acceptodds
Under review as a conference paper at ICLR 2027

MoSE-Counseling: A Mixture of Safety Experts Framework for Large Language Model-Based Psychological Counseling

Abstract

Large language models (LLMs) create new opportunities for scalable psychological counseling, yet general-purpose LLMs lack specialized safety mechanisms: they frequently miss implicit crisis signals such as suicidal ideation and respond with trivializing or boundary-crossing replies, undermining reliability in high-stakes settings. We propose MoSECounseling, a two-stage Mixture of Safety Experts framework. Stage 1 builds a structured, evidence-grounded client profile (demographics and four severity-rated psychological dimensions) using a constrained prompting scheme designed to limit hallucination. Stage 2 routes the dialogue through specialized LoRA experts—risk detection, ethical guidance, safe response, and a coordination expert—using a profileconditioned, risk-monotone router with a profile-independent safety floor. The floor prevents an imperfect profile from overriding a crisis signal that has already been detected, but does not eliminate screening false negatives. Across SOS-HL-1K, PsyQA, and , MoSE-Counseling achieves consistent, statistically significant improvements in safety and professionalism compared to strong fine-tuned and domain-specific baselines. We deliberately frame these as benchmark improvements rather than evidence of clinical equivalence: our results do not establish real-world safety for crisis intervention, and high-risk detection remains imperfect. We therefore report controlled baselines that isolate the design’s contribution from prompting and fine-tuning, a safety-specific evaluation of missed crises and harmful responses, expanded ablations and robustness analysis, a detailed professional-evaluation protocol, and an ethics discussion. MoSE-Counseling is intended as a supplementary screening and drafting aid under human oversight, not a replacement for licensed professionals.

Primary area:
generative models
Keywords:
Psychological Counseling, Large Language Model, Safety Experts

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

MoSE-Counseling is a two-stage pipeline for Chinese counseling benchmarks. Stage 1 extracts a structured client profile with evidence spans. Stage 2 routes among LoRA experts for risk, ethical critique, and response generation, then always runs a coordination expert. A profile-independent screening score can force the risk expert on, and a max-latch keeps a crisis score from decaying across later calm turns. The paper evaluates the system on held-out splits of SOS-HL-1K, PsyQA, and D4, plus a 300-case set it calls Unseen-Crisis.

Critical strengths

Section 4.4 states that every MoSE entry on SOS-HL-1K comes from the standalone screening head, and that changing the profile or the routing thresholds cannot change those labels. Table 4 then measures routing itself: the routed system misses 16.3% of crisis signals against 15.7% when all three experts run, at 336 ms against 462 ms. The professional protocol is also concrete: five licensed counselors, 120 held-out cases per dataset, responses anonymized and shuffled, with ICC(2,k) = 0.83.

Critical weaknesses

Motivation and novelty. The problem is real: counseling models miss crisis signals and cross boundaries. The system is an assembly of existing pieces: evidence-constrained extraction, threshold routing, task-specific LoRA experts, and a max in place of an exponential moving average. The persistence audit shows that a weight of 0.7 retains 0.7^N of a unit signal; that identity is not a result. Closely related Chinese crisis evaluation on SOS-HL-1K and PsyQA, and multi-agent counseling pipelines that already separate profiling from response generation, are absent from the references. A conservative if-then router on known datasets does not meet the novelty bar of a top conference.

Experiments. The number the architecture is not allowed to change is the number used to claim the architecture works. Qwen-MoSE reaches 78.02 against 76.40 for the best fine-tuned baseline. Table 2 then reports risk-class F1 of 68.5 against 63.5 for token-routed MoE-LoRA and calls the 5-point gap evidence for specialized experts and conservative routing. Section 4.4 says those SOS-HL-1K labels are produced by the screening head alone. The comparisons that do include routing are small. Professional scores rise by about 0.16 to 0.20 on a 0–10 scale (8.76 / 8.93 / 9.10 against 8.60 / 8.75 / 8.90), with no significance test. Removing coordination costs 0.08, 0.08, and 0.15. On Unseen-Crisis, routed MoSE has risk-class F1 67.0, while single-LoRA plus Llama Guard has 67.9; the paper compares MoSE with single-LoRA at 65.2 and calls the drop from 68.5 to 67.0 a meaningful domain shift. Unseen-Crisis has no citation. The code supplement states that the manuscript does not record c_min or τ_floor.

Writing. The scope limits are repeated and clear: the router is not a safety guarantee, and the study is not a clinical trial. The same pages define the headline metric as independent of routing and then attribute its gain to routing. That contradiction is the part a reader cannot resolve.

Questions for authors

With the screening head held fixed, which main-text comparison shows that conservative routing changes a safety outcome by more than the 0.6-point miss gap in Table 4, and is that difference significant? Where is Unseen-Crisis released, and why is Llama Guard’s 67.9 risk-class F1 not the comparison on that set? I would need both answers before this was a venue-level contribution.

Reviewer_LC4BOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MoSE-Counseling, a framework with four LoRA experts for risk detection, ethical guidance, response generation, and coordination. It uses risk-based routing and persistent risk memory, and reports improvements on Chinese risk-classification and counseling-response qualities.

Critical strengths
  1. This paper cares about an important aspect of building counseling agents - safety and ethical guidance. The division of responsibilities is intuitive.
  2. The evaluation involved expert ratings.
Critical weaknesses
  1. The necessity of their 4 experts is not sufficiently justified. Some of these tasks already have simpler baselines. For example, the SOS-HL-1K paper reports ~82% F1 with BERT, although the evaluation setup differs. Could task-specific models or few-shot prompting handle risk and quality scoring, with one LLM using their outputs to generate the final response?
  2. The reported improvements appear modest. The paper reports significance for selected automatic metrics, but whether the gains are robust to sampling variability or persist against stronger prompting baselines is unclear.
  3. Some experiment details are missing, e.g. the exact configurations and prompts for Table 2 for fair comparisons.

Minor Issues: Please check the citation formatting, especially the use of \citet{} versus \citep{}.

Questions for authors

What evidence or prior work supports the design of asymmetric update? Could it be simply recomputing the profile from the full conversation history at each turn?

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

MoSE-Counseling is a two-stage Chinese counseling pipeline combining evidence-grounded client profiles with specialized risk, ethical-guidance, response, and coordination LoRA experts. A profile-independent screening head imposes a risk-expert activation floor, while a maximum-based risk update preserves vigilance after a crisis disclosure. The study reports held-out classification and response metrics, blinded counselor ratings, controlled adaptation comparisons, and additional routing, transfer, and persistence experiments.

Critical strengths
  1. Five licensed counselors blind-rated outputs on 120 held-out cases per dataset. The adaptation comparisons use the same backbone, safety prompt, and client profile, making the results more credible.
  2. The paper show how routing and persistent risk tracking affect safety, latency, and over-intervention. These results clarify the design’s benefits and costs within the stated benchmark scope.
Critical weaknesses
  1. The routing contribution is unclear: SOS-HL-1K predictions come only from an independent screening head (Section 4.4), so the +5.0 risk-F1 gain cannot demonstrate a routing benefit as claimed in Section 4.6. A detector baseline with matched training and readout is needed.
  2. Key safety evaluation details are missing: The 300-case Unseen-Crisis set lacks a clear source or construction protocol (Section 4.1). Tables 3–5 omit sample sizes, failure-label definitions, and uncertainty estimates; Table 7 omits its promised confidence intervals. These gaps make the safety and transfer results difficult to assess.
Questions for authors

Please identify Unseen-Crisis and supply the sample counts, failure-label definitions, sequence construction, and paired intervals for Tables 3–5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.