{"id":"986d9ed6-c220-4dcd-93e6-020910d0d680","arxiv_id":"2412.15242","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Script-Based Dialog Policy Planning lets an LLM therapist move through explicit script states, and 100 simulated dialogs show it is feasible with trade-offs between efficiency and script adherence.","lead":"This paper proposes a way to make LLM-powered therapy chatbots follow a therapist-written script with visible, step-by-step states instead of freewheeling conversation. The authors tested two versions in 100 simulated therapy sessions and showed the approach is workable, with trade-offs between efficiency and sticking to the script.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated LLM validator sits under every effectiveness number: low correct section completion rates (45.5%/35.7%) are explained away without human ground truth, so the central deterministic constraint claim is not yet evidenced.","rationale":"I read the paper in good faith. The architecture is clearly described, the code is released, and examples show plausibly natural conversations; those are real assets. My concern is not that the idea is impossible, but that the specific quantitative evidence used to support 'deterministic constraint' is self-referential: an LLM patient, an LLM therapist, and an LLM validator, with no human ground truth at any point. The paper itself acknowledges the tension when it refuses to draw a firm conclusion from its own completion metric. Thus the central claim is conditional: the architecture may be feasible, but whether it reliably constrains behavior remains untested. This matches the reader's CONDITIONAL verdict, so I do not propose a different verdict. The concrete test of human annotation would convert the conditional into either a supported feasibility claim or a clear negative result.","tokens_in":10370,"tokens_out":4570,"duration_ms":47438,"concrete_test":"Use the released GitHub repository and dataset to sample about 50 section-transition events across the 100 dialogs. Have two independent annotators with clinical/psychology training judge 'section completed' and 'switch coherent' using the same criteria given to the Validator LLM in the technical appendix. Compute Cohen's kappa between annotators and between each annotator and the Validator LLM. Then recompute the two effectiveness rows of Table 1 from human labels. If human completion rates are near the validator's 45.5%/35.7%, the deterministic-constraint claim fails; if they are above about 80%, the validator is biased and the paper's dismissal is justified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the script acts as a deterministic component, constraining the LLM's behavior in desirable ways. The only quantitative evidence for this constraint is the 'Percentage of Correct Section Completions' and 'Percentage of Coherent Section Switches' in Table 1, both produced by a gpt-4o Validator LLM with no human gold standard, inter-rater agreement check, or calibration. The completion rates are strikingly low: 45.5% (variant A) and 35.7% (variant B). The authors do not take these numbers at face value; instead, in Section 4 they reinterpret them, stating that non-correct completions mostly reflect the Dialog LLM 'following its instructions slightly less accurately for the sake of addressing a user utterance and/or holding a fluent conversation,' and conclude that they 'can't clearly conclude that variant A is completing instructions more correctly.' This reinterpretation is the load-bearing step: if the validator is merely strict, then the architecture may indeed constrain well and the low scores are artifacts; if the validator is accurate, then the script fails to constrain the agent on most transitions. Without human annotation of a sample of the same dialogs, the feasibility claim rests on the authors' own qualitative impression of transcripts.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Script-Based Dialog Policy Planning, a hybrid architecture for LLM-powered conversational agents in behavioral health. The core idea is to inject an expert-written, state-structured script into the LLM prompt so that the agent moves through a finite set of therapeutic states, with an explicit assessment of section completion and subsequent state transition. Two implementation variants are compared: a single-LLM variant (A) using an adapted ProCoT scheme, and a multi-LLM variant (B) using separate Assessor, Dispatcher, and Dialog LLM actors. The authors generate 100 synthetic conversations with LLM-simulated patients and evaluate efficiency (per-turn duration, token usage) and effectiveness (validator-assessed correct section completions, coherent section switches, and script-conform dispatchings). Results indicate that variant A is more efficient and variant B is more script-adherent, while overall completion rates are low. The paper concludes that the architecture is generally feasible and provides a foundation for future AI Therapist development.","tokens_in":10616,"tokens_out":3593,"duration_ms":38917,"significance":"If the feasibility claim is upheld, the architecture addresses a real gap in LLM-based conversational agents for clinical contexts: integrating expert-defined rules, evidence-based practice adherence, and inspectable state transitions while retaining conversational fluency. The paper provides a concrete, reproducible implementation with open-source code and transparent reporting of metrics. The two-variant comparison offers practical guidance on the efficiency/effectiveness trade-off between single-actor and multi-actor designs. However, the significance is tempered by the unvalidated LLM-based evaluation: all effectiveness numbers depend on a single gpt-4o validator with no human ground truth, and the central 'deterministic constraint' claim is not yet quantitatively supported. As a technical demonstration, the work is valuable; as an evidence base for an 'AI Therapist,' it is preliminary and requires substantial validation.","major_comments":[{"comment":"All effectiveness metrics—'Percentage of Correct Section Completions,' 'Percentage of Coherent Section Switches,' and the validator's reasons—are produced by a single gpt-4o Validator LLM with no human-labeled gold standard, no inter-annotator agreement check, and no calibration against real therapy transcripts. The low completion rates (45.5% for A, 35.7% for B) are reinterpreted in Section 4 as artifacts of a strict validator that penalizes fluency-preserving deviations. This reinterpretation is load-bearing: if the validator is merely strict, the architecture may constrain well; if the validator is accurate, the script fails to constrain on most transitions. Without human annotation of a sample of the same dialogs (e.g., 20% of sections, with Cohen's kappa between validator and human raters), the comparative effectiveness claims and the 'deterministic constraint' claim are not supported.","section":"§3.4, §3.5, Table 1"},{"comment":"The comparison between variants A and B, and the efficiency/effectiveness conclusions, rest on point estimates without any statistical significance tests, confidence intervals, or error bars. With only 50 dialogs per variant and the observed differences (e.g., 45.5% vs. 35.7% correct completions; 2,800 ms vs. 3,520 ms average duration), the authors cannot rule out variance as the source of the differences. Additionally, there is no baseline condition, such as an unscripted proactive agent or a directly enforced finite-state machine, to demonstrate that the script actually constrains behavior beyond what an unconstrained LLM would do. The paper should either add a baseline comparison or substantially weaken the claim that the script is the cause of the observed behavior.","section":"§3.6, §4, Table 1"},{"comment":"The term 'deterministic component' is used in the abstract and Section 2.3 to describe the script, but the implementation does not enforce deterministic behavior: the LLM's adherence to the script is probabilistic and depends on prompt interpretation, as evidenced by the 13.2% (variant A) and 5.8% (variant B) non-script-conform dispatchings in Table 1. Temperature=0 does not guarantee determinism. The claim that the script 'constrains the LLM's behavior in desirable ways' should be reframed as a probabilistic, prompt-dependent constraint, or the architecture should include a separate enforcement mechanism (e.g., a rule-based checker that rejects non-conforming transitions). As written, the central claim overstates the determinism of the proposed system.","section":"Abstract, §2.3"},{"comment":"The use of LLM-simulated patients, without any human users, limits the external validity of the feasibility claim. The paper acknowledges this limitation in the Conclusion, but the feasibility conclusion in Section 4 ('we conclude that our proposed architecture is generally feasible') is based solely on synthetic interactions. The patient model is not validated against real patient behavior, and the paper provides no evidence that the simulated patients' utterances—including 'challenge the therapist's advice' behaviors—are representative of actual therapy conversations. A more cautious conclusion, such as 'feasible in simulated settings,' would be more accurate at this stage.","section":"§3.3, §4"}],"minor_comments":[{"comment":"There is a typo: 'acccess' should be 'access' in the description of variant B.","section":"§3.2"},{"comment":"The claim 'We cannot identify any obvious deficiencies with respect to our outlined requirements' is vague; the paper should specify the criteria used in the qualitative review of transcripts, or provide a more detailed breakdown of what was checked.","section":"§4"},{"comment":"The paper states that 'a JSON-formatted script might be easier to interpret for the software, but is not necessary as LLMs can be asked to interpret unformatted text.' The chosen script format should be shown explicitly, since the appendix reference alone is not sufficient for the reader to assess the actual input used.","section":"§2.3"},{"comment":"The table presents averages without variance or distribution information; reporting standard deviations or per-dialog ranges would be informative, even if significance testing is deferred.","section":"§4, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper discloses a conflict of interest (authors affiliated with Aury, which develops an AI-based mental health chatbot). This is relevant for the editor when considering the framing of the feasibility claim, though the disclosure itself is adequate. The main technical concern is the unvalidated LLM validator and the lack of a baseline; these are fixable with additional analysis and more cautious wording. The paper is within scope for a CS/CL venue but would benefit from a more rigorous evaluation before being presented as a foundation for clinical deployment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a feasibility proposal, not an efficacy study, and if you read it that way it mostly works. The new bit is combining a natural-language script with section-level instructions and a completion assessment step into a finite-state dialog policy, while letting the LLM keep conversational freedom inside each section. They compare a single-LLM variant (ProCoT-like) with a multi-actor variant (Ask-an-Expert-like). The architecture is described carefully, the requirements section is thoughtful, and they ship the code, prompts, and all 100 synthetic dialogs on GitHub. Credit where due: this is reproducible, honest engineering.\n\nThe soft spot is exactly where the stress-test note points. Table 1's effectiveness numbers come from a gpt-4o validator labeling section completions and coherence, with no human gold standard, no inter-rater check, no calibration. The correct-completion rates are low (45.5% and 35.7%), and the authors reinterpret them as the dialog LLM following instructions slightly less accurately for the sake of fluency. That reinterpretation may be right—reading the examples, it looks plausible—but it is unsupported without a human-labeled sample of the same dialogs. So the deterministic-constraint claim is not yet evidenced. It is also not disproven; the transcripts suggest the architecture does keep conversations on-script in a loose sense.\n\nOther, smaller issues: no baseline (e.g., plain ProCoT without the script), no error bars or significance tests, and patients are LLM-simulated. The authors admit all of this in Section 5, which is good, but it means the comparative claims are preliminary.\n\nWho is this for? Anyone building practical LLM therapy agents or hybrid dialog management systems. The paper is useful for its architecture and for naming the evaluation problem, not for its numbers. It deserves a serious referee, but the revision path needs human validation of the validator, at least one baseline, and statistical reporting. If I were the editor I would send it out with a major-revision expectation rather than desk-reject it.","headline":"A clear, honest feasibility architecture for script-constrained LLM therapy agents; the effectiveness numbers rest on an unvalidated LLM judge.","tokens_in":11118,"tokens_out":1817,"would_cite":true,"duration_ms":19376,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Script-Based Dialog Policy Planning is a feasible architecture for an AI Therapist: an expert-written script injected into prompts moves the LLM through a finite set of therapeutic states while retaining fluency.","keywords":["script-based dialog policy planning","AI therapist","LLM conversational agents","dialog policy","prompting","finite-state script","mental health chatbot","inspectability"],"falsifier":"Have a small group of licensed psychotherapists rate a sample of the 100 generated dialogs on the same three effectiveness criteria and measure agreement with the Validator LLM; if the agreement is near chance, the reported effectiveness comparisons between the two variants are unsupported.","tokens_in":10174,"feed_emoji":"🧠","tokens_out":8235,"duration_ms":73103,"temperature":0.7,"pith_summary":"The paper introduces a hybrid design for therapy chatbots: a deterministic, expert-written script defines a finite set of conversation states and the rules for moving between them, and the script is injected into the LLM's prompt so that the model acts as the policy executor. On each turn the agent assesses whether the current section is complete, optionally selects the next section, and generates a response; this makes the decision path inspectable, a requirement the paper argues is essential for risk management and clinical evaluation. Two variants are implemented, one using a single LLM for all steps and one splitting the steps across Assessor, Dispatcher, and Dialog LLMs, and both are tested on 100 conversations with LLM-simulated patients. The paper concludes from these tests that the architecture is generally feasible, and reports that the single-LLM variant is more efficient and more coherent at section switches while the multi-LLM variant conforms to the script's transition rules more strictly.","feed_headline":"A script can steer LLM therapists through fixed states","feed_subtitle":"In 100 simulated patient sessions, a scripted LLM follows a therapeutic plan while keeping transitions visible.","key_machinery":"The Script, a natural-language text defining a finite set of states ('sections'), each with a series of tasks and explicit transition rules such as '[→ 4]' (proceed to Section 4). The script is inspired by rule-based finite-state dialog management and by psychotherapeutic treatment manuals, and it is provided to the LLM via prompting rather than as a separate program. The per-turn mechanism, Script-Based Dialog Policy Planning, consists of four steps: assess whether the current section is completed; optionally reason and plan; optionally decide on the next section; generate the response. The script is what makes the agent's state explicit and its transitions inspectable, and the two implementation variants differ only in whether these steps are performed by one LLM or split across three.","core_discovery":"The central discovery is that a script can act as a 'deterministic component' that constrains an otherwise fluent LLM to follow a therapeutic plan, producing explicit, inspectable state transitions. In the experiments, both variants moved reliably through the script's eight sections (introduction, engagement, exploration, selection, three exercises, ending) in natural-sounding dialogues, and the quantitative comparison matched the paper's stated hypothesis: the single-LLM variant was faster and cheaper per turn by about 20% in duration and 23% in input tokens, while the multi-LLM variant achieved higher script-conform dispatchings (94.2% vs 86.8%) but lower coherent section switches (87.6% vs 96.9%). The paper interprets the low correct-section-completion scores (45.5% and 35.7%) not as clear failures of instruction-following but as cases where the model prioritized fluency over literal completion, and it identifies the tradeoff between following the patient and following the script as a question for future work.","pith_inferences":["If the validator's judgments survive a human-rating check, the same architecture could formalize any manualized psychotherapy, since a script is essentially a machine-readable treatment manual.","The coherence drop in the multi-LLM variant suggests that injecting a new section at the moment of a switch disrupts continuity; a testable extension would be to have the Dispatcher phrase the new section as a continuation of the previous one.","A natural next experiment is a Wizard-of-Oz study with human users to test whether the script's states remain alignable to real patient behavior, especially unexpected utterances.","The pattern of choosing Section 8 (ending) when the simulated patient rejected all offered exercises suggests the script needs explicit rules for patient refusal; the paper does not propose such rules."],"forward_implications":["An expert-written script can be authored and iterated in plain language by domain experts with no programming background, giving therapists direct control over the agent's behavior.","Prompt-based script injection yields an inspectable trace of state transitions, which the paper argues is necessary for risk management and clinical evaluation of an AI Therapist.","The single-LLM variant is more efficient and produces more coherent section switches, while the multi-LLM variant is more script-adherent; the choice between them depends on whether fluency or rule-following is prioritized.","Because instructions are given at the section level, the agent can take multiple turns to complete a section and can temporarily deviate to follow user utterances, which the paper argues preserves natural conversation.","The low correct-section-completion scores show that literal adherence to the script is imperfect, and the tradeoff between following the patient and following the script is left as an open question for future work."],"supporting_citations":[{"why":"Defines the responsible-development requirements for clinical LLMs that motivate requirements (3)-(5) of the AI Therapist architecture.","marker":"(Stade et al. 2024)"},{"why":"Supplies the dialog-management background, including finite-state machines and hybrid rule-based/probabilistic approaches that inspire the script.","marker":"(Harms et al. 2019)"},{"why":"Proposes the ProCoT proactive prompting scheme that Variant A adapts.","marker":"(Deng et al. 2023b)"},{"why":"Proposes the Ask-an-Expert prompting scheme that Variant B adapts.","marker":"(Zhang, Naradowsky, and Miyao 2023)"},{"why":"Describes the psychologist-developed chatbot intervention from which the exemplary script's flow and phrasing are derived.","marker":"(Langhammer et al. 2021)"},{"why":"Provides the five patient case examples used to seed the LLM-simulated patients.","marker":"(American Psychological Association 2008)"},{"why":"Surveys proactive dialogue systems and defines proactivity and evaluation context for the agent requirements.","marker":"(Deng et al. 2023a)"}],"fun_headline_variants":["Scripted LLM therapist: fixed states, inspectable paths","LLM therapy gets a script: 100 sessions test the plan","AI therapist follows a script, state by state","Scripting LLM therapists: 94% script adherence, at a cost","Can a script keep LLM therapy on track?"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Validator LLM's judgments about correct section completions and coherent section switches are trustworthy, and those judgments were never checked against human raters or real therapy transcripts.","fun_headline_variants_meta":{"raw":{"variants":["Scripted LLM therapist: fixed states, inspectable paths","LLM therapy gets a script: 100 sessions test the plan","AI therapist follows a script, state by state","Scripting LLM therapists: 94% script adherence, at a cost","Can a script keep LLM therapy on track?"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000133,"raw_usage":{"total_tokens":1150,"prompt_tokens":971,"completion_tokens":179,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":95}},"tokens_in":587,"tokens_out":179,"duration_ms":2543,"temperature":1.0,"reasoning_tokens":95,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:22:21.385523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have a small group of licensed psychotherapists rate a sample of the 100 generated dialogs on the same three effectiveness criteria and measure agreement with the Validator LLM; if the agreement is near chance, the reported effectiveness comparisons between the two variants are unsupported.","supporting_citations":[{"cited_title":"C.; Stirman, S","cited_arxiv_id":null,"evidence_quote":"Defines the responsible-development requirements for clinical LLMs that motivate requirements (3)-(5) of the AI Therapist architecture."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the dialog-management background, including finite-state machines and hybrid rule-based/probabilistic approaches that inspire the script."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proposes the Ask-an-Expert prompting scheme that Variant B adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the psychologist-developed chatbot intervention from which the exemplary script's flow and phrasing are derived."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the five patient case examples used to seed the LLM-simulated patients."}],"review_version":1}