acceptodds
Desk rejected at ICLR 2027

First, Do No Harm: AI Supervisor Scaffolds Novice Growth in Counselor Education

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

The most dangerous mistakes a novice counselor makes are not the obvious ones: they are utterances that sound caring while quietly violating professional ethics and leaving vulnerable clients less protected. We build an AI supervisor that does not replace novice counselors, but grows them—teaching them to internalize ethical violations they would otherwise never notice. What makes this supervisor non-trivial is not detection but teaching: it must locate the ethical-violating utterance, diagnose the ethical violation against APA principles, and deliver feedback that explains not just what went wrong, but why it is risky and how to respond differently. The core obstacle is that (1) ethical violations are by nature unlabeled in real clinical data, and (2) existing AI counselors trained only to match correct answers will never learn to teach. We resolve both at once: a controllable AI novice that intentionally enacts predefined mistake categories makes supervision labels a natural byproduct of generation, yielding ETHICSCAFF, a human-in-the-loop dataset; and GRPO under a Novice Growth Reward (NGR) optimizes the supervisor not for answer correctness but for whether a weaker novice model actually improves after reading its explanation. Across patient-disjoint tests, our best models reach 94.37% F1 for violated-principle classification and 74.24% F1 for violation localization; GRPO further improves localization precision, enabling the supervisor to surface context-dependent ethical risks that keyword or retrieval baselines can miss. Expert ratings favor our teaching-oriented explanations for professional depth and actionability, while supervised novices improve across all six counseling metrics. In a four-week pilot, our feedback produces the highest self-efficacy trajectory among all conditions, suggesting guidance that helps learners turn detected problems into safer responses.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large language models, LLM supervision, counselor education, mental health AI, human-in-the-loop dataset, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_E32HOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors focus on developing models that teach humans more efficiently rather than performing tasks themselves, specifically considering counselor training as a real-world application. To do so, they first created a dataset for this task, with different scenarios simulating practitioner-client interactions, annotated with common errors and the feedback a junior practitioner would receive. They then trained models using SFT on this dataset to align them with the structure and knowledge of the task, and finally fine-tuned them using GRPO, based on the improvement of a student LLM in solving the task after receiving the generated feedback. They analyze the performance improvements achieved by the trained model on this task, evaluating how well it identifies errors and provides feedback. Finally, they also tested the feedback provided by the trained model with Master's students and showed that it helped improve their performance over time.

Critical strengths
  • The paper is generally well written and well motivated. It is easy to read and follow, without being filled with unnecessary details that are not particularly informative for the reader. It is also very well structured, and the reader can, in a sense, anticipate what comes next, which greatly helps with understanding the paper without feeling lost.

  • The authors use the downstream helpfulness of the model's generated output as a training signal. This is interesting and perfectly aligned with their objective. I appreciate this elegant solution.

  • I think the experimental setup is very well designed, also covering a real-world use case in which the model is used to train Master's students, with positive results.

Critical weaknesses
  • Sometimes, important details needed to properly follow the paper are missing from the main text. They are present in the appendix, but, for instance, discussing the results related to VGR without explaining what VGR is (even at a very high level) makes it difficult to follow (similar examples apply to MITI, PSC, and WA). The same issue also occurs in some parts of the method section (e.g., Prec, R_loc, etc.).

  • The discussion of the results is rather basic. The authors mostly describe what is already presented in the figures and tables without providing deeper insights into what those results mean. For instance, why is the improvement in Task 1 much more significant than in Task 2?

  • Figure 4 is not particularly informative. It is more visually appealing than useful. Statistical measures (e.g., Cohen's kappa) would be much more useful and give the reader a clearer picture of the results.

Questions for authors

See the questions above.

  • If you observe a strong correlation between human raters and the model you used in Section 4.3 to evaluate your models' answers, why not use its preferences to directly optimize the models as well?

  • How well would a model trained to perform the counselor task itself perform? Would it perform better than a student receiving feedback or a model receiving feedback? Is giving feedback more difficult than performing the task itself?

4 comments
AuthorsOpenReview ↗

Response to Reviewer E32H (1/2)

We sincerely thank you for your appreciation of our solution as "elegant." Your suggestions help us clarify the method and strengthen the evaluation.


Weaknesses


  • W1. Key terms and notation are insufficiently explained in the main text.

    • We agree that information needed to understand the method and results should be introduced clearly where it first appears. We will explain the VGR process, spell out MITI, PSC, and WAI and describe what they measure, and define the relevant symbols near the reward formulation.
  • W2. Why does Task 1 (classification) improve more than Task 2 (localization)?

    • We hypothesize that the two tasks benefit differently from the training data. The data are generated around 15 mistake categories, which may help the model learn the boundaries between dialogue cues and error categories relatively quickly. Localization, however, requires identifying an average of 2.74 problematic utterances in dialogues averaging 12.08 turns. The same type of mistake can appear in different wording, and its ethical meaning depends on the surrounding context; localization therefore also requires selecting the right utterances and interpreting them in context.

    • In Table 3, the model before our training already achieves 91.04% recall on localization, suggesting that it usually finds candidate problematic utterances. Further improvement depends on reducing false positives and identifying the exact problematic utterances. We therefore hypothesize that learning category boundaries contributes to the larger gain in classification, while contextual ambiguity and the already high baseline recall constrain the gain in localization.

  • W2. Deeper discussions

    • We thank the reviewer for encouraging a deeper discussion of the results. We have reconsidered their pedagogical implications and will add the following interpretation to the discussion: The core teaching value of feedback lies in helping novices translate abstract ethical principles into actionable counseling responses through scaffolded guidance. Effective supervision first identifies a concrete problem, then explains the relevant principle and potential harm, and finally suggests a feasible revision. Our results are consistent with this interpretation: relative to the novice without feedback, our feedback produces larger gains than baseline feedback on MITI (+1.60 vs. +0.54), PSC (+1.03 vs. +0.33), and WAI (+1.38 vs. +0.37). In the four-week human pilot, self-efficacy in our feedback group continued to rise in later weeks, whereas the baseline-feedback group tended to plateau. These results support the educational value of the feedback.
  • W3. Thank you for suggesting agreement statistics.

    • On the 300 test samples, Cohen's κ between human and model judgments is 0.72 for Objectivity & Fairness, 0.88 for Constructiveness & Actionability, 0.82 for Professional Depth, 0.78 for Comprehensiveness, and 0.66 for Clarity & Structure. The mean κ across the five criteria is 0.77, indicating substantial agreement. We will add these values to Figure 4.
AuthorsOpenReview ↗

Response to Reviewer E32H (2/2)


Questions


  • Q1. Thank you for raising this question.

    • Our central claim is that the supervisor both produces professionally sound feedback and helps novices improve through that feedback. We evaluate this claim from two complementary perspectives. The expert perspective assesses whether feedback is professional, objective, comprehensive, clear, and actionable, as examined by the expert and model evaluations in Section 4.3. The novice perspective asks whether learners can use the feedback to improve their judgments and responses, as examined by task performance before and after feedback and by downstream counseling performance.

    • The agreement between human and model ratings noted by the reviewer supports the expert perspective on feedback quality. Yet an explanation that experts understand and endorse may rely on knowledge a novice has not acquired, or may not address the novice's particular misunderstanding. Highly rated feedback may therefore leave a novice unsure how to judge the case or revise a response. Section 4.3 provides evidence for important aspects of feedback quality, but rating agreement alone does not establish its practical benefit to novices.

    • We chose NGR to incorporate the novice perspective directly into the supervisor's optimization objective. We keep the simulated novice model fixed, compare its ethical-violation localization and classification performance before and after reading explanatory feedback, and use the resulting gain as a reward in GRPO. This provides a direct signal for optimizing explanations that novices can understand and use, in line with our claim about instructional guidance.

  • Q2. Thank you for raising this question.

    • We additionally evaluated the ETHICSCAFF-trained Qwen3-8B supervisor acting directly as a counselor, using the same six counseling quality metrics as in Table 4. It achieves the highest scores among the compared conditions on EFT-TFS, HTAIS, PSC, and TES, exceeding the trained counselor model PsyDT on those four metrics. This indicates that the trained supervisor also performs well when counseling directly.

    • However, the novice receiving its feedback scores higher than the supervisor counseling directly on two dimensions: 7.22 versus 6.24 on MITI (Motivational Interviewing Treatment Integrity), and 6.06 versus 5.16 on WAI (Working Alliance Inventory). Providing feedback and counseling directly have different objectives and output requirements, and the supervisor shows different strengths in these two roles. A teacher need not earn the highest exam score to explain concepts clearly and help a student learn. This is why we assess a supervisor's teaching value by whether its feedback helps learners improve their performance. These results alone do not establish which of the two tasks is intrinsically harder.

Table 4: Performance Comparison Across Counseling Quality Metrics Under Different Counselor Conditions and Feedback Settings. All metrics are higher-is-better.

ConditionMethodEFT-TFSHTAISMITIPSCTESWAI
No FeedbackNovice Counselor2.214.795.625.987.654.68
No FeedbackNormal Counselor (PsyDT)3.015.347.357.148.736.00
With FeedbackBaseline Feedback2.785.216.166.317.805.05
With FeedbackOur Feedback2.975.357.227.018.546.06
Supervisor Trained on ETHICSCAFFQwen3-8B Full4.135.406.247.909.585.16
Improvement over NoviceBaseline Feedback+0.57+0.42+0.54+0.33+0.15+0.37
Improvement over NoviceOur Feedback+0.76+0.56+1.60+1.03+0.89+1.38

We hope these clarifications address the reviewer's concerns, and we would be grateful if they could be considered in the final assessment.

I thank the authors for the additional clarification, and most of my concerns have been addressed. I will increase my score to 4.

AuthorsOpenReview ↗

Further Insights on Teaching-Oriented AI Supervision

Thank you for your positive reassessment and thoughtful feedback! Your questions helped sharpen two insights, which we briefly highlight below.


  • Finding & Insight 1: The largest feedback gains occur precisely where the supervisor lacks a direct-task advantage.
    • A notable, counterintuitive finding is that the dimensions on which the supervisor helps novices improve the most are not those on which it has the strongest direct-task advantage. When acting directly as a counselor, the supervisor achieves the highest scores among the compared conditions on EFT-TFS, HTAIS, PSC, and TES. On MITI and WAI, however, the novice receiving its feedback scores 7.22 and 6.06, exceeding the supervisor’s direct scores of 6.24 and 5.16. In this sense, the novice surpasses the supervisor in specific dimensions, not independently, with the supervisor’s guidance. These are also the two metrics with the largest gains from our feedback relative to no feedback: +1.60 and +1.38, respectively. In other words, the dimensions on which the supervisor is outperformed in direct counseling are precisely those on which its feedback yields the greatest novice improvement.

    • MITI concerns collaborative interviewing, while WAI concerns agreement on therapeutic goals and tasks and the counselor–client bond. One possible explanation is that producing a strong response and diagnosing, explaining, and guiding the revision of another’s response draw on partially different capabilities. For example, the supervisor may identify overly directive responses or failures to consider the client’s perspective and suggest adjustments that the novice incorporates into its own response. Effective feedback does not perform the task on the learner’s behalf; it enables the learner to perform it better, even, in some dimensions, better than the supervisor itself.

    • This finding suggests that instructional effectiveness cannot be inferred solely from a model’s own task performance. Instructional models require an independent assessment of how much they help learners improve, beyond how well they perform themselves. This motivates NGR: by directly optimizing for performance gains in a frozen simulated novice after feedback, we target not merely professionally sound feedback, but feedback that learners can understand and use.


  • Finding & Insight 2: Bridging principles and practice: contextualization as a core supervisory capability
    • The challenge in ethics instruction is not merely helping novices know the rules, but helping them understand how those rules apply to a specific interaction. “Respect client autonomy” is a general principle; recognizing when a suggestion becomes a decision made on the client’s behalf requires interpreting the wording, surrounding dialogue, and the client’s expressed preferences. A supervisor’s key role is therefore to bridge abstract principles and concrete practice: explaining what a principle means in the present context and translating it into targeted guidance for revision.

    • ETHICSCAFF scaffolds this bridge through structured feedback format that connects problematic utterances with relevant ethical principles, potential consequences, and revision directions. NGR uses gains in a frozen simulated novice’s violation localization and classification after reading the feedback as an optimization signal. The objective thus goes beyond producing professionally sound explanations of principles to asking whether novices can use those explanations in case-specific judgments. Contextualization is therefore not merely a way of presenting structured feedback, but an instructional capability that connects clinical principles with novice action.

    • In the four-week human study, 36 counseling-psychology master’s students were assigned to three conditions: no feedback, feedback from a base model, or feedback from our supervisor. The group receiving our feedback showed the largest cumulative increase in self-reported counseling self-efficacy, with continued gains in later weeks. These findings provide complementary evidence from actual trainees for the value of this bridging approach: its benefits are reflected not only in simulated novices’ task performance, but also in trainees’ confidence in counseling practice. The implication is that supervisors need not only clinical principles, but also the ability to translate those general ones into context-specific guidance that novices can understand and use.

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

This paper introduces EthicScaff, a synthetic dataset for counselor supervision, and trains an AI supervisor using SFT and GRPO to help novice counselors understand and apply ethical principles. The training rewards explanations that help a weaker model locate and classify counseling mistakes.

Critical strengths
  1. The paper addresses a meaningful problem - helping novice counselors recognize ethical mistakes and improve their responses.
  2. The evaluation is very comprehensive. It covers detection, expert feedback ratings, and even a four-week study with counseling students to see if the feedback is helpful.
Critical weaknesses

The evaluation is thorough. My main question is whether a strong, recent LLM could achieve similar results with prompting alone:

  1. The models in Table 2 are somewhat dated, so it is hard to tell how the method compares with more recent models.

  2. It’s also unclear how much training adds. A strong, recent LLM prompted to act as the AI supervisor might do just as well with the same ethical principles and feedback instructions.

Questions for authors

Were participants (Section 4.5) in the two feedback conditions blinded to model identity?

1 comment
AuthorsOpenReview ↗

Response to Reviewer MfLb

We thank the reviewer for recognizing the motivation, experimental design, and methodology of our work. Your suggestions help us clarify the method and strengthen the evaluation.


Response to Weaknesses


We thank the reviewer for suggesting the inclusion of recent strong models to more directly assess the benefits of task-specific training. We evaluated 7 additional models, with the results presented below. Qwen3-8B Full, trained on ETHICSCAFF achieves the highest scores on eight of the nine metrics. These results support the value of ETHICSCAFF and its training pipeline.

ModelT1_AccT1_PrecT1_RecallT1_F1T2_PrecT2_RecallT2_F1T2_JaccardT2_EM
gpt-6-sol64.6472.3664.6461.7360.9596.4171.8759.8014.32
gpt-6-luna56.9467.5956.9453.0161.8791.6770.2058.4615.28
claude-opus-5.556.1567.6056.1552.2960.0795.4370.8858.8013.69
DeepSeek-V4.1-Flash-101763.1070.0763.1059.9561.4297.8472.5060.4715.03
Qwen3.8-27B40.1759.9940.1734.5861.3890.2868.8256.9212.82
Qwen3.5-35B-A3B45.0959.9945.0937.2360.5497.3771.9259.8714.42
gemma-4-12B40.0656.5240.0632.5660.2695.2870.6258.3211.97
Qwen3-8B Base (Table 2)23.3344.1623.3322.1854.7860.7653.5640.523.87
Qwen3-8B SFT (Table 2)73.6182.1773.6175.2055.9992.2666.8555.2612.29
Qwen3-8B Full (Table 2)88.9891.1188.9888.6366.7091.4574.2463.0318.40

Response to Question


Yes. Participants in both feedback conditions were not informed of the model identity or whether they were receiving feedback from the base model or our trained supervisor. This procedure was intended to reduce expectation bias associated with model identity.


We hope these clarifications address the reviewer's concerns, and we would be grateful if they could be considered in the final assessment.

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

This paper targets AI assisted training for novice psychological counselors. Its core contribution is optimizing feedback for learner improvement instead of static answer matching. It provides a feasible synthetic data pipeline for domain specific supervision tasks where real annotated human data is not available. It presents a structured supervision paradigm, including locating problematic utterance, classifying ethical violation, explaining harm mechanism, and giving concrete improvement suggestions.

Critical strengths
  1. This paper shifts optimization target from “generate correct critique text” to “generate feedback that teaches”, which aligns better with the actual supervision goal. 2)This paper proposes to generate dialogues by simulating mistake‑prone novices and behavior‑consistent virtual patients, to produce natural, context‑dependent subtle ethical errors. Combined with validator‑guided refinement and clinical expert auditing, ETHICSCAFF fills the gap of pedagogical supervision datasets for counselor training. 3)The paper proposes a NGR learning reward which is determined by whether feedback improves the downstream task performance of a weaker novice agent. This structured formulation pushes models beyond vague high‑level criticism and enforces concrete evidence‑grounded supervision.
Critical weaknesses
  1. All datasets are generated via LLM simulation, even with expert‑defined rules and human curation. It works well for synthetic test patients within the same simulation framework. However, the patterns of mistakes may not fully capture the full complexity, long‑term contextual dynamics and real‑world clinical nuance seen in authentic therapy sessions. The empirical results reveal that our improvements are smaller relative to PsyDT. This indicates that the synthetic dataset delivers meaningful benefits, while also exhibiting clear performance ceilings and associated limitations. Furthermore, patient‑disjoint split only generalizes to the unseen simulated patient profiles within the same synthetic framework, not real clinical data. Performance drop is expected when migrating to real clinical transcripts, as real‑world mistakes are more unstructured, multi‑turn cumulative, and not strictly mapped to the predefined mistake categories. 2)NGR uses a frozen LLM agent to approximate human novice improvement. There exists a distribution gap: what improves a simulated model may not always transfer equally to human counseling trainees with diverse backgrounds, prior knowledge, and cognitive patterns. The human pilot is preliminary. 3)NGR’s effectiveness is bounded by simulated‑novice proxy. It works for the simulated agent, but improvement for simulated model does not perfectly guarantee human trainee improvement. The human experiment is small‑scale (36 participants), only measures self‑reported self‑efficacy, no objective long‑term clinical behavior measurement.
Questions for authors

1)The predefined ethical violation categories focus on novice‑level linguistic ethical missteps. It does not cover full‑scope psychotherapy supervision needs: multi‑session case conceptualization, complex differential diagnosis, modality‑specific intervention strategies, jurisdiction‑specific legal standards. It targets ethics education only, not comprehensive clinical skill training. Dataset quality relies heavily on multiple auxiliary LLMs for simulation, validation and refinement plus clinical expert human review. Performance is bounded by the quality of these upstream components. The pipeline has high computational and human‑labor overhead. The full ETHICSCAFF dataset uses controlled research‑only release, which raises barrier for other researchers to reproduce and extend the work. 2) The paper does not test failure modes: e.g., ambiguous dialogues, adversarial inputs, highly implicit violations, or cases where multiple overlapping ethical mistakes co‑occur. It lacks analysis of false‑positive / false‑negative failure patterns of the AI supervisor.

2 comments
AuthorsOpenReview ↗

Response to Reviewer bfZh (1/2)

We thank the reviewer for recognizing our teaching-oriented supervision framework and the value of ETHICSCAFF. We first clarify two possible misunderstandings that underlie several of the concerns.


Clarifications


  • C1. Educational training and clinical practice have different scopes. (W1, W3, Q1)

Real counseling is complex. It involves long, context-dependent interactions in which mistakes are often ambiguous and cumulative. Education, however, requires clear and standardized targets: novices must first learn to reliably recognize well-defined ethical violations before they can handle the full complexity of clinical practice. We therefore design controlled, repeatable exercises with simulated clients around explicit ethical principles and learning objectives. We start with ethical "red lines", the fundamental mistakes that may harm clients. In this setting, our supervisor scaffolds learning by locating the problematic utterance, explaining its potential harm, and suggesting a revision, which helps novices bridge theoretical knowledge and practical responses. The simplification is a deliberate pedagogical choice, not an attempt to replicate real sessions.

  • C2. In Table 4, PsyDT is a reference for counseling quality, not a baseline for our supervisor. (W1)

A counselor model is evaluated primarily on the quality of its direct responses to clients, whereas a supervisor model is evaluated on whether its feedback helps novices improve their performance. Accordingly, the model compared with PsyDT in Table 4 is the novice counselor after receiving our feedback. This comparison examines whether supervision helps novices narrow the gap with a professionally trained counselor model. The novice's performance approaching PsyDT on some metrics reflects progress toward this educational goal. Educational gains should also be assessed through improvements over the no-feedback and baseline-feedback conditions.


Weaknesses


  • W1 (Synthetic data).

Public clinical dialogues with reliable mistake annotations are scarce, and using them requires privacy safeguards, informed consent, and ethical review. We therefore rely on simulated dialogues. As acknowledged in Appendix A.1, they simplify real situations; as clarified in C1, this simplification is intentional for novice ethics education. To test whether the benefits extend beyond the training data, we evaluate on two levels. Within the simulation framework, the patient-disjoint split tests generalization to unseen client contexts. Beyond simulation, our four-week human pilot provides preliminary evidence with actual trainees: participants who received our feedback showed more favorable changes in self-efficacy.

  • W2 & W3 (Simulated-novice proxy).

NGR rewards learner improvement. A frozen, weaker novice model first localizes and classifies the violation, then answers again after reading the supervisor's feedback. The improvement between the two attempts is the reward, which we optimize with GRPO. Freezing the novice yields a stable, repeatable signal focused on the supervisor's teaching ability. Ablations show that NGR improves both classification and localization.

Repeated human evaluation across many RL iterations is impractical [1]. LLM-based user simulators are also widely used for repeatable evaluation in emotional dialogue, social simulation, and customer service. We acknowledge that models and human trainees differ in prior knowledge, cognition, and how they interpret feedback. We therefore treat NGR as a tractable training proxy, with the human pilot providing an initial check of transfer to real trainees. More diverse trainee populations, objective behavioral measures, and longitudinal follow-up are important next steps.

[1] Tian X, Lin Q. Simulating Human Participants with Large Language Models: Principles, Limitations, and Recommendations[J]. Journal of Psychological Science, 2026, 49(4): 770-782.

AuthorsOpenReview ↗

Response to Reviewer bfZh (2/2)


Questions


  • Q1.

Scope: See C1. Appendix A.2 states that the taxonomy "provides an ethics-centered foundation rather than an exhaustive supervision curriculum." It lists multi-session case formulation, modality-specific intervention timing, and jurisdiction-specific standards as future extensions.

Cost: Inaccurate labels could mislead novices on sensitive ethical issues. We therefore combine model-based validation, VGR, clinical-psychologist revision, and joint expert review. This upfront cost is amortized over repeated use of the dataset and supervisor.

Access: The full dataset contains sensitive examples of inappropriate counselor behavior. It is therefore available through a research-use request and a data-use agreement to reduce the risk of misuse.

  • Q2.

Consistent with C1, each dialogue centers on a single predefined mistake to give novices a clear learning target: locate one problem, understand its harm, and revise the response. Mixing overlapping or ambiguous errors from the outset would complicate annotation and make model errors harder to attribute. Natural extensions include multi-label prediction with per-violation probabilities and cited evidence, as well as recommendations prioritized by potential harm and urgency.

We hope these clarifications address the reviewer's concerns, and we would be grateful if they could be considered in the final assessment.

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

Related papers

Discussion (0)

Sign in to comment.