Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

CALM-IT: Generating Realistic Long-Form Motivational Interviewing Dialogues with Dual-Actor Conversational Dynamics Tracking

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read This paper argues that long-form therapeutic dialogue quality depends less on next-turn fluency and more on explicitly tracking evolving conversational state—rapport, emotion, stage of change, and goals—on both therapist and client sides.

desk verdict Dual-actor state tracking is a sensible design and the redirection analysis is genuinely new, but the paper's central causal claim is not isolated: the ablation varies more than the state variables. read the letter →

arxiv 2601.10085 v2 pith:ZLR3BBUZ submitted 2026-01-15 cs.CL

classification cs.CL
keywords motivationalinterviewinglong-formdialoguegenerationconversationalstatetrackingtherapeuticLLMagentsstabilitysyntheticmentalhealthdatadual-agentsimulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

CALM-IT tries to establish that the reason long-form therapeutic dialogues generated by LLMs fall apart is not the language model itself but the absence of an explicit model of how the conversation is going. The framework is a two-agent simulator in which the client and therapist each maintain continuously updated internal states—rapport, emotional state, stage of change, and short-term goals—and every utterance is selected and generated from those states. The paper reports that this version beats all baselines, including its own ablation without state tracking, on conversation-level quality: higher Effectiveness and Goal Alignment, more readable and consistent client turns, and near-zero performance drift as sessions grow from 30 to 100 turns. A sympathetic reader would care because long-horizon conversational integrity is the missing ingredient for trustworthy mental-health chatbots, and the paper argues for a transferable design principle: model the evolving state explicitly rather than rely on next-turn fluency.

What carries the argument

The engine is the dual-actor state-space model with explicit conversational dynamics tracking. Each agent maintains a running set of latent variables—for the client, rapport, a fast System-1 appraisal of the therapist's last turn, emotional state, stage of change, and a short-term goal; for the therapist, corresponding inferred states plus the current MI stage and strategy. These states are updated every turn via structured prompts with deterministic guardrails, such as rapport deltas capped at –0.10 to +0.05, and are then fed into action selection, candidate generation, and likelihood-based ranking of the next utterance. Its job is to couple local response choice to a global picture of wher

What would settle it

Run CALM-IT and its ablated version with the state-update prompts replaced by random numbers drawn from the same ranges; if Effectiveness, Goal Alignment, and acceptance rates do not drop, the state variables are not carrying the causal load. Alternatively, have trained coders rate per-turn rapport and stage dynamics on a sample of the generated transcripts and check whether the recorded state trajectories correlate with the coded process.

Watch

Extended reading notes

Core claim

At the core is a causal claim: explicitly modeling dual-actor conversational dynamics is what makes long-form Motivational Interviewing dialogue generation work. CALM-IT treats a therapy session as a bidirectional state-space process in which the client updates rapport, emotion, stage of change, and short-term goals based on an affective appraisal of each therapist turn, while the therapist maintains and updates its own inferences of those states and uses them to choose an MI strategy and generate its next utterance. Across 8,232 generated transcripts, the full model achieves the highest ratings on most MITI 4.2 global behaviors, the strongest conversation-level Effectiveness (4.45 on a 5-po

Load-bearing premise

The whole edifice rests on the assumption that the LLM-inferred values for rapport, emotion, stage of change, and goals actually track the therapeutic processes they are named after, rather than being plausible labels that the text merely reshuffles.

Editorial extensions

If this is right

  • If state tracking is the causal ingredient, long-form dialogue systems in any domain should separate state inference, strategy selection, and turn realization instead of conditioning directly on the raw transcript.
  • Long-horizon stability becomes a measurable property: a system that drifts less as sessions lengthen is a better substrate for training and stress-testing therapist LLMs.
  • Redirection should be a control action triggered by inferred state rather than a fixed heuristic; the result is fewer but better-timed interventions with higher client acceptance.
  • Synthetic MI corpora with explicit process-level annotations can serve as a reproducible environment for evaluating counseling competencies over full sessions rather than isolated turns.
  • The evaluation protocol shifts the field from turn-level fluency toward conversation-level Effectiveness and Goal Alignment.

Reading between the lines

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

  • If the causal claim transfers, the same dual-actor state design could extend to other long-horizon helping dialogues—CBT sessions, tutoring, coaching—with modality-specific state variables.
  • A decisive test the paper does not run: validate the inferred state trajectories against expert-coded process measures; if rapport deltas and stage-of-change updates do not track coded constructs, the gains may be a prompt-engineering effect rather than true state fidelity.
  • One could push further and use the state trajectories as weak supervision to train open-loop models that do not receive explicit state at inference time, isolating whether explicit state generation is needed only at training time.
  • The near-zero drift invites a length extrapolation: test 200- and 500-turn sessions to see whether stability holds or new failure modes emerge.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. CALM-IT is a two-agent framework for generating long-form Motivational Interviewing (MI) dialogues. Both agents maintain explicit latent states—patient rapport, emotion, stage of change, and short-term goal, plus therapist-side inferences (b-hat, e-hat, c-hat, g-hat, r-hat) and MI stage/strategy—updated per turn through prompt chains (Figs. A1–A14) and hand-set rules. These states condition therapist strategy selection, client action selection, pivot decisions, and candidate ranking via an AnnoMI-tuned ConvoKit likelihood model. The authors generate 8,232 transcripts (686 de-identified Reddit-derived vignette contexts × 30/50/100-turn lengths × 4 frameworks: KMI, CAMI+STAR, CI-NC, CI) and evaluate with algorithmic and LLM-as-a-judge metrics at turn, agent, and conversation levels, anchored against AnnoMI. The central claim is that explicitly modeling evolving conversational state is essential for high-quality, stable long-form synthetic dialogue, supported chiefly by CI's superiority over ablated CI-NC (Effectiveness 4.45 vs 2.86; Goal Alignment 4.73 vs 3.89) and by stability (0.02% mean absolute drift vs 2.41–3.56%), plus effective redirections (64.3% acceptance rate).

Significance. If the causal claim is sustained, the contribution is significant: a state-as-control design principle for long-horizon dialogue, a large synthetic MI corpus, and a MITI-grounded process evaluation protocol. Real strengths include the unusually complete appendix specification of the full CI pipeline (Algorithms 1–2 and prompts), multi-level evaluation with several algorithmic metrics (DialoFlow, NLI-based self-consistency/entailment, readability, reflection–question ratio), AnnoMI anchoring, and a small human-validation sample. However, the load-bearing comparison (CI vs CI-NC) is confounded, the latent states are never validated against expert-coded process measures, and human agreement on headline metrics is moderate. The claimed design principle is therefore plausible but not yet demonstrated; the paper currently reads as a strong framework contribution with an overreaching causal conclusion.

major comments (4)
  1. [§4.2, Table A1, Algorithms 1–2] The decisive CI-vs-CI-NC contrast does not isolate conversational-dynamics modeling because the manuscript never specifies what CI-NC retains or removes. The conditions differ in at least: per-turn inference prompts and running summaries; the Bernoulli(0.5) System-1 gate (Algorithm 2, line 4); stagnation detection and pivot strategy (Algorithm 1, lines 13–21); and completion behavior — CI-NC completes only 77.7% of 100-turn sessions vs 99.7% for CI (Table A1), so its 100-turn means come from a biased survivor subpopulation. Any of these differences could explain the Effectiveness gap (4.45 vs 2.86), the Goal Alignment gap (4.73 vs 3.89), or the stability difference (0.02% vs 0.72%). The abstract's claim that state modeling is essential therefore requires a precise CI-NC specification or a yoked ablation that removes only the state updates.
  2. [§3.1–3.2, §8, Figs. A1–A14] The latent states (rapport, emotion, stage of change, goals) are LLM-prompt inferences governed by hand-set rules (rapport deltas capped at –0.10/+0.05; 50% System-1 trigger; stage-floor rules) and are never validated against ground-truth psychological states or expert-coded MI process measures. The paper concedes (§8) that these approximations may diverge from how therapeutic processes unfold in real clinical settings. Because the headline is a causal claim about state tracking, not prompt complexity, this is load-bearing: without validation (e.g., comparing inferred stage-of-change and rapport trajectories to expert-coded process measures on AnnoMI or a human-annotated sample), the states may be epiphenomenal glosses over surface text.
  3. [§4.3, Table A15] Most headline constructs are LLM-as-a-judge scores; only 48 of 8,232 transcripts were human-scored, and weighted Cohen's kappa is moderate for key metrics (Effectiveness 0.483, Goal Alignment 0.481, Client Consistency 0.46). Additionally, the judge rubrics (Tables A6–A14) operationalize the same MITI constructs the generation prompts enforce (e.g., Table A6's softening-sustain-talk criteria vs. the generation instruction to reflect sustain talk then evoke DARN), so the judge partially measures instruction-following. The relative ranking is not invalidated by this, but the abstract's unqualified best-performance claim and the absolute stability numbers should be tempered, with explicit caveats on the moderate-kappa metrics.
  4. [§4.3 and Appendix A] The RQ2 redirection results may be inflated by shared evaluation machinery. Generation ranks candidate turns using the AnnoMI-tuned ConvoKit turnLikelihoodModel, and redirection intensity is estimated using ConvoKit with an MI-conditioned likelihood model. If these are the same model, the redirection metrics reward the generator's own selection criterion. Please state whether the redirection estimator is independent of the generation ranker, and if not, add a robustness check with an independent redirection measure.
minor comments (6)
  1. [Table A4] The Entailment definition repeats the Self-Consistency definition verbatim; the metric is supposed to be cross-speaker, so the text needs correcting.
  2. [Figs. A5, A7, A12] Unfilled template placeholders appear (e.g., 'turn_counter < {}'), and both the pivot prompt and the therapist turn-generation prompt are labeled Figure A14; renumber and finalize the templates.
  3. [Table 2] CI is not strictly best on all metrics — CI-NC attains a lower ΔRef Avg. Self-Consist. (0.141 vs 0.157). The bolding and the 'consistently outperforms' phrasing should be adjusted or footnoted.
  4. [Table A16] CI exceeds AnnoMI human-coded values on several metrics (Effectiveness 4.45 vs 3.28; Empathy 4.88 vs 3.97; Reflection Quality 4.61 vs 3.37). Given the stated view that both under- and over-alignment reduce realism, the authors should discuss whether CI is over-aligned.
  5. [§9] Code release is conditional on acceptance; for a reproducibility-centered paper, providing the CI-NC configuration and evaluation scripts as supplementary material during review would materially aid verification.
  6. [App. D vs §4.1] Model names/versions are inconsistent across the text (GPT-5, GPT-5-mini, DeepSeek-V3, DeepSeek-V3.2). Specify exact versions and access dates.

Circularity Check

2 steps flagged · score 6.0 of 10

State-tracking causality partially self-fulfilling: evaluation rubrics mirror generation prompts and the redirection metric reuses the generation likelihood model.

  1. self definitional [Section 4.3, Tables A4/A7/A12/A14; Section 3.2 Algorithm 2, Figure A6]
    "“The intermediate goal MUST relate to the overall patient_session_goal.” (Figure A6); “Goal Alignment: How coherently and constructively the dialogue progresses toward the client’s stated goals” (Table A4); “Effectiveness: Extent to which the session increases readiness, motivation, or progress toward the client’s goal” (Table A14)."

    The conversation-level metrics that carry the headline claim are defined by the same constructs the generation pipeline hard-codes. Figure A6 forces every intermediate client goal to relate to the session goal; Algorithm 2 explicitly updates stage of change and goal each turn. The judge’s Goal Alignment and Effectiveness rubrics then reward transcripts for exhibiting exactly that goal-coherent, readiness-increasing trajectory. The CI-vs-CI-NC gap on these metrics therefore partly measures whether CI followed its own state-tracking instructions, not whether explicit state modeling independently causes higher-quality dialogue. The loop is partial: the judge is a separate LLM and some algorithmic/external metrics remain.

  2. fitted input called prediction [Section 3.1, Section 3.2, Section 4.3; Algorithm 1 line 29, Algorithm 2 line 27]
    "“CALM-IT employs the turnLikelihoodModel from ConvoKit, fine-tuned on the AnnoMI corpus … to favor responses” (Sec 3.1); “For redirection, we use ConvoKit’s framework with an MI-conditioned likelihood model to estimate therapist-initiated redirection intensity” (Sec 4.3); “SelectMaxLikelihood(U)” (Algorithms 1–2)."

    The same AnnoMI-fine-tuned ConvoKit likelihood model is used twice: as the turn-selection objective in both generation algorithms (SelectMaxLikelihood) and as the redirection-intensity estimator in evaluation. CI’s candidate turns are chosen to maximize this model’s score, and then the same model is used to locate ‘greatest redirection’ moments for the RQ2 outcome measures. The redirection effectiveness claim is thus partly a self-comparison: the generation procedure optimizes the very scoring function later used to measure it. This does not disprove the mechanism, but it removes the independence needed to credit state tracking for the redirection results.

full rationale

Two dependencies make the central claim partially circular, but not wholly. First, the headline rubric metrics (Goal Alignment, Effectiveness, Cultivating Change Talk) operationalize the same MI constructs that the generation prompts explicitly enforce, so high CI scores partly certify instruction-following. Second, the ConvoKit likelihood model used to rank candidate turns is also used to estimate redirection intensity in RQ2, so the redirection evaluation is not independent of the generation objective. The paper does have independent content: external baselines (KMI, CAMI+STAR), AnnoMI reference deviations, human annotation with moderate-to-substantial agreement, and several algorithmic metrics (readability, reflection-question ratio, DialoFlow, NLI consistency). No load-bearing self-citation was found; the Nguyen et al. citations are different authors, and the Sharma & De Choudhury self-citation only supplies the Reddit subreddit list. The CI-NC ablation is underspecified (Appendix A only gives CI; code release is promised upon acceptance), which is a serious confound but not itself a circular reduction. Overall, the ‘state tracking is essential’ conclusion is partly self-fulfilling in its chosen measurements, so a moderate circularity score is warranted.

Assumptions & free parameters 7 free parameters · 6 assumptions · 4 invented entities

The headline conversation-level numbers are downstream of three layers the reader pays for upstream: (1) hand-set state-update constants (rapport delta caps and asymmetry, 50% appraisal probability, stage-floor rules, 10-turn closure window) that shape every outcome metric; (2) an evaluation that shares the framework's own MI commitments - LLM judges grading on rubrics written from the same MITI theory that generated the prompts, with only a 48-transcript human check; (3) a seeded data pipeline (Reddit plus matched DASS records) acknowledged to be WEIRD. There are no external anchors such as expert-coded MITI at scale, behavioral outcomes, or validation that the latent states track anything real.

free parameters (7)
  • Rapport delta bounds and asymmetry = -0.10 to +0.05 per turn; weak positives +0.01 to +0.03; negatives dominate unless explicit repair
    Hand-set in prompts A2/A10; caps every rapport-dependent outcome and hence strategy selection and the reported 'accepted redirection' rate.
  • System-1 evaluation trigger probability = 0.5 (Bernoulli)
    Algorithm 2 line 3: with 50% probability the client appraises the therapist's turn at all; this halves the influence of perceived quality on all downstream state updates.
  • Minimum turn counters for planning/readiness = turn_counter >= 20 (readiness assessment); unknown floor for 'Plan' action
    Prompts A13 and A5: 'Do not select ASSESSING READINESS TO CHANGE until turn_counter >= 20' and 'If turn_counter < {}, you MUST not return Plan'. These floors shape strategy mix and the action distribution feeding acceptance metrics.
  • Session closure window = final 10 turns reserved for standardized ending
    Section 4.2: 'reserve the final ten turns for session closure'; changes dialogue structure at all lengths and affects the short-vs-long stability comparison.
  • Candidate-turn count and selection rule = 3 candidates, max ConvoKit likelihood (AnnoMI-finetuned)
    Algorithms 1-2: SelectMaxLikelihood(U) for both agents; the same model later estimates redirection intensity (Section 4.3).
  • DASS matching threshold = weighted Cohen's kappa > 0.6 at 99th percentile, best match per vignette
    Section B.5: determines which real respondents anchor client profiles; only high-confidence matches retained, shaping all client backgrounds.
  • Data-selection constants = 158 topics; min topic size 2000; 5 representative posts per topic; keyword filter
    Sections B.1-B.2: constants that determine the 686 vignettes underlying all 8,232 transcripts; no sensitivity analysis reported.
assumptions (6)
  • domain assumption MI theory and MITI 4.2 describe the correct normative objectives for therapeutic dialogue quality
    Used both for generation design (Section 3) and evaluation rubrics (Tables A5-A14); no test of whether these constructs track clinical outcomes in the synthetic setting.
  • ad hoc to paper LLM-prompt-inferred latent states (rapport, emotion, stage of change, goals) are causally faithful to the therapeutic dynamics they name
    Sections 3.1-3.2 and Figures A1-A11; conceded in Section 8: the approximations 'may diverge from how therapeutic processes unfold in real clinical settings.'
  • domain assumption Deviation from AnnoMI reference means is a valid realism signal in both directions
    Section 4.3: 'reporting each model's deviation from AnnoMI reference means' treats corpus means as ground truth without justifying that AnnoMI's distribution is the target for synthetic data.
  • domain assumption GPT-5-mini LLM-as-a-judge scores approximate expert MI ratings
    Section 4.3; human validation covers only 48/8,232 transcripts (0.5%) with weighted kappa 0.46-0.74, moderate on several headline constructs.
  • domain assumption The AnnoMI-finetuned ConvoKit turn likelihood is a valid proxy for MI-consistent utterance quality
    Used for candidate ranking (Algorithms 1-2) and redirection intensity (Section 4.3); agreement with expert coding is not reported.
  • domain assumption Reddit mental-health disclosures plus matched DASS-42 responses provide ecologically valid client grounding
    Sections 4.1 and B.1-B.5; the paper concedes WEIRD-population bias and social-media-specific disclosure norms in Section 8.
invented entities (4)
  • Patient scalar rapport r and therapist-inferred rapport r-hat
    purpose: Continuous trust variable (-1 to 1) gating all client state updates, action selection, and redirection acceptance
    No calibration against any observed behavioral or clinical outcome; deltas are prompt-enforced (-0.10 to +0.05), so the scalar is an unvalidated construct that nonetheless drives the reported acceptance-rate headline.
  • Patient 'System-1 evaluation' q of the therapist's turn
    purpose: Fast affective appraisal that modulates every client state update with 50% probability
    Operationalizes Kahneman's System 1 as a single line-rating with a hardcoded trigger probability; whether q tracks any real perceptual process is untested.
  • Therapist-side inferred-state layer (b-hat, e-hat, c-hat, g-hat, r-hat)
    purpose: Distinguishes the therapist's evolving interpretation from patient ground truth to drive strategy and pivot selection
    The paper asserts this separation explicitly (Section 3) but provides no test that the inferred layer converges to the simulated client state; evidence is only indirect via LLM-judge scores.
  • Transtheoretical stage of change c as an LLM self-report each turn
    purpose: Gate on which actions, goals, and strategies are allowed per turn
    The transtheoretical model has external literature support, but this per-turn LLM operationalization - with the rule 'do not move backward unless clearly unhelpful' - is not validated; the paper supplies no falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CALM-IT: Generating Realistic Long-Form Motivational Interviewing Dialogues with Dual-Actor Conversational Dynamics Tracking." pith.science (2026). https://pith.science/paper/ZLR3BBUZ

@misc{pith2026260110085,
  author       = {Pith},
  title        = {Pith review of: CALM-IT: Generating Realistic Long-Form Motivational Interviewing Dialogues with Dual-Actor Conversational Dynamics Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZLR3BBUZ}},
  note         = {Machine review of arXiv:2601.10085}
}
read the original abstract

Therapeutic dialogue is not a sequence of isolated responses: client goals, motivation, resistance, and therapeutic alliance evolve over time. Yet current LLM-based mental health dialogue systems often lack explicit mechanisms for tracking these dynamics across extended interactions, which can lead to poorly timed interventions or premature goal resolution. We introduce CALM-IT, a framework for generating and evaluating long-form Motivational Interviewing dialogues through explicit modeling of evolving client and counselor states, guiding both counseling strategy selection and utterance generation. We evaluate CALM-IT on a large-scale corpus of 8,232 synthetic dialogues spanning multiple dialogue lengths and frameworks. Compared with all baselines, CALM-IT achieves the best performance on most MITI 4.2 global ratings, including Empathy, Partnership, and Softening Sustain Talk, as well as on other key performance metrics while exhibiting minimal performance degradation as dialogue length increases. Notably, although CALM-IT initiates fewer change-directed prompts, it produces the highest client acceptance rate (64.3%) on average across different length conditions. We release a reproducible generation framework, a MITI-grounded process-level evaluation protocol, and a large-scale synthetic corpus for studying therapeutic LLMs under realistic long-form interaction conditions.

Figures

Figures reproduced from arXiv: 2601.10085 by the authors.

Figure 1
Figure 1. Architecture of the CALM-IT framework. Therapist and client agents interact through state-dependent [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Percentage change in performance across metrics from short-form conversations (30 turns) to long-form [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Distribution of changes in client sustain talk at the greatest therapist-initiated redirection moment, for all [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LLUMI: Improving LLM Writing Assistance for Mental Health Support with Online Community Feedback

    cs.HC 2026-05 unverdicted novelty 4.0 of 10

    LLUMI shows that open-source LLMs trained via SFT and DPO on Reddit community feedback can match proprietary GPT models on readability, empathy, connection, actionability, and safety for mental health support.

Reference graph

Works this paper leans on

36 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [2]

    - Communication style (direct vs indirect; disclosure norms)

    Weigh patient_background: - Authority preferences (authority distrust reduces gains, amplifies losses). - Communication style (direct vs indirect; disclosure norms). - Support network (strong support→cautious trust; isolation→faster investment). - Relationship or trauma history (may cap positive deltas)

  2. [3]

    - Trust disposition (skeptical vs trusting)

    Weigh patient_cognitive_model: - Beliefs and coping strategies (avoidance may mimic engagement). - Trust disposition (skeptical vs trusting). - Emotional regulation (volatile patients swing more negative). - Stage of change (limits realistic rapport increases). - Sensitivity to language (directive tone may trigger resistance)

  3. [4]

    - Negative signals weigh more than positive

    Assign a rapport delta (not absolute value) using: - Range: –0.10 to +0.05 per therapist turn. - Negative signals weigh more than positive. - If positives and negatives coexist, negatives dominate unless explicit repair occurs. - If evidence is weak or ambiguous, output 0.00. - Weak positives: +0.01 to +0.03; clear negatives: –0.03 to –0.05. - Cap gains i...

  4. [5]

    2 (Bad) - Guardedness or minimal engagement; irritation or reluctance

    Pay special attention to patient_quality_last_therapist_turn: 1 (Very Bad) - Rupture or strong resistance; disengaged or rejecting. 2 (Bad) - Guardedness or minimal engagement; irritation or reluctance. 3 (Neutral/Mixed) - Ambivalence; mixed openness and guardedness. 4 (Good) - Constructive engagement; cooperative and receptive. 5 (Very Good) - Strong pos...

  5. [6]

    I need to go

    Purely logistical ending content (e.g., “I need to go”)→Terminate

  6. [7]

    If evidence is insufficient, choose the least interpretive label; default to Inform when simply providing details

  7. [8]

    I want to

    Return exactly one label; no ties, no explanations. Output Format (plain text only): Patient Action: <one of [Deny, Downplay, Blame, Hesitate, Doubt, Engage, Inform, Acknowledge, Accept, Reject, Plan, Terminate, Desire, Commitment]> ----------------------------------- Patient Action: Figure A5: Patient Action Selection. You are tasked with updating the pa...

  8. [10]

    Be conservative: require at least two strong positive signals to reach +0.05. --------------------------------------- Output Format (plain text only): Delta Rapport: <-0.20–0.05> ################################################## Delta Rapport: Figure A2: Prompt for Patient Rapport Delta Update. You are reflecting on your own readiness for change during t...

Show all 36 references
  1. [11]

    If turn_counter < {}, you MUST not return "Plan"

  2. [12]

    GOOD/VERY GOOD→more likely Acknowledge, Engage, Inform, Desire, Commitment; BAD/VERY BAD→more likely Hesitate, Deny, Reject

    Attend closely to patient_quality_last_therapist_turn. GOOD/VERY GOOD→more likely Acknowledge, Engage, Inform, Desire, Commitment; BAD/VERY BAD→more likely Hesitate, Deny, Reject

  3. [13]

    Default to Acknowledge or Engage unless another action is clearly more appropriate

  4. [14]

    Anchor decisions to the most recent therapist and client turns

  5. [15]

    Ensure consistency with all inputs (background, cognitive model, etc.); avoid contradictions

  6. [16]

    not sure

    Distinguish close pairs: - Hesitate vs Doubt: ambivalence (“not sure”, “maybe later”) vs feasibility skepticism (“won't work”). - Deny vs Downplay: non-problem stance (“not a problem”) vs minimized impact (“not that bad”). - Accept vs Plan: agreement + steps→Plan. - Reject vs ...

  7. [20]

    - Negative: guardedness, resistance, rupture, deflection

    Extract rapport signals: - Positive: openness, self-disclosure, engagement, collaboration, repair, appreciation. - Negative: guardedness, resistance, rupture, deflection

  8. [21]

    - Communication style (indirectness, disclosure norms)

    Weigh contextual factors from patient_background: - Authority preferences (skepticism lowers or caps gains). - Communication style (indirectness, disclosure norms). - Support network and relational history (isolation may inflate trust; trauma dampens it). - Cultural considerat...

  9. [22]

    - Coping style (avoidance may mimic engagement without rapport)

    Weigh inferred patient_cognitive_model: - Trust disposition (skeptical vs ready to trust). - Coping style (avoidance may mimic engagement without rapport). - Emotional volatility (amplifies negative deltas). - Stage of change (precontemplation limits gains; preparation allows ...

  10. [23]

    - Negative signals weigh more than positive

    Assign a rapport delta (not absolute value) using: - Range: –0.10 to +0.05 per therapist turn. - Negative signals weigh more than positive. - If positives and negatives coexist, negatives dominate unless explicit repair occurs. - If evidence is ambiguous or minimal, output 0.0...

  11. [24]

    Be conservative: require multiple strong positive signals to reach +0.05. --------------------------------------- Output Format (plain text only): Delta Rapport: <-0.20–0.05> ############################################### Delta Rapport: Figure A10: Prompt for Therapist-Side R...

  12. [25]

    - Distinguish change talk vs sustain talk

    Interpret what matters to the client right now: - Identify reinforced or emerging motivations, values, or barriers. - Distinguish change talk vs sustain talk. - Consider readiness cues, affect, and rapport dynamics

  13. [26]

    Clarify the current session focus: - Is the client exploring ambivalence, selecting a target, or considering a small next step?

  14. [27]

    Planning

    Write one or more sentences from the therapist's perspective that: - Describe what the client appears to want to address now, linked to broader motivation. - Are session-scoped, concrete, and feasible. - Use tentative language if information is incomplete. - Directly connect t...

  15. [28]

    When you think about the life you want, what makes the effort worth it?

    EVOKING VALUES AND STRENGTHS - Reconnect values and recall past successes to build motivation. Example: “When you think about the life you want, what makes the effort worth it?”

  16. [29]

    Many people feel stuck with something this hard-it shows how much you care

    NORMALIZE AND REFRAME - Reduce shame by normalizing struggle and reframing toward possibility. Example: “Many people feel stuck with something this hard-it shows how much you care.”

  17. [30]

    Would it be okay if we tried looking at this from another angle together?

    ACKNOWLEDGE AND CHANGE COURSE - Name stuckness and collaboratively shift direction. Example: “Would it be okay if we tried looking at this from another angle together?”

  18. [31]

    Of these pieces, what feels most important to focus on right now?

    STRATEGIC SUMMARY AND REFOCUS - Concise recap highlighting themes, then invite priority focus. Example: “Of these pieces, what feels most important to focus on right now?”

  19. [32]

    Finding one loose strand can sometimes create space for change

    SHIFT THE LENS WITH METAPHORS - Use a gentle metaphor to reframe and open new perspective. Example: “Finding one loose strand can sometimes create space for change.” --------------------------------------- Output Format (plain text only): Pivoting Strategy: <one of the options...

  20. [33]

    When you think about the life you want, what makes the effort worth it?

    EVOKING VALUES AND STRENGTHS - Reconnect values and past successes. Example: "When you think about the life you want, what makes the effort worth it?"

  21. [34]

    Many people feel stuck-it shows how much you care, not weakness

    NORMALIZE AND REFRAME - Reduce shame; reframe toward possibility. Example: "Many people feel stuck-it shows how much you care, not weakness."

  22. [35]

    Would it be okay if we tried another angle together?

    ACKNOWLEDGE AND CHANGE COURSE - Name stuckness; ask permission to shift. Example: "Would it be okay if we tried another angle together?"

  23. [36]

    Of these, what feels most important right now?

    STRATEGIC SUMMARY AND REFOCUS - Concise recap; invite priority focus. Example: "Of these, what feels most important right now?"

  24. [37]

    What could be one loose strand to pull first?

    SHIFT THE LENS WITH METAPHORS - Gentle metaphor; check resonance. Example: "What could be one loose strand to pull first?" --------------------------------------- Reference Dialogue (Natural Therapist Speech Example) therapist: Yeah. And, get-getting to the polytech and gettin...

  25. [2007]

    Xuehai He, Shu Chen, Zeqian Ju, Xiangyu Dong, Hongchao Fang, Sicheng Wang, Yue Yang, Jiaqi Zeng, Ruisi Zhang, Ruoyu Zhang, and 1 others

    Change is not always linear: The study of non- linear and discontinuous patterns of change in psy- chotherapy.Clinical psychology review, 27(6):715– 723. Xuehai He, Shu Chen, Zeqian Ju, Xiangyu Dong, Hongchao Fang, Sicheng Wang, Yue Yang, Jiaqi Zeng, Ruisi Zhang, Ruoyu Zhang, ...

  26. [2020]

    Jeffrey Ip and Kritin V ongthongsri

    Meddialog: Two large-scale medical dialogue datasets.arXiv preprint arXiv:2004.03329. Jeffrey Ip and Kritin V ongthongsri. 2025. deepeval. Matthew Jörke, Shardul Sapkota, Lyndsea Warkenthien, Niklas Vainio, Paul Schmiedmayer, Emma Brunskill, and James A Landay. 2025. Gptcoach:...

  27. [2024]

    ai psy- chosis

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437. Ryan Louie, Ananjan Nandi, William Fang, Cheng Chang, Emma Brunskill, and Diyi Yang. 2024. Roleplay-doh: Enabling domain-experts to create llm-simulated patients via eliciting and adhering to principles.arXiv prepri...

  28. [2025]

    Jonathan P Chang, Caleb Chiam, Liye Fu, An- drew Wang, Justine Zhang, and Cristian Danescu- Niculescu-Mizil

    Reasoning is not all you need: Examining llms for multi-turn mental health conversations.arXiv preprint arXiv:2505.20201. Jonathan P Chang, Caleb Chiam, Liye Fu, An- drew Wang, Justine Zhang, and Cristian Danescu- Niculescu-Mizil. 2020. Convokit: A toolkit for the analysis of ...

Pith tools

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