acceptodds
Under review as a conference paper at ICLR 2027

PsyMem: Role-Asymmetric Information Flow for Longitudinal Counseling Dialogue Synthesis

Abstract

Longitudinal counseling-style dialogue research requires data that capture substantive interaction within sessions and continuity across sessions, yet authentic linked records are difffcult to obtain and share. We introduce PsyMem, a synthesis framework combining persona-grounded client simulation, reffectionguided utterance generation, and role-asymmetric cross-session state updates. The client retains a ffxed private persona and an evolving concern state, while the counselor carries forward memory derived from preceding dialogues. PsyMem produces 150 linked four-session sequences comprising 600 sessions, with an average of 61 turns per session. We assess within-session quality using a 15- item WAI-derived observer rubric and introduce the Memory Continuity Analysis Framework (MCAF) to evaluate cross-session continuity in stored role summaries. Under GPT-4o evaluation, PsyMem scores 6.04 on WAIOS, versus 5.71 for the best evaluated synthetic baseline on this metric, and 4.96/5 on MCAF. Removing cross-session state lowers MCAF by about 3.9 points; among variants retaining such state (MCAF: 4.94–4.95), the role-asymmetric design attains the highest observed WAIOS (6.04 vs. 5.97–6.02). Across six open-source base models, models ffne-tuned on PsyMem-generated dialogues perform best in most MentalChat16K comparisons, while WAIOS results vary by model and evaluator.

Primary area:
datasets and benchmarks
Keywords:
large language models, cross-session continuity, synthetic data generation, dialogue evaluation, psychological dialogue synthesis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KNe8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work tackles multi-session counseling by building both a simulated client and a counselor. The client updates its state after every session using the session dialogue, its state before that session, and its persona. The counselor updates its state using the session dialogue and its state before the session; it doesn't get access to the client's states or persona. They use this setup, along with 150 client personas, to create a dialogue dataset consisting of 4 sessions per client. They experiment with finetuning different models on this created dataset and compare their performance with the same models finetuned on other datasets. The authors also introduce a metric to measure cross-session coherence.

Critical strengths
  1. The approach looks interesting and simulates what would happen in real multi-session encounters. The client's state is affected by the conversation and their character, while the counselor can only learn through the conversation and can only infer the client's evolving state from it.

  2. The proposed MCAF metric is an important addition as a measure of cross-session coherence. To my knowledge, no other metric exists for this purpose.

  3. The released multi-session datasets might be used to build better synthetic clients/counselors.

Critical weaknesses
  1. The MCAF metric, which is an important contribution of this work, is not human-validated. I did notice that this is acknowledged in the limitations section, but the authors human-validated the within-session metrics, which is great, so it raises the question of why they skipped validating the reliability of their proposed cross-session metric.

  2. The ablation suggests that other approaches for state tracking perform similarly to the more complex pipeline proposed by the authors, which means that the added complexity might be unnecessary and using summaries of previous sessions might suffice.

Questions for authors

for the ablation study, did the authors compute statistical significance? If so, are these differences (for the memory-enabled configurations) statistically significant?

3 comments
AuthorsOpenReview ↗

Response to Reviewer KNe8

We thank the reviewer for the thoughtful comments.

W1 — Human validation of MCAF. We supplemented a human evaluation of the cross-session metric. Eleven evaluators performed a double-blind evaluation on 30 randomly sampled groups: 10 PsyMem, 10 native human PsychEval [1], and 10 negative examples. Negatives were built by concatenating Session 1 and Session 4 of the same case, or sessions from unrelated cases; their surface topics may appear coherent, but they contain identity and memory inconsistencies.

We used boundary anchoring: PsychEval as the high-quality reference, concatenated negatives as the low-quality reference, with PsyMem between them for calibration. Agreement between human ratings and MCAF at both anchors supports MCAF's validity. Evaluators used a 1–5 scale (1 = lacking coherence, 5 = coherent) across four MCAF-aligned dimensions: memory consistency, state evolution, alliance continuity, and overall coherence.

Model VariantMemory (Q1)State Evol. (Q2)Alliance (Q3)Overall (Q4)
PsyMem554.54.5
PsychEval5555
Negative Examples1121

Mean over 10 dialogues/condition; Q1–Q3 are the core dimensions underlying MCAF.

Results: PsyMem 4.5/5, PsychEval 5/5, negatives 1/5; MCAF also assigns lower scores to negatives. Human and automatic MCAF ratings show consistent trends at both anchors, supporting MCAF's reliability. Representative excerpts (PsyMem / PsychEval / negative) will be included in the revision.

W2 — Is the added complexity unnecessary? MCAF is slightly higher for PsyMem (4.96 vs. 4.95), and scores are close across configurations; MCAF primarily measures memory coherence and may not fully capture differences in generation behavior. We therefore also analyze interaction length and task alignment.

VariantWAIOS Goal ↑WAIOS Task ↑MCAF Avg. ↑Words/Utt. ↑Turns ↑
PsyMem (Full)5.955.924.9610561
Plain Text5.89 (−0.06)5.84 (−0.08)4.94 (−0.02)92 (−13)23 (−38)
Summary5.93 (−0.02)5.92 (±0)4.95 (−0.01)87 (−18)33 (−28)
Shared Memory5.89 (−0.06)5.75 (−0.17)4.95 (−0.01)85 (−20)34 (−27)

Full in absolute values; parentheses = difference vs. Full; ±0 = no change.

The difference in interaction length is clearer: PsyMem averages 61 turns vs. 33 (Summary) and 34 (Shared Memory), with 105/87/85 words per utterance. On WAIOS Task, PsyMem matches Summary (5.92 vs. 5.92) and is 0.17 above Shared Memory (5.75). This is consistent with the role-asymmetric design: client-private state and counselor dialogue-derived memory are updated separately, so the two roles need to keep exchanging new information in later sessions. When summaries or memory are shared, interaction length drops substantially (28 and 27 fewer turns), suggesting that information asymmetry is one factor sustaining long-horizon interaction, rather than redundant complexity.

Q1 — Statistical significance in the ablation. Given the number of variants and limited per-variant sample size, we primarily report effect sizes rather than making per-variant significance claims. In the larger-scale downstream fine-tuning experiments we did conduct a systematic statistical analysis, including multi-judge ratings and correlation analysis (e.g., GPT-4o vs. DeepSeek-V4-Flash/Qwen3.6), to assess stability. From the table above, the role-asymmetric memory module maintains strong coherence while mainly affecting interaction behavior and task alignment: the clearest effect is interaction length (61 vs. 33/34 turns, i.e., 28/27 fewer; 105 vs. 87/85 words per utterance), with WAIOS Task matched to Summary (5.92) and +0.17 over Shared Memory (5.75).

References:

[1] Pan et al. (2026). PsychEval. Findings of ACL 2026, pp. 22192–22212.

I hope our clarification was helpful. If anything remains unclear, we would be glad to provide a brief additional clarification before the discussion period closes.

I thank the authors for their response.

W1: Thanks for the clarification.

W2: The response here reinforces my point: there really isn't a noticeable difference between the proposed approach and simpler approaches; the only significant differences are in words/utterance and number of turns. One cannot argue that it's better (or worse) to have a bigger value for either of these two metrics; what matters is the effectiveness of the session, which is measured by the other metrics.

On statistical significance. I disagree: If I understand correctly from this work, there are 150 personas; one can still measure statistical significance with that number.

Overall, I appreciate the work and maintain my score.

Reviewer_RshvOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

To address the lack of linkage across sessions in counseling dialogue, this paper proposes a dialogue simulation and corpus generation method that combines persona-grounded client simulation, reflection-guided utterance generation, and role-asymmetric cross-session state updates. Its core contribution is an information-asymmetric way of modeling the client's and the counselor's memory separately. The authors further release a cross-session dataset PsyMem as well as an evaluation framework, MCAF. This work is therefore of value for advancing research on AI counseling models.

Critical strengths
  1. The research motivation, which includes using an asymmetric memory mechanism to simulate both client and counselor behaviors and employing dedicated variables such as the concern state to represent the client's evolving concerns, is valuable and aligns well with real-world application scenarios.
  2. The authors validate the proposed method across multiple different fine-tuned model architectures, and the proposed MCAF also provides a helpful reference for evaluating cross-session psychological dialogue corpora.
  3. The authors honestly discuss the limitations of their current approach and provide an anonymized open-source code repository.
Critical weaknesses
  1. While information between the Counselor and Client should indeed be asymmetric, a professional counselor will not remain permanently oblivious to the client's underlying psychological issues. The empathy and suggestions offered by the counselor should professionally align with the client's latent psychological problems. Such a fundamental constraint is missing in this work. The counselor's advice can completely diverge from the client's actual issues without the pipeline ever detecting it, this directly undermines the professional quality and reliability of the synthesized corpus.
  2. The behavioral guidelines for the Counselor in Section 3.3 are vague: inspecting the appendix prompt reveals that selecting among the four strategies is left entirely to the model's discretion, lacking theoretical justification or selection logging. The concept of "session phase" it relies on is also ill-defined. It appears merely as textual ranges in the prompt, without passing turn counts or phase labels to the model, nor applying any turn-count verification.
  3. All samples in the PsyMem dataset consist of exactly 4 sessions, and are generated using the exact same prompt template; during a session, the Counselor does not know the current session index, and cross-session transitions rely entirely on p_i and m_i. This setup is overly simplistic, making it difficult to capture progression or developmental trajectories across multiple sessions, thus failing to simulate realistic counseling scenarios. Moreover, p_i only has a formal definition (Eq. 1) , a substantive/content definition is missing. In other words, what value space it can take, what distribution it follows, and how it evolves across the four sessions remain completely unknowable.
  4. The underlying model used across all disclosed stages of generation and evaluation is GPT-4o. Although the authors acknowledge the risk, the MCAF and the persona identity check rely solely on GPT-4o as a single judge, and self-preference and self-bias in LLM-as-a-judge have been widely documented.
  5. The baselines in the main experiment are all single-session data, none of them spans multiple sessions, such a comparison is missing and should be reported. The authors should evaluate against existing open-source multi-session counseling frameworks cited in their own related work, such as PsychEval.

[1]Panickssery et al. LLM Evaluators Recognize and Favor Their Own Generations. NeurIPS 2024. [2]Wataoka et al. Self-Preference Bias in LLM-as-a-Judge. arXiv:2410.21819, 2024. [3]Ye et al. Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge. ICLR 2025.

Questions for authors

In Appendix F.2, the memory extraction prompt instructs the model to generate summaries " based on the current session content and with appropriate reference to the previous memory." This directly enforces cross-summary consistency during extraction, which is the exact property MCAF's four dimensions evaluate. Does this constitute a form of evaluation leakage? Would MCAF inherently assign high scores to almost any memory extracted using this prompt?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Rshv

We thank the reviewer.

W1 — Counselor advice vs. latent issues. We target general psychological/emotional support, not clinical psychopathology, and keep role asymmetry to model a counselor inferring state from dialogue only. Generation (App. F.1): the first 20–25 turns elicit background, personality, history, and needs without early conclusions/advice; pacing follows the injected Resistance Level; ≤1 scenario-specific suggestion per turn, no generic advice. Evaluation: MCAF is a strict post-hoc diagnostic (cross-session demographic contradictions count as different individuals; incoherent progression scores 1). PsyMem reaches Within-session Alignment 4.95/5.0 and MCAF 4.96/5.0 (human 5.0); no obvious deviation is observed.

W2 — Strategy selection, phase, logging. All constraints are in the App. F.1 Counselor prompt. (1) The four strategies are theory-grounded: Empathy (humanistic; medium/high resistance, emotion-tied, not used as a standalone single-turn response), Reframing (CBT; maladaptive thoughts in real context), Guidance (motivational interviewing; ≤1 context-specific suggestion/turn), Actionable Support (later; concrete execution details); routing uses the Resistance Level. (2) We avoid hard if turn > 25: coding: App. F.1 uses soft temporal anchors (early 20–25, middle 20–25, late 10–15 turns) because counseling is fluid; the model infers phases for boundary robustness. (3) No [Phase: X] labels are emitted; transparency comes from the generate-check-rework loop, with rework flags/failure reasons logged in re{idx}_{visit}.json, plus MCAF Alignment.

W3 — Four sessions and pip_i. Sessions end by natural closure or truncation (App. A.3), not fixed length; each averages 61 turns (~244 per sequence). Four sessions match the 4–6 session short-term support cycle [1]. Beyond Eq. (1), pip_i is a structured state with three fields: Core Concern & Scenarios (primary concern → 3–5 scenarios: events, emotions, behaviors), Resistance Level ({low, medium, high}, controlling behavior evolution), Facts & Suggestions (key facts, received suggestions). For FclientF_{client}, after each session an LLM combines the original background with the last three turns; structured prompts extract facts and issues (cognitive progress) to update the state. pip_i thus follows explicit state-transition logic, not an unspecified distribution.

W4 — Single-judge bias. We do not rely solely on GPT-4o: DeepSeek-V4-Flash and Qwen3.6 are additional automatic evaluators (Sec. 4.2, Table 17, App. E.1). Inter-judge correlations are relatively high (WAIOS: r = 0.79 with DeepSeek, 0.74 with Qwen3.6; MentalChat16K: r = 0.92). Automatic and human WAIOS ratings correlate moderately (r = 0.58, p = 0.025, n = 15 dataset–subscale pairs); ranking is consistent.

W5 — Native multi-session baselines. We add two non-concatenated native multi-session references, PsychEval (human) [2] and Psy-Chronicle (synthetic) [3]; AnnaAgent/MusPsy release no directly evaluable native synthetic datasets.

DatasetTypeMCAF ↑PPL ↓Distinct-2 ↑
PsychEvalNative Human5.021.80.51
Psy-ChronicleNative Synthetic3.823.10.57
PsyMem (Ours)Native Synthetic4.924.80.52

MCAF does not reward only PsyMem (4.9 vs. 5.0/3.8).

Q1 — Evaluation leakage? "Appropriately reference prior memory" (App. F.2) models semantic continuity in real long-term counseling, not metric inflation; leakage would mean injecting prior-session raw text into generation or evaluation. (i) Counselor memory is structured (facts, personality, suggestions, issue evolution), not verbatim concatenation. (ii) The suggestions field is not injected into the counselor's current context (Sec. 3.4); it serves later updates and post-hoc diagnostics — removing cross-session state drops MCAF by ~3.91 and the identity indicator by 28 points (Table 1). (iii) MCAF is post-hoc, not a generation objective, and discriminates native data (4.9/5.0/3.8). Human validation: 11 evaluators (App. G.1) rated 30 groups (10 each: PsyMem, PsychEval, negatives), 1–5, four MCAF dimensions.

Model VariantMemory (Q1)State Evol. (Q2)Alliance (Q3)Overall (Q4)
PsyMem554.54.5
PsychEval5555
Negative Examples1121

Means over 10 dialogues/condition.

Results: PsyMem 4.5/5, PsychEval 5/5, negatives 1/5; MCAF assigns 1.0 to negatives. Human and MCAF trends align; no obvious leakage.

References:

[1] Robinson et al. (2020). The dose-response effect in routinely delivered psychological therapies. Psychotherapy Research 30(1):79–96. [2] Pan et al. (2026). PsychEval. Findings of ACL 2026, pp. 22192–22212. [3] Gou (2026). Psy-Chronicle. arXiv:2605.22140.

I hope our clarification was helpful. If anything remains unclear, we would be glad to provide a brief additional clarification before the discussion period closes.

Dear reviewer:

Thank you again for your tremendous effort and valuable feedback. We have responded carefully and thoroughly to your main questions and hope you are satisfied with our response. Given the limited time for rebuttal, we would like to know if you have any further concerns, which we would be happy to address. We look forward to receiving your further feedback.

Sincerely,

Authors

Reviewer_6Zj6OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PsyMem, a framework for synthesizing longitudinal counseling dialogues through role-asymmetric information flow, maintaining a private client persona and evolving concern state alongside dialogue-derived counselor memory. The authors generate 600 sessions across 150 four-session sequences and propose MCAF to evaluate cross-session continuity. Experiments show improved dialogue-quality scores and strong summary-level continuity, while downstream fine-tuning results are promising but vary across models and evaluators.

Critical strengths
  1. Multi-session settings and new evaluation metrics. Current counseling dialouges are generally of single session. This paper propose a multi-session dataset while providing several evaluation metrics.
  2. Clear framework design: PsyMem explicitly separates the client's private state from the counselor's dialogue-derived memory, providing a structured approach to role-specific information management.
  3. Experiments cover multiple datasets and settings, with detailed descriptions. Human evaluations of dataset quality are also included.The authors also provide source code of the project.
Critical weaknesses
  1. Unfair comparision in multi-session settings. One major contribution of this paper stands on multi-session dataset generation and evaluation. However, the comparision between PsyMem and other baselines (e.g. PsyQA) is unfair from my perspective. Baseline datasets are of single session and the authors combined 4 individual sessions to construct the multi-session dialogue. Besides, MCAF is also quite tailored for PsyMem, which makes it not surprising to see the evaluation results of MCAF. The authors are encouraged to use MCAF on other multi-session datasets, or construct PsyMem-individual with 4 independent sessions as another baseline and re-calculate the evaluation metrics.
  2. Unclear inprovement in ablation studies. In Table 1, ablation results of MCAF results among different settings are only within minor improvements. Some ablation settings even surpass the proposed methods. It's still unclear whether these evaluation results are statistically significant among various settings. Besides, the authors are encouraged to provide more details about the design of Plain Text, Summary, Shared Memory and emphasize the difference between these ablation settings and PsyMem.
Questions for authors
  1. How did baseline LLMs fine-tuned and generated dialogues? Did you fine-tuned separate models as counselor and user or just one single model to generate the whole dialogue?
  2. What is the stopping criteria of each session in datsset generation process.
  3. How do the authors establish that MCAF reliably measures meaningful longitudinal continuity rather than superficial consistency between summaries? Could the metric be validated using human judgments, more challenging negative examples, or complete dialogue transcripts?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 6Zj6

We thank the reviewer for the detailed comments.

W1 — Fairness of the multi-session comparison. Prior cross-session work (MusPsy, AnnaAgent) does not consistently release reproducible native multi-session data; differences in scenario, language, and temporal structure preclude comparison under identical MCAF/WAIOS and fine-tuning protocols. We keep concatenated single-session baselines as a controlled comparison and add two non-concatenated native multi-session references, PsychEval (human) [1] and Psy-Chronicle (synthetic) [2]. PsyMem scores 4.9, close to PsychEval (5.0) and above Psy-Chronicle (3.8): MCAF does not reward only PsyMem. Distinct-2 is comparable (0.52 vs. 0.51); higher PPL (24.8 vs. 21.8) is descriptive only, as PPL is length/lexicon-dependent.

DatasetTypeMCAF ↑PPL ↓Distinct-2 ↑
PsychEvalNative Human5.021.80.51
Psy-ChronicleNative Synthetic3.823.10.57
PsyMem (Ours)Native Synthetic4.924.80.52

MCAF 1–5; ↑/↓ = higher/lower better.

W2 — Ablation magnitude and variant design. Given limited per-variant samples, we report effect sizes rather than per-variant significance. The clearest separation is interaction length: 61 vs. 33 (Summary) and 34 (Shared Memory) turns, with 105/87/85 words per utterance; Summary matches Full on WAIOS Task (5.92) at ~54% of its turns. As MCAF mainly measures memory coherence, we also analyze length and task alignment.

VariantWAIOS Goal ↑WAIOS Task ↑MCAF Avg. ↑Words/Utt. ↑Turns ↑
PsyMem (Full)5.955.924.9610561
Plain Text5.89 (−0.06)5.84 (−0.08)4.94 (−0.02)92 (−13)23 (−38)
Summary5.93 (−0.02)5.92 (±0)4.95 (−0.01)87 (−18)33 (−28)
Shared Memory5.89 (−0.06)5.75 (−0.17)4.95 (−0.01)85 (−20)34 (−27)

Full in absolute values; parentheses = difference vs. Full; ±0 = no change.

The variants differ in the granularity and direction of cross-session information transfer: Plain Text concatenates full prior raw dialogues (noisy, subject to truncation); Summary shares LLM summaries with both roles (removes redundancy, discards structured state); Shared Memory uses one shared structured memory (removes role asymmetry); PsyMem (Full) lets the counselor see only dialogue-derived memory while the client keeps a private persona and concern state. These details will be added to the revision. Separate updates lead the two roles to exchange new information later, sustaining longer interactions.

Q1 — Fine-tuning / generation setup. (i) Generation: GPT-4o with distinct system prompts drives Client and Counselor role-play. (ii) Fine-tuning: a single model [3], LoRA-tuned on one base model; different prompts condition the same model for both roles — not two separate models (Appendix B.1). (iii) Evaluation: the fine-tuned model is instantiated as both roles via different prompts (the PsyMem entry in the original table).

Q2 — Session stopping criteria. (1) Natural Closure (first) — a session ends when either role considers the issue clarified, or a termination keyword fires (e.g., "Let's stop here for today"; Appendix F.1, code release). (2) Hard Limit — absent natural closure, at turn 100 the counselor proposes closing at turn 99; at the 150-turn bound it is truncated to control cost and avoid loops.

Q3 — MCAF validity beyond summary consistency. We added a human evaluation with challenging negatives: 11 evaluators (Appendix G.1) rated 30 groups (10 PsyMem, 10 PsychEval, 10 negatives splicing Session 1 with Session 4 of the same case or unrelated cases) on 1–5 scales across four MCAF-aligned dimensions. Boundary anchoring: PsychEval as high-quality and negatives as low-quality anchors; agreement of human ratings and MCAF at both anchors supports validity.

Model VariantMemory (Q1)State Evol. (Q2)Alliance (Q3)Overall (Q4)
PsyMem554.54.5
PsychEval5555
Negative Examples1121

Mean over 10 dialogues/condition; Q1–Q3 are the core MCAF dimensions.

Results: negatives 1/5, PsyMem 4.5/5, PsychEval 5/5; MCAF likewise scores negatives low. Human and MCAF trends align — MCAF captures longitudinal continuity rather than surface consistency.

References:

[1] Pan et al. (2026). PsychEval. Findings of ACL 2026, pp. 22192–22212. [2] Gou (2026). Psy-Chronicle. arXiv:2605.22140. [3] Qiu & Lan (2026). Interactive Agents. *SEM 2026, pp. 410–427.

I hope our clarification was helpful. If anything remains unclear, we would be glad to provide a brief additional clarification before the discussion period closes.

Dear reviewer:

Thank you again for your tremendous effort and valuable feedback. We have responded carefully and thoroughly to your main questions and hope you are satisfied with our response. Given the limited time for rebuttal, we would like to know if you have any further concerns, which we would be happy to address. We look forward to receiving your further feedback.

Sincerely,

Authors

Thanks for your reply. I would like to keep my rating after seeing the discussions from other reviewers.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair,

We thank you for handling our submission, and Reviewers 6Zj6, Rshv, and KNe8 for careful, technically engaged reviews. We submitted point-by-point responses with new experiments or analyses for every major concern. Reviewers 6Zj6 and KNe8 engaged with our responses in the discussion. Reviewer Rshv did not comment after our response, and no further discussion took place on that thread. We ask you to weigh this when deciding, and to assess our responses directly where they differ from the reviews.

Strengths

Role-asymmetric information flow. PsyMem keeps the client's persona and evolving concern state private while the counselor sees only dialogue-derived memory, so the counselor infers state from the interaction itself. This is the structural source of sustained long-horizon exchange: 61 turns on average versus 33 (Summary) and 34 (Shared Memory).

Longitudinal data and evaluation. We release 600 sessions across 150 four-session sequences and introduce MCAF for cross-session continuity, validated against human judgments with boundary anchoring and challenging negatives, and checked across multiple automatic judges.

Breadth of evidence. Multiple datasets and fine-tuning settings, two native non-concatenated multi-session references, component ablations, an 11-annotator human study, and released source code.

Common concerns and how we addressed them

Fairness of the multi-session comparison, and whether MCAF only rewards PsyMem. We added two non-concatenated native multi-session references: PsychEval (human) [1] scores 5.0 on MCAF, Psy-Chronicle (synthetic) [2] 3.8, and PsyMem 4.9; Distinct-2 is comparable (0.51 / 0.57 / 0.52). Concatenated single-session baselines are retained only as a controlled comparison, and PPL is reported as descriptive because it depends on length and lexicon. This directly addresses reviewer 6Zj6's W1, and reviewer Rshv's W5.

Whether the added complexity is justified, and whether ablation differences are meaningful. With limited per-variant samples we report effect sizes rather than per-variant significance claims, and we define all four variants by the granularity and direction of cross-session information transfer. The clearest separation is interaction length (61 vs. 33/34 turns; 105 vs. 87/85 words per utterance), while WAIOS Task matches Summary (5.92) and leads Shared Memory by 0.17. In the larger downstream fine-tuning study we additionally ran multi-judge correlation analysis (r = 0.79 / 0.74 / 0.92 across judges; human–automatic r = 0.58, p = 0.025). This directly addresses reviewer 6Zj6's W2, and reviewer KNe8's W2 and Q1.

Whether MCAF measures longitudinal continuity rather than surface consistency, and whether evaluation leaks. Eleven evaluators double-blind rated 30 groups (10 PsyMem, 10 native human PsychEval, 10 negatives splicing Session 1 with Session 4 or unrelated cases) on four MCAF-aligned dimensions. With boundary anchoring, negatives score 1/5, PsyMem 4.5/5 and PsychEval 5/5, and MCAF likewise ranks negatives lowest. Counselor memory is structured rather than verbatim, the suggestions field is not injected into the current context, MCAF is post-hoc rather than a generation objective, and removing cross-session state lowers MCAF by ~3.91 and the identity indicator by 28 points. This directly addresses reviewer 6Zj6's Q3, reviewer KNe8's W1, and reviewer Rshv's Q1.

Framework and generation details. We clarified the theory-grounded strategy routing via Resistance Level, soft temporal anchors instead of hard turn thresholds, rework logging in re{idx}_{visit}.json, the three-field structured state pip_i and the FclientF_{client} update rule, and the 4–6 session short-term support rationale for four sessions [3]. Sessions end by natural closure, with a turn-100 closing prompt and a turn-150 truncation bound; generation uses role-specific GPT-4o prompts, and fine-tuning uses a single LoRA model conditioned by prompts for both roles rather than two separate models. This directly addresses reviewer Rshv's W1, W2 and W3, and reviewer 6Zj6's Q1 and Q2.

Reliability of the automatic evaluators. Beyond GPT-4o, DeepSeek-V4-Flash and Qwen3.6 serve as additional judges; inter-judge correlations are relatively high (WAIOS r = 0.79 / 0.74; MentalChat16K r = 0.92) and the ranking is consistent. This directly addresses reviewer Rshv's W4.

We will revise the manuscript accordingly and are glad to provide any further material.

Sincerely, Authors

References:

[1] Pan et al. (2026). PsychEval. Findings of ACL 2026, pp. 22192–22212. [2] Gou (2026). Psy-Chronicle. arXiv:2605.22140. [3] Robinson et al. (2020). The dose-response effect in routinely delivered psychological therapies. Psychotherapy Research 30(1):79–96.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.