Psy-Chronicle: Synthesizing Campus Counseling Grounded in a Student's Semester Timeline

arXiv:2605.22140 · cs.CL · Submitted 2026-05-21 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Psy-Chronicle: Synthesizing Campus Counseling Grounded in a Student's Semester Timeline".

Jane: Here's the gist of the paper. The gist The framework first constructs a four-dimensional student profile consisting of basic background, personality tendencies, family and social support, and core psychological conflicts.

Tom: Psy-Chronicle Framework Psy-Chronicle is proposed as a structured data generation framework for synthesizing long-horizon campus psychological counseling dialogues.

Jane: It models counseling as a longterm process driven by a stable student profile, a temporal stress event graph, cross-session counseling interactions, and structured memory updates For each synthetic student i, Psy-Chronicle first constructs a structured student profile: Pi = Bi, Ti, Si, Ci (1) where Bi, Ti, Si, and Ci denote the student’s demographic and academic background, personality tendency, family and social support background, and core psychological conflict, respectively Based on the student profile, the framework then generates a temporal stress event graph: Gi = (Vi, Ei) where Vi = vi,1, vi,2,..., vi,T is a sequence of stress event nodes and Ei represents temporal and developmental dependencies among events Psy-Chronicle then simulates cross-session counseling along the event graph.

Tom: At session t, the dialogue is generated based on the student profile Pi, the current stress event vi,t, and the accumulated historical memory Mi,t−1: Di,t = Φ(Pi, vi,t, Mi,t−1) (3) After each session, Psy-Chronicle compresses the dialogue into a structured memory summary and updates the historical memory: mi,t = Update(Mi,t−1, Ψ(Di,t, vi,t)) (4) The overall closed-loop process is Pi → Gi → Di,t T t=1 → Mi,t (5) Components of the Framework The framework consists of four stages: student profile construction, temporal stress event graph generation, cross-session counseling simulation, and dataset post-processing with quality control The student profile construction follows a “stratified sampling–constrained generation–consistency filtering” pipeline The core prompt used for profile generation requires the output to include basic information, personality traits, family and social background, and a core psychological conflict related to the stress domain The temporal stress event graph generation takes the student profile as input and generates a list of events covering a 16-week semester Each event node records the event content, occurrence week, stress intensity, and post-event psychological state The core prompt for event graph generation requires events to be logically coherent and reflect the Core Conflict in the profile Counseling Simulation and Memory Integration Psy-Chronicle simulates cross-session counseling along the event trajectory, where each session corresponds to a current stress event while inheriting relevant information from previous sessions through structured memory The student agent generates help-seeking expressions conditioned on the current event, personality tendency, family and social support background, core conflict, and historical counseling memory To improve the structure and progression of counselor responses, a progressive CBT-guided prompting strategy is introduced Each session is divided into three phases: establishment and exploration, deep exploration, and CBT-informed intervention The student agent prompt instructs the model to produce state-driven expression constrained by the current event and psychological impact The counselor agent is guided by a progressive CBT stage instruction, which dictates specific tasks for each phase, such as "Ask at most one question in each response" and using Socratic guidance After each round, Psy-Chronicle compresses the dialogue into a structured memory summary and updates the historical memory: Mi,t = Update(Mi,t−1, mi,t) (13) This updated memory is then used as input for the next counseling session to enable cross-session continuity and reduce factual drift Dataset and Evaluation Based on Psy-Chronicle, CPCD is constructed, a Chinese long-horizon dialogue dataset for college psychological counseling scenarios The dataset covers multiple high-frequency campus stress domains and explicitly preserves the correspondence among student profiles, stress event chains, and counseling dialogues CPCD-Bench is proposed as a long-horizon campus psychological counseling benchmark that evaluates model capabilities from three aspects: session-level response, long-horizon memory recall, and temporal-causal reasoning The evaluation tasks include Session-level Response (SR), Memory Recall (MR), and Temporal-Causal Reasoning (TCR) Evaluation for SR focuses on Empathy, Coherence, and Professionalism MR tests factual retrieval from long counseling histories, focusing on Accuracy, Completeness, Temporal Consistency, and No Hallucination TCR evaluates the ability to organize and explain the long-term development of student distress through Temporal Accuracy and Causal Coherence Experimental results show that CPCD effectively improves session-level response generation and long-horizon memory recall for models with the same base architecture Ablation studies verify the importance of temporal stress event graphs and structured memory mechanisms for event-chain modeling and cross-session historical information utilization Psy-Chronicle provides a structured path for long-horizon campus psychological counseling data synthesis and offers foundational resources for event-driven psychological support, cross-session counseling memory modeling, and campus mental health model training The paper concludes that models fine-tuned on CPCD effectively improve session-level counseling response generation and long-horizon memory recall under the same base model setting However, improvements on TCR remain limited, suggesting that event-chain organization, causal explanation, and long-context information integration are key problems to be addressed in the future Future work may introduce explicit event-chain annotations or graph-structured reasoning mechanisms to enhance models’ understanding of the evolution of psychological distress The paper also discusses ethical considerations, noting that CPCD is synthetic and should be used with clear safety boundaries The final dataset is obtained after post-processing and quality control, including checks for structural completeness, semantic consistency, diversity, campus-scenario appropriateness, and safety boundaries The paper concludes that longhorizon campus psychological counseling modeling still faces significant challenges regarding temporal-causal reasoning The overall findings indicate that the structured pipeline provides more effective long-horizon counseling training signals than direct dialogue generation The paper emphasizes that event-chain organization and causal explanation are key challenges in long-horizon psychological counseling modeling Future research may explore more effective memory compression or hierarchical retrieval methods so that smaller models can more stably handle cross-session counseling histories The paper concludes that Psy-Chronicle provides a structured path for long-horizon campus psychological counseling data synthesis and offers foundational resources for event-driven psychological support, cross-session counseling memory modeling, and campus mental health model training The final dataset is obtained after post-processing and quality control, including checks for structural completeness, semantic consistency, diversity, campus-scenario appropriateness, and safety boundaries The paper concludes that longhorizon campus psychological counseling modeling still faces significant challenges regarding temporal-causal reasoning The overall findings indicate that the structured pipeline provides more effective long-horizon counseling training signals than direct dialogue generation The paper concludes that event-chain organization and causal explanation are key challenges in long-horizon psychological counseling modeling Future research may introduce explicit event-chain annotations or graph-structured reasoning mechanisms to enhance models’ understanding of the evolution of psychological distress<ref:26

Paper summary: Tom: So, wrapping up on Psy-Chronicle: it provides a structured path for long-horizon campus psychological counseling data synthesis and offers foundational resources for event-driven psychological support and cross-session counseling memory modeling >

Jane: The paper focuses heavily on building this pipeline—the student profile, the temporal stress event graph, the interactive simulation—and then validating it with CPCD and CPCD-Bench >

Lu: The main implication is that we need to move away from training models just on single turns or short dialogues toward training them on data that explicitly models the sequence and dependencies of life events >

Meng: I see. It suggests that for real-world applications in mental health, we need to focus less on massive datasets of short chats and more on structured, long-horizon simulation pipelines like this one >

Tom: Exactly. The finding that models fine-tuned on CPCD improve session-level counseling response generation and long-horizon memory recall under the same base model setting is a solid result for how we structure training signals >

Jane: But we also have to keep in mind that the paper itself flags limitations, specifically regarding temporal-causal reasoning, which means getting models to explain the *why* behind the long-term changes is still a challenge >

Conclusion: Tom: So, we’ve been looking at Psy-Chronicle, and to wrap up on this segment: this framework takes counseling data from just single sessions and builds something that lasts across an entire semester for a student >

Jane: Yeah, so it's basically modeling the whole timeline of a student's stress and how those past events affect what they say in later sessions >

Tom: The authors are trying to ground this synthesis in a concrete timeline, using things like the academic schedule and specific stress events that happen over weeks >

Lu: What’s interesting here is how they structure the input. They start with a profile—demographics, personality, family stuff—and then map out these temporal stress nodes for the whole semester >

Meng: I wonder how robust that timeline generation actually is when you try to scale it up to a huge university counseling center >

Jane: That’s a good point. The paper shows they use a "stratified sampling" approach for building those profiles, and then they generate events that are supposed to be logically coherent with the core conflict in the profile >

Tom: And on the results side, they tested this against existing single-session corpora, and Psy-Chronicle outperformed them in working alliance scores >

Lu: The real kicker is when you look at how a model trained on this CPCD dataset actually performs; it connects events from earlier weeks more often than a model trained without those dependencies >

Jane: So what does that mean for the world, Tom? It means we’re moving toward training AI that understands long-term psychological patterns rather than just reacting to the immediate moment >

Tom: Yeah, it suggests that for real-world mental health AI, you need to show it the whole story of a student's experience >

Jane: Right. And while they showed improvement in response and memory recall, the authors themselves admit there’s still work left to do on temporal-causal reasoning—really explaining *why* things change over time >

Chaogui Gou

University of Science and Technology Beijing

cs.CL

Submitted: 2026-05-21

Updated: 2026-10-20

Code: https://github.com/EdwinUSTB/Psy-Chronicle

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: The gist The framework first constructs a four-dimensional student profile consisting of basic background, personality tendencies, family and social support, and core psychological conflicts.

Key concepts

Student Profile Pi: This is a four-dimensional structure representing a synthetic student, including basic demographics, personality traits, family/social support systems, and core psychological conflicts. This profile serves as the stable foundation for generating all subsequent events and counseling interactions throughout the simulation.
Temporal Stress Event Graph Gi: This graph maps out a semester's progression by sequencing stress events (Vi) with their temporal and developmental dependencies (Ei). It ensures that the simulated counseling sessions follow a logical, evolving timeline where later issues are influenced by earlier ones in the student's life.
Structured Memory Mi,t: This mechanism carries historical information across counseling sessions. After each session, Psy-Chronicle compresses the dialogue into a summary and updates this memory. This allows subsequent sessions to reference past events and maintain continuity without needing to re-read the entire history.
CPCD-Bench: This is a benchmark designed to test AI models on long-horizon counseling tasks. It evaluates three areas: responding appropriately in a single session, accurately recalling information from long histories (Memory Recall), and understanding the cause-and-effect relationship over time (Temporal-Causal Reasoning).

Terminology

Causal Reasoning — 2× in this paper · explained in 2 episodes

Summary

The gist The framework first constructs a four-dimensional student profile consisting of basic background, personality tendencies, family and social support, and core psychological conflicts.

Psy-Chronicle Framework

Psy-Chronicle is proposed as a structured data generation framework for synthesizing long-horizon campus psychological counseling dialogues. It models counseling as a longterm process driven by a stable student profile, a temporal stress event graph, cross-session counseling interactions, and structured memory updates For each synthetic student i, Psy-Chronicle first constructs a structured student profile: Pi =

Bi, Ti, Si, Ci (1) where Bi, Ti, Si, and Ci denote the student’s demographic and academic background, personality tendency, family and social support background, and core psychological conflict, respectively Based on the student profile, the framework then generates a temporal stress event graph: Gi = (Vi, Ei) where Vi =

vi,1, vi,2,..., vi,T is a sequence of stress event nodes and Ei represents temporal and developmental dependencies among events Psy-Chronicle then simulates cross-session counseling along the event graph. At session t, the dialogue is generated based on the student profile Pi, the current stress event vi,t, and the accumulated historical memory Mi,t−1: Di,t = Φ(Pi, vi,t, Mi,t−1) (3) After each session, Psy-Chronicle compresses the dialogue into a structured memory summary and updates the historical memory: mi,t = Update(Mi,t−1, Ψ(Di,t, vi,t)) (4) The overall closed-loop process is Pi → Gi →

Di,t T t=1 →

Mi,t (5)

Components of the Framework

The framework consists of four stages: student profile construction, temporal stress event graph generation, cross-session counseling simulation, and dataset post-processing with quality control The student profile construction follows a “stratified sampling–constrained generation–consistency filtering” pipeline The core prompt used for profile generation requires the output to include basic information, personality traits, family and social background, and a core psychological conflict related to the stress domain The temporal stress event graph generation takes the student profile as input and generates a list of events covering a 16-week semester Each event node records the event content, occurrence week, stress intensity, and post-event psychological state The core prompt for event graph generation requires events to be logically coherent and reflect the Core Conflict in the profile

Counseling Simulation and Memory Integration

Psy-Chronicle simulates cross-session counseling along the event trajectory, where each session corresponds to a current stress event while inheriting relevant information from previous sessions through structured memory The student agent generates help-seeking expressions conditioned on the current event, personality tendency, family and social support background, core conflict, and historical counseling memory To improve the structure and progression of counselor responses, a progressive CBT-guided prompting strategy is introduced Each session is divided into three phases: establishment and exploration, deep exploration, and CBT-informed intervention The student agent prompt instructs the model to produce state-driven expression constrained by the current event and psychological impact The counselor agent is guided by a progressive CBT stage instruction, which dictates specific tasks for each phase, such as Ask at most one question in each response and using Socratic guidance After each round, Psy-Chronicle compresses the dialogue into a structured memory summary and updates the historical memory: Mi,t = Update(Mi,t−1, mi,t) (13) This updated memory is then used as input for the next counseling session to enable cross-session continuity and reduce factual drift

Dataset and Evaluation

Based on Psy-Chronicle, CPCD is constructed, a Chinese long-horizon dialogue dataset for college psychological counseling scenarios The dataset covers multiple high-frequency campus stress domains and explicitly preserves the correspondence among student profiles, stress event chains, and counseling dialogues CPCD-Bench is proposed as a long-horizon campus psychological counseling benchmark that evaluates model capabilities from three aspects: session-level response, long-horizon memory recall, and temporal-causal reasoning The evaluation tasks include Session-level Response (SR), Memory Recall (MR), and Temporal-Causal Reasoning (TCR) Evaluation for SR focuses on Empathy, Coherence, and Professionalism MR tests factual retrieval from long counseling histories, focusing on Accuracy, Completeness, Temporal Consistency, and No Hallucination TCR evaluates the ability to organize and explain the long-term development of student distress through Temporal Accuracy and Causal Coherence Experimental results show that CPCD effectively improves session-level response generation and long-horizon memory recall for models with the same base architecture Ablation studies verify the importance of temporal stress event graphs and structured memory mechanisms for event-chain modeling and cross-session historical information utilization

Conclusion

Psy-Chronicle provides a structured path for long-horizon campus psychological counseling data synthesis and offers foundational resources for event-driven psychological support, cross-session counseling memory modeling, and campus mental health model training The paper concludes that models fine-tuned on CPCD effectively improve session-level counseling response generation and long-horizon memory recall under the same base model setting However, improvements on TCR remain limited, suggesting that event-chain organization, causal explanation, and long-context information integration are key problems to be addressed in the future Future work may introduce explicit event-chain annotations or graph-structured reasoning mechanisms to enhance models’ understanding of the evolution of psychological distress The paper also discusses ethical considerations, noting that CPCD is synthetic and should be used with clear safety boundaries The final dataset is obtained after post-processing and quality control, including checks for structural completeness, semantic consistency, diversity, campus-scenario appropriateness, and safety boundaries The paper concludes that longhorizon campus psychological counseling modeling still faces significant challenges regarding temporal-causal reasoning The overall findings indicate that the structured pipeline provides more effective long-horizon counseling training signals than direct dialogue generation The paper emphasizes that event-chain organization and causal explanation are key challenges in long-horizon psychological counseling modeling Future research may explore more effective memory compression or hierarchical retrieval methods so that smaller models can more stably handle cross-session counseling histories The paper concludes that Psy-Chronicle provides a structured path for long-horizon campus psychological counseling data synthesis and offers foundational resources for event-driven psychological support, cross-session counseling memory modeling, and campus mental health model training The final dataset is obtained after post-processing and quality control, including checks for structural completeness, semantic consistency, diversity, campus-scenario appropriateness, and safety boundaries The paper concludes that longhorizon campus psychological counseling modeling still faces significant challenges regarding temporal-causal reasoning The overall findings indicate that the structured pipeline provides more effective long-horizon counseling training signals than direct dialogue generation The paper concludes that event-chain organization and causal explanation are key challenges in long-horizon psychological counseling modeling Future research may introduce explicit event-chain annotations or graph-structured reasoning mechanisms to enhance models’ understanding of the evolution of psychological distress<ref:26

Improvements for AI systems

Bold Header: Structured Long-Horizon Data Generation Framework

This framework integrates student profiles, a temporal stress event graph, cross-session counseling simulation, and structured memory integration into a unified pipeline, enabling the generation of dialogues with longterm backgrounds, temporal continuity, and logical problem evolution.

Bold Header: CPCD Dataset for Contextual Training

The construction of CPCD provides a data foundation for studying longhorizon psychological counseling modeling by explicitly preserving the correspondence among student profiles, stress event chains, and counseling dialogues, making it the first publicly available Chinese long-horizon dialogue dataset for this domain.

Bold Header: CPCD-Bench for Multi-Dimensional Evaluation

CPCD-Bench evaluates models across three dimensions—session-level response, long-horizon memory recall, and temporalcausal reasoning—revealing that current models still face significant challenges in event-chain organization and causal explanation.

Bold Header: Session Response Enhancement via CBT Simulation

The simulation of counseling sessions utilizes a progressive CBT-guided prompting strategy, where the counselor agent follows three stages—rapport building and problem exploration, deep empathy and clarification, and CBT-informed intervention—to generate structured, stage-wise responses.

Bold Header: Long-Term Memory Retention Mechanism

The system employs a memory update mechanism where the updated memory is then used as part of the context for subsequent sessions, enabling cross-session continuity and reducing factual drift during long-horizon dialogue generation.

Bold Header: Temporal Causal Reasoning Improvement

Ablation studies confirm that removing the event graph consistently degrades performance, especially on MR and TCR, demonstrating that integrating the temporal stress event graph is crucial for models to learn event progression and developmental dependencies.