acceptodds
Under review as a conference paper at ICLR 2027

HealingAxis: Large Language Models Can Innately TherapyMaxxx Clinical Psychotherapeutic Interventions

Abstract

AI chatbots powered by Large Language Models are the largest providers of mental health support, with emotional support as their top use-case. Despite this scale, the inner workings of models providing psychotherapeutic support remain largely un- known. In this paper, we identify a low-dimensional direction, a Healing Axis, as- sociated with the application of psychotherapeutic techniques from a dozen modal- ities studied including CBT, EFT, AEDP, psychodynamic, schema, and other ther- apies. For each of the six models examined, ideal activation steering increases by +22-232% the number of verifiable clinical skill criteria the model exhibits in its response; while negative steering collapses responses toward zero clinical skills. We propose the GoodBadTherapy dataset with 2,000+ contrasting pairs establishing the Healing Axis, and the SkilledTherapyBenchmark to evaluate the performance of models on deliberate-practice therapeutic skillsets. Using mecha- nistic interpretability methods, we demonstrate the Healing Axis is composed of interpretable features and sub-axes are meta-therapeutic techniques. Additionally, the Healing Axis on average generalizes beyond the modalities it’s derived from, demonstrating a single vector has cross-modality applicability. More broadly, as models become embedded in the mental health support landscape, this work offers a path to monitor and improve the skillfulness of their clinical interventions.

Primary area:
interpretability and explainable AI
Keywords:
Psychotherapy, LLMs, Mechanistic interpretability, Deliberate Practice

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_F6pxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The useful part is that they do not stop at “the representations separate.” They intervene and look at behavior, and they do it on more than one model family. That is a better starting point than another output-only comparison.

The pairs come from actual training books, and they do try to kill the obvious alternative explanations. Style is matched on the surface, and they also run a version where both sides are LLM-written. AUC stays very high after that, so I do not think this is just human textbook prose versus model prose.

Critical strengths

The useful part is that they do not stop at “the representations separate.” They intervene and look at behavior, and they do it on more than one model family. That is a better starting point than another output-only comparison.

The pairs come from actual training books, and they do try to kill the obvious alternative explanations. Style is matched on the surface, and they also run a version where both sides are LLM-written. AUC stays very high after that, so I do not think this is just human textbook prose versus model prose.

Critical weaknesses

I am not convinced the judge is counting skills. Telling a textbook answer from a synthetic misstep is a different task from labeling individual skills in a free generation. Figure 1 is the example that stuck with me: a short acknowledgment and a clarification question are given ten skills. Several of those look like the same move under different names. Unless someone who trains or practices this work has checked the generations, a higher count does not tell me the response is a better intervention.

The steering result is also thinner than the writeup suggests. The gains come from searching layers and strengths on twelve inputs per setting, and it is not clear that this search was frozen before the numbers being reported. There is no therapy-prompt baseline, no generic empathy or helpfulness direction, and no random direction matched on norm. Right now I cannot tell whether the axis is about psychotherapy skill or about sounding warmer and more supportive.

Questions for authors

Have clinicians or trainers scored the generated responses, both for whether each skill is really there and for whether it fits the moment? For the response in Figure 1, were the ten labels checked one by one, and how did you treat skills that describe the same utterance?

Do you have a comparison to a therapy prompt, to an empathy or helpfulness direction, or to random directions with the same norm? I would want that before reading the gains as specific to the Healing Axis.

Reviewer_w6PVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper identifies an internal direction in LLMs that represents therapeutic skills. They first build a dataset of good and bad therapy responses, then show these responses are separable in the models internal representations. Next, they use and manipulate activation steering to explore different response and sub-axes using interpretability methods.

Critical strengths

the idea is very interesting, identifying an internal direction in LLMs representing therapeutic skills and using activation steering to improve responses.

Critical weaknesses
  • The outline of the paper requires significant improvement. The paper starts with dataset construction before mentioning the scientific research question, so the reader encounters the methodology before fully understanding the scientific question. The contributions and bigger picture are not clear;

  • The authors move between different scientific claims without mentioning the connections between them.

  • All the claims and results depend on the validity of the dataset; however, the paper lacks a systematic review or validation of GoodBadTherapy by independent clinical experts.

Questions for authors

To my understanding, your LLMs were not explicitly prompted to act as therapists. I'm curious whether you also experimented with prompting the models to provide therapeutic support. Would the Healing Axis still provide the same improvements, or could a well-designed prompt achieve similar results?

Reviewer_z8JCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper is about “Healing Axis” derived from expert-written good therapy responses and LLM-generated bad responses, and uses activation steering to increase therapeutic skills. The authors show that this direction separates the two response groups, that activation steering along it increases the number of therapeutic skills detected in model outputs, and that it partially generalizes across therapy modalities.

Critical strengths

The paper studies an important and underexplored problem at the intersection of mechanistic interpretability and mental-health support. It covers multiple therapy modalities and open-weight models, provides extensive analyses and ablations, and offers an interesting direction for understanding and controlling therapeutic language in LLMs.

Critical weaknesses

My concerns are about construct validity and measurement validity.

Unvalidated evaluation judge: Layer selection, steering-strength selection, and cross-modality evaluation all rely on skill counts from the same GLM-5.2 judge. As acknowledged in Section 6.1, its agreement with human experts, other judge models, or external benchmarks has not been established. Therefore, the results demonstrate higher judge scores rather than improved therapeutic quality.

Limited diversity of synthetic negative responses: The bad responses are generated by an LLM using a small set of predefined missteps, which may produce concentrated content and linguistic patterns. The Healing Axis may therefore capture artifacts of the generation process rather than a general direction representing therapeutic quality. External validation on human-produced high- and low-quality therapy responses or expert-labeled LLM responses can help reduce concerns about this.

Limitations of single-turn evaluation: Many therapeutic skills depend on conversational history, appropriate timing, adaptation to the client’s reactions, and the development of a therapeutic relationship. These properties cannot be reliably evaluated from a single client utterance and therapist response.

Questions for authors
  • Questions for the authors: Can the authors validate the judge and steering results using blinded ratings from clinical experts or external human-annotated datasets such as CounselBench?
  • Does the Healing Axis generalize to human-produced high- and low-quality therapy conversations without retraining (the highlowquality therapy dataset )?
  • How diverse are the synthetic negative responses, and what evidence rules out content, style, or generation-source artifacts?
  • Can the approach be evaluated in multi-turn conversations where timing, adaptation, and therapeutic alliance can be assessed?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.