acceptodds
Withdrawn from ICLR 2027

UNFOLD: Patient-State Prediction and Progressive Profile Disclosure for Long-Horizon Patient Simulation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-horizon patient simulation requires more than profile consistency: a simulated patient may remain factually faithful while exhibiting implausible behavioral changes or revealing personal information prematurely, repeatedly, or excessively. We frame these trajectory-level failures as two coupled temporal controls: patient-state prediction, which models how attitude, emotion, and topic evolve with interaction, and progressive profile disclosure, which controls when patient-specific information becomes available. We introduce UNFOLD, a psychotherapy patient simulator that predicts the next behavioral state from structured interaction history and controls access to independently addressable profile evidence before response generation. On 8,397 held-out patient turns, UNFOLD improves attitude, emotion, and topic prediction over a supervised text-encoder reference. Across 30 matched long-horizon simulations, it produces broader, more gradual, and less repetitive information trajectories than full-profile simulators, with differences persisting under longer baseline horizons. Mechanism-targeted ablations isolate the roles of state control, progressive selection, and the disclosure bottleneck, while blinded therapist raters prefer UNFOLD overall in pairwise evaluation. Together, these results suggest that long-horizon simulation should treat both behavioral state and information availability as explicit time-varying controls, a principle that may extend to interactive systems combining evolving behavior with persistent information.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Patient Simulation, Large Language Models, Long-Horizon Dialogue, Behavioral Dynamics, Progressive Disclosure

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_wXP4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces UNFOLD, a psychotherapy long-horizon patient simulator that explicitly separates how patient state and information availability evolve over time. UNFOLD uses task-specific prediction heads for patient state and offline profile construction for information availability. On 30 long-horizon simulations, UNFOLD outperforms baselines.

Critical strengths

The broad motivation of modeling behavioral state and information availability for long-horizon patient simulation makes sense.

Critical weaknesses

The LLMs used in this paper are outdated. (GPT-4o-mini, Gemini-2.5-Pro, Qwen2.5-72B-Instruct). SOTA LLMs with agentic capabilities (GPT-6 Astra, Claude Opus 5.5, Qwen3.8) should be evaluated. It's very possible that newer models with agentic capabilities can outperform UNFOLD with basic scaffolding or prompts, without the machinery that UNFOLD introduces like task-specific prediction heads (Figure 2) or filtering/extraction/merging (Figure 3).

Questions for authors
  1. Can you use newer LLMs?
  2. Why is it necessary to split out task-specific heads (Figure 2) and do a bespoke pipeline for offline profile construction (Figure 3)? Does it perform better than putting these instructions/guidelines in the prompt?
  3. Why are there only 30 matched long-horizon simulations?
Reviewer_zJtNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

UNFOLD controls both how a simulated patient responds and which profile details are available at each turn. It combines a structured attitude/emotion/topic predictor with a history-aware disclosure policy. The authors train on annotated psychotherapy dialogues and evaluate next-state prediction on 8,397 held-out turns, long simulations on 30 profiles, and blinded comparisons by three therapist raters. The system produces broader, later, and less repetitive disclosure than full-profile baselines. The paper also includes 200-turn baseline runs and component ablations.

Critical strengths
  1. The paper identifies a useful distinction: a patient can stay consistent with a profile while revealing it in an implausible way. Making disclosure an explicit control addresses this problem directly. The ablations support the design: removing selection lowers breadth from .558 to .307, and removing the bottleneck raises repetition from .528 to .723 (Table 3).

  2. Several evaluation choices strengthen the results. The differences persist when full-profile baselines are allowed 200 turns (Table 15), so they cannot be explained simply by the original 60-turn cap. The profile matcher has a human audit with 93.3% F1 (Appendix D.4). The therapist study also provides evidence beyond the automatic metrics: 61 wins, 18 ties, and 5 losses across 84 dimension-specific comparisons, with Fleiss' kappa .504 (Tables 17–18).

Critical weaknesses
  1. My main concern is how the predictor inputs were annotated. Appendix B.6 says that annotation uses two following turns as well as the current and preceding turns (lines 895–899). The predictor receives historical patient labels and the current therapist's labels. If those features come from the same annotation pass, they may indirectly contain information from the patient response being predicted. Giving the predictor only earlier labels would not remove that information. This may already be handled in the implementation, but the paper does not make it clear. The distinction between how target labels and input features are constructed is essential for interpreting Table 1.
  2. The realism evidence is less decisive than the disclosure metrics. The selector directly encourages breadth, delayed disclosure, and less repetition, but more of each is not always more realistic. Against GPT-4o-mini, which is also UNFOLD's generator, the therapist study has 9 wins, 18 ties, and 1 loss; response dynamics has only 1 win and 6 ties (Table 17). The long-horizon assessment uses six short checkpoints rather than full sessions. The baseline comparison also changes the predictor and stopping rule alongside profile access. The 200-turn control helps, but a matched retrieval baseline with the same predictor and stopping rule would make the contribution more convincing.
Questions for authors
  1. For Table 1, were all input-side historical labels and current therapist labels re-annotated without access to the target patient response or later turns? Please distinguish target-label construction from feature construction and describe the online annotation procedure. If bidirectional labels were reused as features, what happens under a strictly causal reconstruction?
  2. How much of the next-state advantage remains against a simple MLP or logistic model given exactly the same structured features? Since the priors and projected summaries in Equations 6–7 are instance-independent, such a control would clarify the benefit of the predictor architecture versus its representation of the input.
  3. Can you compare disclosure policies using the same state predictor, generator, context window, and common stopping rule, including a simple relevance-plus-novelty retrieval policy? Do the main results persist over a common fixed prefix?
  4. Can the therapist comparison report uncertainty for the same-backbone GPT-4o-mini contrast, and assess complete sessions or longer continuous segments? This would materially strengthen the connection between improved trajectory metrics and perceived realism.
Reviewer_88NPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces UNFOLD, a framework for long-horizon psychotherapy patient simulation that explicitly models patient-state evolution and progressive profile disclosure. It predicts the next patient attitude, emotion, and clinical topic from structured interaction history and therapist actions, while a history-aware selection policy controls which patient-specific evidence is made available to the response generator. The authors evaluate state prediction on 8,397 held-out patient turns, compare information trajectories across 30 matched simulations, conduct component ablations, and report blinded pairwise evaluations by therapist raters. The results suggest improvements in state prediction over a supervised text-encoder reference and broader, more gradual, and less repetitive information trajectories than the evaluated full-profile baselines.

Critical strengths

1.Explicit decomposition of behavioral dynamics and information availability. UNFOLD separates the prediction of patient behavior from the selection of patient-specific evidence. This decomposition offers a potentially useful way to address two distinct problems in long-horizon simulation: implausible behavioral changes and poorly timed or repetitive information disclosure.

2.Multi-faceted experimental evaluation. The paper evaluates next-state prediction on held-out dialogue turns, measures information trajectories in matched simulations, performs mechanism-targeted ablations, and includes blinded therapist assessments. These experiments provide useful evidence about the behavior of the proposed pipeline, although stronger comparisons are needed to establish its incremental contribution.

Critical weaknesses

1.Insufficient differentiation from prior work on dynamic patient-state modeling. The paper motivates UNFOLD partly through the need to model how patient behavior evolves in response to interaction. However, its relationship to existing work is not sufficiently developed. PUPPET (Xu et al., ACL 2026) studies intervention-conditioned psychological state transitions through a neural-symbolic architecture, while AnnaAgent (Wang et al., Findings of ACL 2025) addresses dynamic emotional and cognitive states and multi-session memory in counseling simulation. These works are relevant to the motivation and technical positioning of UNFOLD, although thematic overlap alone does not establish methodological redundancy. The manuscript should explain more precisely how UNFOLD's state representation, prediction mechanism, conditioning signals, and evaluation setting differ from these approaches. In the current version, the lack of this analysis makes it difficult to determine the extent of the incremental contribution of the state-prediction component.

2.The distinct contribution of progressive profile disclosure is not sufficiently established. Progressive disclosure is the second central component of UNFOLD. Patients With Personality (Schlager et al., 2026, arXiv:2606.17441) is relevant because it studies patient simulation with controlled personality and selective information disclosure. Its relevance warrants consideration when positioning UNFOLD's disclosure mechanism, but the absence of a citation to an arXiv preprint should not, by itself, constitute a reason for rejection. The substantive issue is whether UNFOLD's history-aware evidence-selection policy, and its coupling with predicted patient states, provides a meaningful methodological advance over existing approaches. The current experiments primarily compare UNFOLD with full-profile prompting baselines and do not directly resolve this question. The paper therefore lacks sufficient evidence to establish the distinct contribution of progressive disclosure and its integration with state prediction.

Questions for authors

1.How does UNFOLD differ from PUPPET and AnnaAgent in modeling patient-state transitions, particularly in state representation and conditioning on therapist actions and dialogue history? 2.What is new about UNFOLD’s progressive disclosure mechanism compared with Patients With Personality? Is the main novelty the selection policy itself or its integration with patient-state prediction? 3.Can the authors provide a controlled comparison to show that combining state prediction with progressive disclosure offers benefits beyond either component alone or existing approaches? If the authors address these questions in a revision and clearly establish the technical differences from these related works, I would be open to increasing my score.

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

Related papers

Discussion (0)

Sign in to comment.