acceptodds
Under review as a conference paper at ICLR 2027

Beyond Compliance: resistance-informed motivation alignment for Challenging Psychological Client Simulation

Abstract

Large language models (LLMs) are increasingly used to simulate clients for scalable counselor training and evaluation. However, existing simulators often inherit the compliance bias of instruction-tuned LLMs, producing clients who are unrealistically open, receptive, and emotionally stable. We introduce ResistClient, a resistance-informed client simulator that models both how client resistance is expressed and why it emerges. Drawing on Client Resistance Theory, we construct the Resistance-informed Psychological Conversation (RPC) dataset by augmenting real counseling dialogues with diverse, context-dependent resistance patterns. Fine-tuning on RPC enables ResistClient to generate challenging behaviors beyond generic compliance. To ground these behaviors in coherent psychological processes, ResistClient represents client profiles using the clinical 5P case formulation, explicitly infers the client’s current motivation from the profile and dialogue context, and generates responses conditioned on this motivation. Expert-informed process rewards further align motivational reasoning and response generation with human preferences while promoting decision–response consistency. Extensive automatic and expert evaluations show that ResistClient substantially outperforms existing simulators in challenge fidelity, behavioral plausibility, and reasoning coherence. Moreover, evaluations with ResistClient reveal substantial limitations in existing psychological LLMs, including frequent client resistance, therapeutic drift, and limited progress. These findings demonstrate that resistance-aware simulation provides a more realistic and demanding environment for evaluating and improving psychological LLMs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Client Simulation, Psychological Conversation, Motivation Reasoning, Reinforcement Learning, Mental Health

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aTzwOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes ResistClient, a resistance-aware client simulator for psychological counseling. The paper argues that existing LLM-based client simulators are usually too cooperative. They often accept counselor suggestions too easily and therefore cannot provide difficult and realistic counseling situations.

Critical strengths

The problem is meaningful. Existing client simulators can be too cooperative, which may make counseling evaluation too easy. Modeling realistic resistance is useful for stress testing psychological LLMs.

The paper uses psychological theory instead of only prompt engineering. The combination of the 5P formulation and resistance categories gives a clearer reason for when and why resistance happens.

Critical weaknesses
  1. The evaluation is highly dependent on the authors' own synthetic RPC framework. This is my main concern. RPC is constructed by using an LLM to rewrite existing counseling conversations according to the proposed resistance taxonomy and 5P profiles. ResistClient is then trained on this dataset and evaluated mainly on held-out RPC profiles using the same resistance definitions.

  2. The comparison with large LLM baselines is not fully fair. GPT-5.1, DeepSeek-V3.2, Kimi and other models mainly use prompting, while ResistClient receives task-specific SFT on RPC and additional expert-supervised GRPO training.

Questions for authors

Please see above.

Reviewer_fZ7POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents ResistClient, a counseling client simulator that generates context-dependent resistance. It constructs RPC from counseling dialogues using 5P profiles and a resistance taxonomy, then trains Qwen3-8B with SFT and offline GRPO using expert process and consistency rewards. Automatic and counselor evaluations report gains over prompting and SFT baselines; the simulator is also used to stress-test counselor LLMs.

Critical strengths
  1. RPC ties resistance to client profiles and counselor interventions, with masked expert reannotation and corrections to compliant-resistance labeling. It provides a concrete training resource covering both resistance and cooperation.
  2. On the same Qwen3-8B backbone and prompts, reported resistance F1 rises from 41.68 to 68.33 after SFT and 74.42 after alignment, alongside higher expert ratings. This supports the usefulness of the full training pipeline.
Critical weaknesses
  1. The ablation does not isolate process credit or the consistency reward. Moving from SFT to ResistClient adds expert-scored trajectories and both reward mechanisms together (Table 1; Section 3.3). Without a control using the same scored trajectories and training budget, the gain cannot be attributed to the specific reward design.
  2. Coh. does not establish motivational coherence: repeating the same reply maximizes the score (Appendix C.2.2). The human A/B comparison also gives no source or matching procedure for the human responses (Table 2). Expert realism ratings and the case studies support plausibility, but the stronger claims about motivational dynamics and quality comparable to human responses remain insufficiently supported.
Questions for authors
  1. Using the same expert-scored trajectories and training budget, what are the results for (i) assigning all rewards at sequence end and (ii) removing the consistency reward?
  2. What human responses were used in Table 2's A/B test, and how were contexts matched? Please report paired win/tie/loss counts and uncertainty, and clarify whether reasoning traces were hidden from realism raters.
Reviewer_4KtMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ResistClient, a resistance-aware client simulator for psychological counseling. It combines a theory-guided dataset, 5P case formulations, supervised fine-tuning, and expert-informed process rewards to generate contextually appropriate resistance. Experiments show improved realism and resistance fidelity and demonstrate its use for stress-testing psychological LLMs.

Critical strengths

The paper addresses an important limitation of existing client simulators: their tendency to be overly compliant and therefore insufficiently challenging for counselor training and evaluation. ResistClient is well motivated by psychological theory and combines 5P case formulations, context-dependent resistance modeling, supervised fine-tuning, and expert-informed process rewards in a coherent framework. The experiments are extensive and show consistent improvements over both prompt-only LLMs and existing client simulators.

Critical weaknesses

The main concern is that training and evaluation rely heavily on the same synthetically constructed RPC framework, so the results may primarily demonstrate alignment with the authors’ resistance taxonomy rather than fidelity to real clients. The human comparison is limited, and the paper does not systematically compare against optimized prompting pipelines. Moreover, modeling each response with a single mutually exclusive reaction type may oversimplify the multidimensional and evolving nature of client resistance. Several metrics, such as conversation length, cooperation rate, and adjacent-turn embedding similarity, are also imperfect proxies for challenge and psychological coherence.

Questions for authors

Could the authors clarify whether the 5P profiles contain information revealed after the evaluated response, and report results using prefix-only profiles? How many samples were included in the human comparison, was the evaluation blinded, and were the differences statistically significant? How does ResistClient compare with stronger training-free baselines using development-set prompt optimization, retrieval, self-consistency, and can the authors provide an ablation isolating the contributions of process rewards and the consistency reward?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.