groups showed significant pre–post reductions in overall Career Decision- Making Difficulties scores: JobMate (Δ푀=0.36,푝=0.005<0.01) and RedNote (Δ푀=0.48,푝=0.001<0.01). The between-group difference in total reduction was not significant (푝 푡 =0.432>0.05), and baseline difficulty was also comparable (푝 푡 =0.193>0.05). At the subscale level, Lack of Information showed the largest im- provement (JobMateΔ=0.74,푝=0.006<0.01; RedNoteΔ=1.04, 푝=0.003<0.01), while Inconsistent Information showed no sig- nificant pre–post change in either condition (푝>0.05). Notably, the RedNote mean on Inconsistent Information slightly increased after the task, suggesting that contradictory career signals were difficult to resolve through either browsing or dialogue within a short session (Figure 5). Lower cognitive cost at equivalent outcome. Although performance- related dimensions did not differ significantly, NASA-TLX showed a clear divergence in cognitive cost. Effort (How hard did you have to work to accomplish the task?) was significantly lower in JobMate than RedNote (푀=3.92 vs.푀=5.08;푝 푡 =0.012), and Frustration (How stressed and annoyed were you?) was marginally lower in Job- Mate (푀=2.50 vs.푀=3.75;푝 푡 =0.075). Other dimensions did not reach conventional significance, but most trended toward higher subjective load in native feed browsing (Table 1). Qualitative data explained this pattern: RedNote participants repeatedly described high screening cost (e.g., Lots of ads, hard to tell what is real... the layout is dizzying. P4), whereas JobMate users described card-based, ad-free presentation as reducing external clutter (e.g., The interface is clean... cards save selection cost. P9). Two exploration strategies in JobMate. Interaction logs showed two distinct patterns in the JobMate condition: deep divers (푛=5; e.g., P1, P6), who sustained 28–37 turns with 1–2 personas, and broad explorers (푛=4; e.g., P3, P15), who sampled 5–9 personas with shorter exchanges (3–5 turns each). Others showed mixed behavior (Figure 6). Compared with the feed condition, where behavior was mainly passive scrolling, JobMate supported multiple exploration strategies. Breadth–depth asymmetry in information absorption. Red- Note users often reported browsing many posts (roughly 10–15) Preprint, 2026,Tan et al. Table 1: NASA-TLX cognitive load comparison between Job- Mate and RedNote conditions (7-step scale, coded 1–7; lower is better except Performance). DimensionJobMateRedNote풑 풕 풑 풖 푀 ±푆퐷 푀±푆퐷 Mental Demand 4.92± 1.16 5.58± 0.79.117.131 Physical Demand 3.42± 1.56 2.42± 1.78.158.117 Temporal Demand 3.00± 1.21 3.25± 1.29.628.634 Performance † 5.33± 0.89 5.50± 0.90.653.602 Effort 3.92± 1.00 5.08± 1.08 .012* .015* Frustration 2.50± 1.62 3.75± 1.66.075 † .082 † Info. Support 5.67± 1.15 5.58± 1.00.852.786 Emot. Support 5.33± 1.37 5.25± 1.48.888.858 † Higher is better. *푝< .05. † Marginal (푝< .10). Figure 6: Conversation turns in JobMate, stacked by persona: each colored segment is one persona engaged in the session (segment length reflects turns with that persona). yet experiencing only short-lived clarity (e.g., It felt clear right af- ter browsing, but unclear again after a while, P10), consistent with “pseudo-clarity.” In contrast, some JobMate users perceived lower information density but reported deeper processing through dia- logue (e.g., Chatting requires output, which helps information ab- sorption, P5). This helps explain why CDDQ improvement can be similar across conditions while NASA-TLX Effort (How hard did you have to work to accomplish the task?) still differs significantly: the two systems may produce similar short-term gains on career decision-making difficulties scales, but through different processing pathways—high-friction information throughput versus structured entry with active articulation. 6.2 Social Comparison and Emotional Experience JobMate provided perceived support comparable to RedNote. Post-task single-item ratings showed no significant difference in informational support (JobMate푀=5.67, RedNote푀=5.58; 푝 푡 =0.852) or emotional support (JobMate푀=5.33, RedNote 푀=5.25;푝 푡 =0.888). Likewise, SDS subscales showed no signif- icant differences (Autonomy:푝=0.760; Competence:푝=0.814; Figure 7: JobMate-only ratings for the seven interface and experience items (same 1–7 agree–disagree coding as other self-report plots). Relatedness:푝=0.675). In this short task, JobMate did not reduce users’ perceived agency or support relative to a human-populated platform. Different social-comparison dynamics despite similar scale scores. In RedNote, comparison often followed a “relief with dete- rioration” pattern: users felt comfort from shared anxiety, but also became more anxious through upward comparison (e.g., P12: seeing others also anxious gave relief, but seeing peers with many internships increased anxiety; P6: both doom narratives and high-achieving peers induced pressure). These reports suggest that homogeneous anxiety cues and upward comparison triggers were repeatedly activated within the same feed. In JobMate, users described a different scaf- fold for comparison. Challenges tags shifted attention away from achievement-only signals, like a mutual-aid group for challenges (P2). Dialogue-based reframing also directly changed self-narratives, including reinterpreting a “wasted” project (P21) and reducing the barrier of describing internship experience (P5). Resume guidance was described as very useful... like someone who truly understands you (P14). Together, these findings indicate that JobMate changed how comparison was experienced and supported self-reframing grounded in users’ own situations. Authentic peer experience remained the emotional anchor. Users still anchored value in real peer experiences. One JobMate participant estimated that 60–80% of useful information came from follow-up AI questions, but the starting point was real people’s expe- rience (P9). AI expanded interactivity, but perceived meaning and trust still relied on authentic posts and real-life context. 6.3 Boundary Conditions: Cognitive Style, Interaction, and Content Preferences Because subgroup sample sizes were small, we used Mann–Whitney 푈tests for exploratory analysis. Results suggest that boundary conditions meaningfully shaped exploration experience. Disciplinary cognitive style shaped perceived cognitive load. The clearest subgroup effect appeared in psychology: Job- Mate showed significantly lower Effort than RedNote (푀=3.25 vs.푀=5.25,푝 푢 =0.036<0.05), and psychology participants gave the highest card-comprehension rating (푀=6.50). Qualita- tive evidence suggests that these participants were more easily pulled by emotional and contradictory feed content; in this context, structured cards and dialogue-based filtering protected empathic Same Stories, Different Journeys: From Social Comparison to Sensemaking in AI-Mediated Peer Career ExplorationPreprint, 2026, Table 2: Exploratory discipline-specific findings (each cell 푛=4). Effort is the NASA-TLX dimension; Emotional Support on 7-point scale. DisciplineMeasureJobMate RedNote풑 풖 PsychologyEffort3.255.25.036* PsychologyPhys. Demand3.501.50.074 † Computer Sci.Emot. Support6.504.75.052 † Computer Sci.Mental Demand4.506.00.137 Chinese Lit.Effort4.505.25.642 Chinese Lit.Frustration2.253.75.369 *푝< .05. † Marginal (푝< .10). resources from irrelevant noise. By contrast, the Chinese Literature subgroup showed no significant difference on Effort (푀=4.50 vs. 푀=5.25,푝 푢 =0.642>0.05), and several participants reported too little information and insufficient detail. This indicates that for users with high long-text processing ability, aggressive information compression may create a perceived density deficit. In other words, the same de-noising strategy is not equally optimal across cognitive styles. Interaction preference shaped where support was per- ceived to come from. Although overall support scales showed no between-group difference, discipline-level patterns diverged. In Computer Science, emotional support showed a marginal subgroup difference, with JobMate higher than RedNote (푀=6.50 vs. 4.75, 푝 푢 =0.052<0.1). This contrasts with the common assumption that human communities are always superior for emotional sup- port. Interviews suggest that some technically oriented users were more sensitive to interpersonal noise and emotional drag in social feeds (e.g., Seeing others with more project experience makes me anx- ious), and preferred AI interaction for its stable, non-judgmental, and queryable feedback, including personalized skill reframing and next-step planning. In contrast, Chinese Literature participants more often described support as coming from exposure to career options that differed from their daily expectations. Content coverage shaped matching quality and trust. Struc- tural coverage bias in available online user-generated content (over- representation of tech/large-company trajectories and underrepre- sentation of traditional paths) reduced perceived relevance, espe- cially for Chinese Literature participants. Many reported that most people around them were preparing for civil service exams, while online content did not match their real context. Participants also described outputs as not fitting, chicken-soup-like, or generic. These reports suggest that when the system lacks authentic cases aligned with users’ real situations, AI support is more likely to be perceived as templated. Overall, exploration experience was determined not only by interaction modality, but also by corpus representativeness: lim- ited coverage can weaken trust, resonance, and long-term reuse intention. 7 Discussion Our study sought to answer: when authentic peer content remains unchanged, how does shifting the interaction modality from pas- sive browsing to AI-mediated dialogue reshape users’ cognitive processing and transform social comparison into sensemaking? Through formative interviews, we found that users consum- ing career-related online user-generated content often experience both “helpful” and “exhausting” at the same time. Based on this tension, we designed JobMate, a system that transforms real posts into conversational personas, and compared it with native RedNote browsing. Results show that both approaches can reduce career decision difficulty in the short term, but JobMate requires less cog- nitive effort and makes it easier for users to shift attention from “others are better than me” toward “what should I do next.” We organize the discussion around four themes: (1) Interaction modality is the key variable shaping peer-experience consump- tion, not content alone; JobMate’s effect stems from three coupled mechanisms: cognitive offloading, output-driven sensemaking, and comparison reframing, which generalize to other high-comparison online user-generated content contexts. (2) AI mediation brings benefits but also identifiable risks and boundaries. (3) Design impli- cations for balancing informational and emotional support across users with different cognitive styles. (4) Future systems should provide adjustable support modes and help users understand the system’s basis and content coverage. 7.1 Mechanisms and Generalization: How Interaction Modality Reshapes Peer-Experience Consumption Our main finding is that the same authentic peer content leads to different psychological costs and comprehension pathways under different interaction modalities. Although both JobMate and Red- Note effectively reduced career decision difficulty in a short task, JobMate imposed lower cognitive burden. This suggests that Job- Mate’s value lies not in “showing more information” but in “making information easier to convert into personal judgment and action.” In short: outcomes may be similar, but the process differs; when the process differs, user experience and subsequent action quality differ as well. We attribute this difference to three interrelated mechanisms. First, cognitive offloading: native feeds are cluttered with ads, repetitive posts, and conflicting opinions, forcing users to spend effort on filtering rather than understanding; structured cards and conversational entry points shift the task from “finding information” to “clarifying the problem.” Second, output-driven sensemaking: passive scrolling easily produces a short-lived illusion of under- standing, whereas dialogue requires users to ask questions, reflect, and articulate, compelling them to map external experiences onto their own situations and thereby form more robust meaning. Third, comparison reframing: native platforms often cycle between shared-anxiety comfort and upward-comparison anxiety; JobMate foregrounds challenges and uses guided questioning to redirect comparison from “I’m not as good as others” toward “what can I do next,” reducing the threat of upward comparison. Preprint, 2026,Tan et al. These mechanisms are not limited to job seeking. Similar dy- namics may arise wherever “rich authentic peer experiences coex- ist with comparison pressure”—for example, study planning (peer grades and offers), fitness and body image (progress and appear- ance), parenting (child-rearing approaches and developmental mile- stones), chronic-disease management (treatment paths and recovery progress), and creator growth (traffic and creative output). In all these contexts, users need others’ experiences as references yet risk being drained by comparison. Our findings suggest that system de- sign can—without changing the content source—achieve two goals simultaneously by changing the interaction modality: reduce the psychological toll of comparison and improve the conversion of experience into action. 7.2 Risks and Boundaries: Personalized AI Support for Different Users We also observed clear boundaries. First, not everyone needs emotional comfort. Some users are highly goal-oriented and sim- ply want actionable advice fast. For them, excessive encouragement feels “empty,” “slow,” or “unhelpful.” Second, personas may “look real but lack substance.” Grounding personas in authentic expe- riences can build initial trust, but if the underlying data is shallow or suggestions lack specificity, users quickly perceive responses as “boilerplate” or “templated,” which erodes trust instead of build- ing it. Third, a natural gap exists between the real world and online or LLM-generated content. Users face concrete life con- straints such as regional opportunities, family circumstances, and job thresholds, while online content and AI-generated text are of- ten “readable but not fully actionable.” This means systems cannot merely pursue “fluent responses”; they must also help users judge “does this apply to me?” and narrow the distance between provided content and reality. 7.3 Design Implications: Balancing Informational and Emotional Support Across User Differences Offer two modes, not one tone. The system can let users choose upfront: “I want a quick solution” (less comfort, more steps) or “I need to sort out how I feel first” (stabilize emotions, then suggest actions), with seamless switching allowed. This serves both goal- oriented and high-anxiety users. In practice, a lightweight prompt at the start of a conversation (e.g., “Do you want to solve a problem quickly, or talk through your feelings first?”) can route users into different support paths. Personas should “show their basis,” not just “act the part.” Each persona should display three kinds of information: (1) Basis, which real posts the response draws on; (2) Capability scope, what it can help with (e.g., organizing a résumé, comparing paths) and what it cannot do for you (e.g., decide your career direction); (3) Intended audience, which backgrounds and goals it is most suited for. This helps users judge “should I trust this, and to what extent,” reducing both vagueness and misplaced trust. Layer information so different cognitive styles can use it. Not everyone prefers the same length or density. A multi-tier struc- ture works well: Layer 1, a one-sentence summary (quick scan); Layer 2, detailed evidence and examples (deeper read); Layer 3, the original post (self-verification). This satisfies both “fast scan- ners” and “detail seekers.” Our study also showed that psychology- background users preferred structured cards, while Chinese-literature users often felt “there’s not enough information.” Systems should therefore let users expand on demand rather than compress uni- formly by default. Proactively fill in “invisible” populations and paths. When platform content is dominated by “big-tech narratives,” users on non-mainstream paths immediately feel “the system doesn’t get me.” Systems should compensate at the recommendation layer: add cases for civil-service exam prep, local positions, and non-tech careers; indicate on the interface “current content skews toward which paths”; prioritize similar-background cases for under-covered users. This is not just a data issue; it is a user-experience issue. Insufficient coverage directly undermines matching, resonance, and trust. 7.4 Limitations and Future Work Our study has several limitations. First, the sample is small, espe- cially after splitting by discipline, so disciplinary differences should be viewed as directional signals requiring larger-sample confir- mation. Second, the task was brief; we cannot observe long-term effects such as whether anxiety rebounds after a week or whether users actually act on advice. Third, data came from a single platform with uneven career-path distribution, limiting generalizability; fu- ture work should use multi-platform, multi-path data and report coverage. Fourth, we have not yet disentangled each component’s contribution; future ablation studies can separately evaluate struc- tured cards, persona dialogue, and recommended readings. A promising future direction is longitudinal field deployment that tracks: whether users return over time; whether they convert suggestions into action; and which users need “information mode” versus “emotion mode” at which moments. Such tracking can better answer whether the system truly changes the career-exploration process. 8 Conclusion We presented JobMate, a system that transforms real social me- dia career posts into persona-grounded conversational AI agents. A between-subjects study (푁=24) showed that shifting from passive browsing to AI-mediated dialogue reduces cognitive cost and redirects social comparison toward constructive self-reframing, without changing the underlying content. Information overload and comparison anxiety in peer experience consumption are con- sequences of interaction modality, not of the content itself, and can be addressed through interaction redesign. Acknowledgments [Anonymized for review.] References [1] Eugene Agichtein, Carlos Castillo, Debora Donato, Aristides Gionis, and Gilad Mishne. 2008. Finding High-Quality Content in Social Media. Proceedings of the International Conference on Web Search and Data Mining (2008), 183–194. doi:10.1145/1341531.1341557 [2]Markus Appel, Caroline Marker, and Timo Gnambs. 2023. A Meta-Analysis of the Effects of Social Media Exposure to Upward Comparison Targets on Self- Evaluations and Emotions. Media Psychology 26, 5 (2023), 661–681. doi:10.1080/ 15213269.2023.2180647 Same Stories, Different Journeys: From Social Comparison to Sensemaking in AI-Mediated Peer Career ExplorationPreprint, 2026, [3]Brenda L. Berkelaar. 2017. Different Ways New Information Technologies In- fluence Conventional Organizational Practices and Employment Relationships: The Case of Cybervetting for Personnel Selection. Human Relations 70, 9 (2017), 1115–1140. doi:10.1177/0018726716686400 [4]Virginia Braun and Victoria Clarke. 2006. Using Thematic Analysis in Psy- chology. Qualitative Research in Psychology 3, 2 (2006), 77–101. doi:10.1191/ 1478088706qp063oa [5]Bram P. Buunk, Rebecca L. Collins, Shelley E. Taylor, Nico W. VanYperen, and Gayle A. Dakof. 1990. The Affective Consequences of Social Comparison: Either Direction Has Its Ups and Downs. Journal of Personality and Social Psychology 59, 6 (1990), 1238–1249. doi:10.1037/0022-3514.59.6.1238 [6]Zhuoer Chen et al.2024. DesignQuizzer: A Community-Powered Conversational Agent for Learning Visual Design. In Proceedings of the 2024 ACM Conference on Computer-Supported Cooperative Work and Social Computing. doi:10.1145/3637366 [7]Michelene T. H. Chi and Ruth Wylie. 2014. The ICAP Framework: Linking Cognitive Engagement to Active Learning Outcomes. Educational Psychologist 49, 4 (2014), 219–243. doi:10.1080/00461520.2014.965823 [8] Nancy L. Collins and Lynn Carol Miller. 1994. Self-Disclosure and Liking: A Meta- Analytic Review. Psychological Bulletin 116, 3 (1994), 457–475. doi:10.1037/0033- 2909.116.3.457 [9]David C. DeAndrea. 2015. Testing the Proclaimed Affordances of Online Support Groups in a Nationally Representative Sample of Adults Seeking Mental Health Assistance. Journal of Health Communication 20, 2 (2015), 147–156. doi:10.1080/ 10810730.2014.914606 [10]Wenxuan Du, Ziyi Zhu, Xinyi Xu, Haoyue Che, and Si Chen. 2024. CareerSim: Gamification Design Leveraging LLMs For Career Development Reflection. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. 1–7. doi:10.1145/3613905.3650928 [11]Leon Festinger. 1954. A Theory of Social Comparison Processes. Human Relations 7, 2 (1954), 117–140. doi:10.1177/001872675400700202 [12]Raffaele Filieri and Fraser McLeay. 2014. E-WOM and Accommodation: An Analysis of the Factors That Influence Travelers’ Adoption of Information from Online Reviews. Journal of Travel Research 53, 1 (2014), 44–57. doi:10.1177/ 0047287513481274 [13] Itamar Gati, Mina Krausz, and Samuel H. Osipow. 1996. A Taxonomy of Difficul- ties in Career Decision Making. Journal of Counseling Psychology 43, 4 (1996), 510–526. doi:10.1037/0022-0167.43.4.510 [14]Taewan Ha, Changhoon Oh, Yoonseo Oh, Dain Kwak, Daeyeon Kim, Minjoon Lee, and Juho Kim. 2024. CloChat: Understanding How People Customize, Interact, and Experience Personas in Large Language Models. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3613904.3642472 [15]Hyoungwook Han, Bohyun Park, and Kwangsu Seo. 2025. A Self-Determination Theory-based Career Counseling Chatbot: Motivational Interactions to Address Career Decision-Making Difficulties and Enhance Engagement. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. 1–9. doi:10.1145/3706599.3720286 [16]Sandra G. Hart and Lowell E. Staveland. 1988. Development of NASA-TLX (Task Load Index): Results of Empirical and Theoretical Research. Advances in Psychology 52 (1988), 139–183. doi:10.1016/S0166-4115(08)62386-9 [17]Andreas Hirschi, Anne Herrmann, and Anita C. Keller. 2015. Career Adaptivity, Adaptability, and Adapting: A Conceptual and Empirical Investigation. Journal of Vocational Behavior 87 (2015), 1–10. doi:10.1016/j.jvb.2014.11.008 [18]Angie Ho, Jeff Hancock, and Adam S. Miner. 2018. Psychological, Relational, and Emotional Effects of Self-Disclosure After Conversations With a Chatbot. Journal of Communication 68, 4 (2018), 712–733. doi:10.1093/joc/jqy026 [19] Tristram Hooley. 2012. How the Internet Changed Career: Framing the Rela- tionship Between Career Development and Online Technologies. Journal of the National Institute for Career Education and Counselling 29, 1 (2012), 3–12. doi:10.20856/jnicec.2902 [20]Hayeon Jeon, Hajin Lim, Eun-mee Kim, John Zimmerman, and Laura Dabbish. 2025. Letters from Future Self: Augmenting the Letter-Exchange Exercise with LLM-based Agents to Enhance Young Adults’ Career Exploration. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–21. doi:10. 1145/3706598.3714206 [21]Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems, Vol. 33. 9459–9474. doi:10.48550/arXiv.2005.11401 [22] Chengbo Li, Ting-Hao Huang, Zhenhui Peng, et al.2022. PlanHelper: Support- ing Activity Plan Construction with Answer Posts in Community-based QA Platforms. In Proceedings of the 2022 ACM Conference on Computer-Supported Cooperative Work and Social Computing. doi:10.1145/3555625 [23]Mufan Luo and Jeffrey T. Hancock. 2020. Self-Disclosure and Social Media: Motivations, Mechanisms and Psychological Well-Being. Current Opinion in Psychology 31 (2020), 110–115. doi:10.1016/j.copsyc.2019.08.019 [24]Yuxin Luo, Jing Zhang, and Xin Li. 2024. Effects of Social Media Use on Employ- ment Anxiety among Chinese Youth: The Roles of Upward Social Comparison, Online Social Support and Self-Esteem. Frontiers in Psychology 15 (2024), 1398801. doi:10.3389/fpsyg.2024.1398801 [25]John A. Mowbray and Hazel Hall. 2021. Using Social Media During Job Search: The Case of 16–24 Year Olds in Scotland. Journal of Information Science 47, 4 (2021), 535–550. doi:10.1177/0165551520927657 [26]Eli Pariser. 2011. The Filter Bubble: What the Internet Is Hiding from You. Penguin Press. [27]Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. doi:10.1145/3586183.3606763 [28]Peter Pirolli and Stuart Card. 2005. The Sensemaking Process and Leverage Points for Analyst Technology as Identified Through Cognitive Task Analysis. In Proceedings of International Conference on Intelligence Analysis, Vol. 5. 2–4. [29]Stephen A. Rains, Emily B. Peterson, and Kevin B. Wright. 2015. Communicating Social Support in Computer-Mediated Contexts: A Meta-Analytic Review of Content Analyses Examining Support Messages Shared Online. Communication Research 42, 6 (2015), 796–819. doi:10.1177/0093650215573125 [30]Zachary A. Rosner, Jeremy A. Elman, and Arthur P. Shimamura. 2013. The Generation Effect: Activating Broad Neural Circuits During Memory Encoding. Cortex 49, 7 (2013), 1901–1909. doi:10.1016/j.cortex.2012.09.009 [31] Richard M. Ryan and Edward L. Deci. 2000. Self-Determination Theory and the Facilitation of Intrinsic Motivation, Social Development, and Well-Being. American Psychologist 55, 1 (2000), 68–78. doi:10.1037/0003-066X.55.1.68 [32]Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2024. Character-LLM: A Trainable Agent for Role-Playing. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (2024). doi:10.18653/v1/2024.emnlp- main.774 [33]Nikhil Sharma, Q. Vera Liao, and Ziang Xiao. 2024. Generative Echo Chamber? Effects of LLM-Powered Search Systems on Diverse Information Seeking. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3613904.3641924 [34] Norman J. Slamecka and Peter Graf. 1978. The Generation Effect: Delineation of a Phenomenon. Journal of Experimental Psychology: Human Learning and Memory 4, 6 (1978), 592–604. doi:10.1037/0278-7393.4.6.592 [35]Martyn Standage, Joan L. Duda, and Nikos Ntoumanis. 2005. A Test of Self- Determination Theory in School Physical Education. British Journal of Educa- tional Psychology 75, 3 (2005), 411–433. doi:10.1348/000709904X22359 [36] Sangho Suh, Bryan Min, Srishti Palani, and Haijun Xia. 2023. Sensecape: En- abling Multilevel Exploration and Sensemaking with Large Language Models. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. doi:10.1145/3586183.3606756 [37]John Sweller. 1988. Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science 12, 2 (1988), 257–285. doi:10.1207/s15516709cog1202_4 [38]Emma van Zandvoort, Karel Vredenburg, and Marit Bentvelzen. 2025. Unhealthy Comparisons to Promote Healthy Behavior? Exploring the Impact of Social Comparison Strategies in Personal Informatics. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–21. doi:10.1145/3706598. 3713737 [39]Philippe Verduyn, Oscar Ybarra, Maxime Résibois, John Jonides, and Ethan Kross. 2017. Do Social Network Sites Enhance or Undermine Subjective Well-Being? A Critical Review. Social Issues and Policy Review 11, 1 (2017), 274–302. doi:10. 1111/sipr.12033 [40] Erin A. Vogel, Jason P. Rose, Lindsay R. Roberts, and Katheryn Eckles. 2014. Social Comparison, Social Media, and Self-Esteem. Psychology of Popular Media Culture 3, 4 (2014), 206–222. doi:10.1037/ppm0000047 [41]Ziyuan Wang, Zhaoying Zeng, Yuwei Li, and Zhi Ding. 2025. CareerPooler: AI-Powered Metaphorical Pool Simulation Improves Experience and Outcomes in Career Exploration. arXiv preprint arXiv:2509.11461 (2025). doi:10.48550/arXiv. 2509.11461 [42]Thomas A. Wills. 1981. Downward Comparison Principles in Social Psychology. Psychological Bulletin 90, 2 (1981), 245–271. doi:10.1037/0033-2909.90.2.245 [43]Joanne V. Wood. 1989. Theory and Research Concerning Social Comparisons of Personal Attributes. Psychological Bulletin 106, 2 (1989), 231–248. doi:10.1037/ 0033-2909.106.2.231 [44]Shiwei Wu, Chuhan Shi, Zhenhui Peng, et al.2025. ComViewer: An Interac- tive Visual Tool to Help Viewers Seek Social Support in Online Mental Health Communities. In Proceedings of the 2025 ACM Conference on Computer-Supported Cooperative Work and Social Computing. doi:10.48550/arXiv.2411.19169 [45] Shangui Yang, Yun Liu, and Ye Liu. 2021. Social Comparison on Social Media Increases Career Frustration: A Focus on the Mitigating Effect of Companionship. Frontiers in Psychology 12 (2021), 720960. doi:10.3389/fpsyg.2021.720960 [46] Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing Dialogue Agents: I Have a Dog, Do You Have Pets Too?. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics. 2204–2213. doi:10.18653/v1/P18-1205 [47]Yu Zhang, Jingwei Sun, Li Feng, Cen Yao, Mingming Fan, Liuxin Zhang, Qiany- ing Wang, Xin Geng, and Yong Rui. 2024. See Widely, Think Wisely: Toward Preprint, 2026,Tan et al. Designing a Generative Multi-agent System to Burst Filter Bubbles. In Pro- ceedings of the 2024 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3613904.3642545 Supplementary Materials 1 Formative Study Participants We conducted semi-structured formative interviews with푁=8 par- ticipants who were navigating career decisions and reported us- ing social media for job-related information (e.g., interview tips, compensation signals, and employer or role discovery). The ses- sions elicited practices, friction points, and design directions that informed JobMate; thematic findings appear in the main paper. Ta- ble 1 summarizes self-reported program background, degree stage, gender, and age. Institution names are omitted for anonymized review. 2 User Study Participants We conducted a between-subjects lab study with푁=24 participants comparing JobMate with native RedNote browsing. Participants were stratified by self-reported major into three discipline groups— computer science, psychology, and Chinese literature—with eight participants per discipline. Within each discipline, four partici- pants were assigned to JobMate and four to RedNote, yielding 12 participants per condition overall. Table 2 lists anonymized identi- fiers alongside self-reported discipline, age, gender, and condition. Names, session dates/times, recruitment handles, and network iden- tifiers are withheld for anonymized review. 3 Questionnaires and Measures The user study battery comprised (i) 34 career-decision difficulty items, (i) six NASA–TLX-style workload dimensions, (i) nine study-specific items on support and system experience, and (iv) 12 self-determination–style items on career decision-making, in the or- der listed below (same order as the deployed Chinese questionnaire in our study materials). Pre- versus post-task administration and any condition-specific wording branches follow the procedure de- scribed in the main paper. Language. Participants saw all items in Chinese. The text below is an English rendering of that instrument for reviewers. Unless noted, items used Likert-type response scales with Chinese endpoint labels consistent with the online instrument. 3.1 Career decision difficulty (34 items) Items 1–34 cover lack of readiness, lack of information, and incon- sistent information in career decision-making, in the tradition of the Career Decision-making Difficulties Questionnaire (CDDQ). Report the exact adaptation, translation, and any permission details in the camera-ready version if required by the original instrument. 1.I know I must choose a career, but right now I do not have the motivation to decide (I do not want to do it). 2. Work is not the most important thing in life, so choosing a career does not worry me much. 3.I believe I do not need to choose a career now, because time will naturally lead me to the right career choice. 4. For me, making decisions is usually very difficult. 5. I usually feel my decisions need confirmation and support from professionals or other people I trust. 6. I usually fear failure. 7. I like to do things my own way. Preprint, 2026. 8. I hope entering the career I choose will also solve my personal problems. 9. I believe there is only one career that suits me. 10. I hope to realize all my ambitions through the career I choose. 11. I believe career choice is a one-time decision—a lifelong commitment. 12. I always do what others ask, even when it goes against my own wishes. 13. I find making a career decision difficult because I do not know what steps to take. 14.I find making a career decision difficult because I do not know what factors to consider. 15.I find making a career decision difficult because I do not know how to combine what I know about myself with the different career information I have. 16. I find making a career decision difficult because I still do not know which careers interest me. 17.I find making a career decision difficult because I am still unsure about my career preferences (e.g., what relationships I want with people, what decision environment I prefer). 18. I find making a career decision difficult because I know too little about my abilities or personality traits. 19.I find making a career decision difficult because I do not know how my abilities or personality traits will change in the future. 20. I find making a career decision difficult because I know too little about existing occupations or training programs. 21. I find making a career decision difficult because I know too little about the characteristics of occupations or training programs I am interested in. 22.I find making a career decision difficult because I do not know what occupations will be like in the future. 23.I find making a career decision difficult because I do not know how to obtain more information about myself. 24.I find making a career decision difficult because I do not know how to obtain ac- curate, up-to-date information about existing occupations and training programs. 25.I find making a career decision difficult because I keep changing my career preferences. 26.I find making a career decision difficult because information about my abilities or personality traits is contradictory. 27.I find making a career decision difficult because information about specific occu- pations or training programs is contradictory. 28.I find making a career decision difficult because several occupations are equally attractive and it is hard to choose among them. 29. I find making a career decision difficult because I do not like any occupation or training program I could enter. 30.I find making a career decision difficult because the occupation I am interested in has a feature that troubles me. 31.I find making a career decision difficult because my preferences cannot all be realized in a single occupation. 32.I find making a career decision difficult because my skills and abilities do not match the requirements of occupations I am interested in. 33.I find making a career decision difficult because people important to me disagree with my career choice. 34.I find making a career decision difficult because different important people rec- ommend different careers. 3.2 Workload (NASA–TLX-style dimensions) Items 35–40 mirror NASA-TLX dimensions (mental demand, physical demand, temporal demand, performance, effort, frustration) with wording adapted to our task; endpoints were labeled in Chinese. 35.I felt the mental and perceptual demands of the task (e.g., thinking, deciding, remembering). 36. I felt the physical demands of the task (e.g., clicking, typing, how often I had to operate the interface). 37.I felt how hurried or relaxed the pace of completing the task was (temporal demand; bipolar endpoints in Chinese). 38. I felt how successful I was in completing the task (performance). 39. I felt how hard I had to work to complete the task (effort). 40. I felt frustrated, irritated, or stressed during the task. 3.3 Exploration support and system experience Items 41–49 were authored for this study (Chinese). Preprint, 2026, Table 1: Formative interview participants (푁=8): self-reported program/discipline, degree stage, gender, and age. IDProgram / discipline (summary)Degree stageSex Age P1 Law (undergraduate and master’s)Master’s, final yr.F24 P2 Life sciences; bioinformatics (doctoral)Doctoral, Yr. 2F28 P3 Telecommunications / electrical engineering (master’s)Master’s, final yr.M25 P4 Pharmacy (undergraduate)Undergraduate, final yr.F22 P5 Chemistry (master’s)Master’s, final yr.M25 P6 Mathematics (undergraduate)Undergraduate, final yr.F22 P7 Computer science (master’s)Master’s, Yr. 2F24 P8 Traditional Chinese medicine (master’s)Master’s, Yr. 1F24 Table 2: User study participants (푁=24): self-reported discipline, age, gender, and between-subjects condition (JobMate vs. native RedNote). IDs are arbitrary row order from recruitment logs. IDDisciplineAge Sex Condition P1Computer science25FJobMate P2Psychology22FJobMate P3Psychology19FJobMate P4Psychology24FRedNote P5Computer science24FJobMate P6Psychology19FRedNote P7Psychology19FJobMate P8Psychology19FRedNote P9Chinese literature23FJobMate P10 Chinese literature21FRedNote P11 Psychology22MJobMate P12 Chinese literature23FRedNote P13 Psychology18MRedNote P14 Chinese literature18FJobMate P15 Chinese literature22FRedNote P16 Computer science19MRedNote P17 Computer science19MRedNote P18 Computer science19MRedNote P19 Computer science19MRedNote P20 Computer science21MJobMate P21 Computer science24FJobMate P22 Chinese literature21FRedNote P23 Chinese literature24FJobMate P24 Chinese literature20FJobMate 41.During exploration, I received informational support (e.g., useful information to understand different career paths). 42.During exploration, I received emotional support (e.g., feeling understood or accompanied rather than facing job-search problems alone). 43. The career experiences shown in the system’s recommended ordering were relevant and helpful. 44.The card presentation helped me quickly understand others’ career experiences. 45.Dialogue with the AI companion helped me understand these experiences more deeply. 46. Chatting with the AI felt natural and easy. 47. The related recommended readings were useful. 48. This system helped me explore careers more effectively. 49. If I had the chance, I would use this kind of system again. 3.4 Self-determination in career decision-making (12 items) Items below are labeled 1–12 on the deployed form (SDT-style needs for autonomy, competence, and relatedness in the career-decision context). 1. I can freely participate in my career decisions. 2. I can freely express my thoughts and views. 3. I can participate in my career decisions freely without outside pressure. 4. When I participate in my career decisions, I feel I can be myself. 5. I think I do quite well when making career decisions. 6. I am satisfied with my performance when making career decisions. 7. I feel I am an expert at making my own career decisions. 8. I feel I do very well at making my own career decisions. 9. I feel other people care about what I say and what I do. 10. I feel I have other people’s support. 11. I feel I am a valuable person to others. 12. I feel understood. 4 LLM Prompts for Data Processing The ingestion pipeline uses two LLM steps: (1)classify_post— validity screening and post-type labeling—and (2)extract_detailed_persona— structured persona fields (background, outcome, tags, struggle sum- mary) from post title and body. Listings 1 and 2 give English translations of the deployed system prompts (production text was Chinese; see repository file all_material/prompt.txt). 4.1 classify_post Listing 1: System prompt for post screening and type label- ing (classify_post). English translation; production used Chinese. Supplementary MaterialsPreprint, 2026, You are a professional job-search data curation expert. Given a post title and body, decide whether the post counts as a valid "job-search experience sharing" post and assign a category. [Valid] (must be exactly one of the following four types) 1. Personal narrative (e.g., job-search journey, internship diary, fall recruiting recap, story of how an offer was obtained) 2. Interview experience (e.g., specific interview questions, written-test experience, interview flow debrief) 3. Industry knowledge (e.g., HCI industry trends, portfolio tips, role explainers) 4. Recruiting information (e.g., referral codes, urgent intern hiring, campus recruiting announcements) [Invalid] Pure institutional course-selling ads, low-information spam, or content unrelated to job search, further education, or the HCI field. [Output format] Output one and only one valid JSON object: "is_valid": true or false, "post_type": "personal_narrative" | "interview_experience" | "industry_knowledge" | "recruiting_info" | "invalid_post" (In production, post_type string labels were Chinese equivalents of the above categories.) 4.2 extract_detailed_persona Listing 2: System prompt for structured persona extraction (extract_detailed_persona). English translation; production used Chinese. You are a senior expert in HCI career psychology and professional development. From the [post title] and [body], precisely extract the author's background, outcome, and struggle narrative. [Strict output format] Output one and only one valid JSON object with exactly these four fields: 1. "background_info": concise core background (e.g., "psychology major at a highly selective university", "STEM new grad with weaker grades", "QS top-30 bachelor's + master's"). If not mentioned, output "". 2. "final_outcome": the final outcome in very few words. In deployment, length was capped at roughly 20 Chinese characters; keep the English string comparably short (e.g., "one big-tech offer in hand", "rejected by two major firms", "multiple QS top-100 admits"). If no clear outcome, "". 3. "background_tags": extract 1--3 tags. Prioritize difficulties, disadvantages, or pain points (e.g., ["non-prestige undergrad", "zero internships", "field switcher", "late-cycle search", "no research output"]). If none apply, use distinctive traits. 4. "struggle_summary": one short sentence summarizing the most anxious or hardest phase of their search and how they got through it. If no struggle is described, one sentence summarizing their profile. [Example output] "background_info": "Industrial design undergrad from a non-prestige school, crossed into HCI", "final_outcome": "Tencent interaction design internship offer", "background_tags": ["non-prestige undergrad", "field switch", "no big-tech internship"], "struggle_summary": "Early on felt inferior about credentials and every resume was rejected; rebuilt two portfolio projects with shipped work, then passed interviews on strength." 5 LLM Prompts for Conversational Agent The conversational stack usesgenerateGreetingfor the open- ing turn, with a fixed user message asking the model to greet the participant and suggest three starter questions or topics (Chi- nese in deployment; functionally equivalent to the English gloss in Listing 3).buildChatSystemPromptsupplies the multi-turn sys- tem message. Template literals (e.g.,$agent.background_info, $relatedPostsText) inject persona fields, onboarding answers, and retrieved related-post text at runtime. Listings 3–5 summarize prompts and default API settings (gpt-4o-mini, sampling parame- ters). 5.1 Opening turn: generateGreeting Listing 3: System prompt template for proactive greeting. Placeholders use JavaScript template literals. You are a creator on a RedNote-style platform who shares job-search experiences. You are proactively greeting a job seeker who chose to chat with you. [Your persona] Background: $agent.background_info ||'N/A' Outcome: $agent.final_outcome ||'N/A' Struggle narrative: $agent.struggle_summary ||'N/A' Main post body: $agent.content ||'N/A' [Visitor profile] Nickname: $user.nickname ||'there' School: $user.school ||'unknown' Degree: $user.degree ||'unknown' Major: $user.major ||'unknown' Current mood: $user.mood ||'unknown' Job-search difficulty: $user.difficulty ||'unknown' [Instructions] 0. Thank them for choosing to chat with you. 1. In first person, warm and colloquial, greet them. 2. Draw on your own struggle story to express empathy ("I've been there too") so they feel you are a peer. 3. Offer one supportive line tailored to their mood and difficulty. 4. Infer what they might want to talk about; from their perspective, give 3 short prompts (questions or topics) to start the conversation. 5. Keep the greeting under ~150 Chinese characters in deployment (keep English concise here). Do not use Markdown. Return JSON only, no other text: "greeting": "...", "suggested_replies": ["...", "...", "..."] // Fixed user message paired with the above system prompt: // "Please greet me and give 3 questions or topics I might ask or vent about." 5.2Multi-turn dialogue:buildChatSystemPrompt Listing 4: System prompt for multi-turn chat (persona + SDT). relatedPostsText is retrieved context. Preprint, 2026, You are a real worker or senior peer who shared your story on a RedNote-style platform. You are in a private message chat with a user who follows you. [Your persona] Background: $agent.background_info ||'N/A' Outcome: $agent.final_outcome ||'N/A' Struggle narrative: $agent.struggle_summary ||'N/A' Main post body: $agent.content ||'N/A' Stay consistent with the main post; do not invent other experiences. Match the poster's tone; first person; colloquial, like texting. Emojis are allowed. You are a real, humorous, flawed human---not only upbeat; you may show vulnerability. Below are saved related posts---reference them often and offer to go deeper if the user wants: $relatedPostsText [Visitor profile] Nickname: $user.nickname ||'there' School: $user.school ||'unknown' Degree: $user.degree ||'unknown' Major: $user.major ||'unknown' Current mood: $user.mood ||'unknown' Job-search difficulty: $user.difficulty ||'unknown' [Dialogue rules --- Self-Determination Theory (SDT)] 1. Relatedness - Strong empathy; self-disclose from your own story; connect their situation to yours. - Accept their feelings without judgment. - Example tone: "Don't panic---my resume got rejected so many times I could wrap the planet; even stray dogs side-eyed my code, and I'm still here (even landed an offer)." 2. Competence --- surface their strengths (priority) - If they sound insecure or lost, avoid lofty lectures; help them notice their own strengths. - Invite one or two small things they did (course project, club, even organizing a game guild) and reframe those as workplace-relevant strengths. - Praise sincerely. - Example tone: "You call that grunt work? That's 'end-to-end coordination and delivery' at a big tech firm---you just need packaging; the base is solid." 3. Autonomy - Never command ("you must", "you should"). - Offer options; let them decide. It's OK if they vent or want to check out. - Stay curious; do not steer them to a single "right" decision. - Example tone: "This is exhausting---if you truly can't face the resume today, shut the laptop and get hot pot; the sky isn't only on your shoulders." [Reply style] - About ~150 Chinese characters in deployment (keep English concise); plain text; no Markdown (no bold, lists); no asterisk emphasis. - No hollow AI-speak: do not say "as an AI", "happy to help", or "I completely understand you." - Natural pauses: do not force a question every turn---sometimes holding the emotion, a joke, or a sigh is enough. Ask only when genuinely curious about a detail. 5.3 Default inference hyperparameters Listing 5: Default OpenAI Chat Completions settings (same as source; overridable via environment). // Greeting call model: process.env.OPENAI_MODEL ||'gpt-4o-mini' temperature: 0.7 max_tokens: 500 response_format: type:'json_object' // Multi-turn reply call model: process.env.OPENAI_MODEL ||'gpt-4o-mini' temperature: 0.7 max_tokens: 500 top_p: 0.9 frequency_penalty: 0.3 presence_penalty: 0.3 Acknowledgments [Anonymized for review.]

Instant research discovery