K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversations
Laura M. Vowels, Matthew J. Vowels, Shivali Sharma, Apoorv Jha, Rehnuma Choudhury, Wasseem El Sarraj, Rachel Francois-Walcott, Aruba Hussain, et al.
cs.CL · cs.AI · cs.LG
Quick verdict
K-Bench is the broadest clinician-calibrated mental-health conversation benchmark to date (125 configs, 200 vignettes), but its headline 94.2% judge agreement is measured on its own calibration set against author-adjudicated consensus.
Summary
K-Bench is a clinician-calibrated, contamination-protected benchmark for evaluating LLMs in multi-turn, high-risk mental health conversations. Every evaluated configuration runs against a fixed cohort of 200 vignettes rendered into conversations of up to 20 talk turns, covering suicide, self-harm, domestic violence, substance misuse and no-risk presentations; vignettes derive from lived-experience source material expanded through a 122-variable factorial design. Scoring is done by a frozen GPT-4o judge against a 7-dimension, 47-component clinician rubric that separates risk recognition, risk exploration and therapeutic conduct. Validation rests on two bodies of evidence: synthetic patient turns overlap closely with a de-identified corpus of 50 real user-AI conversations (nearest-neighbour cosine distance 0.404-0.486), and the judge reproduces clinician consensus on 94.2% of 6,751 eligible item comparisons, above the 89.4% clinician-clinician agreement. Across 125 configurations from 33 base models and 14 providers, overall scores span 81.19-98.96 and combined-risk scores 52.39-96.11; only risk exploration separates strong from weak systems (mean 79.9, range 22.1-93.1). Therapeutic prompting and elevated reasoning both produced configuration-specific, not uniform, gains. Limitations: simulated English-only single-episode conversations, ground truth adjudicated by the study's own leads, no held-out judge validation reported, and no clinical outcome measurement.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 3
● low 0
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Synthetic patient turns generated under a no-risk condition closely approximate real user-AI mental health conversation in embedding space (nearest-neighbour cosine distance 0.404-0.486 across four patient-agent models).
- A frozen GPT-4o judge reproduces clinician consensus on 94.2% of 6,751 eligible item comparisons across seven rubric dimensions, exceeding the 89.4% achieved between clinicians on 22,816 paired comparisons.
- Across 125 configurations from 33 base models and 14 providers, overall K-Bench scores are high (81.19-98.96) but combined-risk scores span 52.39-96.11, with risk exploration the only dimension that separates strong from weak systems.
- Therapeutic prompting and provider-exposed elevated reasoning both produce configuration-specific rather than uniform changes in risk performance, so neither acts as a general safety control.
- Protecting vignettes, disclosure schedules, judge prompts and transcripts from release is necessary to keep the benchmark a valid independent prospective test under AEF-1.
Strengths
Unmatched configuration-scale coverage on a single fixed test cohort
125 model-configuration entries spanning 33 base models and 14 provider namespaces, all evaluated on the same 200 vignettes with prompt, reasoning, version and cost recorded per row (Supplementary Table 1). No prior mental-health conversation benchmark (VERA-MH: 3 provider models; SIM-VAIL: 9 chatbots, 10 turns) approaches this breadth, and it is the only one that makes version/prompt/reasoning a tested variable rather than a footnote.
Risk exploration is separated from supportive tone, and that is where the signal is
Supplementary Table 2: D2 risk exploration mean 79.86 with a 22.09-93.10 range (71-point spread), versus D1 clinical judgement 99.76, D4 supportive conversation 99.77 and D6 cultural competence 99.14. The design therefore exposes failures that an empathic-tone score would hide, and correctly reports the overall leader as different from the combined-risk leader (GPT-5.5 vs Claude Fable 5; Gemini 3.1 Flash Lite moves from 91st overall to 41st on combined risk).
Clinician grounding is substantial and procedurally documented
Vignette construction is systematic rather than hand-written
Explicit about what it does NOT measure
Weaknesses
The realism evidence covers only the condition the benchmark is not about
Ground truth is authored, not independent
Four of seven dimensions sit at the ceiling, making the headline score near-degenerate
Severity and co-occurrence contrasts are presented as findings without a model that could support them
Nothing required to audit the result is available
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
K-Bench sets out to be the reference instrument for deciding whether a large language model can hold a safe conversation with someone in evolving psychological distress. Its scope claim is large and mostly delivered: 125 model-configuration entries representing 33 base models across 14 provider namespaces, each run on one fixed cohort of 200 vignettes rendered into conversations of up to 20 talk turns, covering suicide, self-harm, domestic violence, substance misuse and no-risk presentations, including risks that emerge indirectly, escalate, or co-occur. It states three research questions - whether synthetic patients and automated scoring are realistic, reliable and scalable (RQ1); how performance varies by risk domain, severity and co-occurrence (RQ2); and how prompting and reasoning configurations change risk performance (RQ3) - and answers all three with numbers.
Two claims deserve scrutiny because the rest of the paper rests on them. First, that synthetic patient turns approximate real human-AI conversation. Second, that a frozen GPT-4o judge reproduces clinician consensus on 94.2% of 6,751 eligible item comparisons, exceeding clinician-clinician agreement of 89.4% on 22,816 pairs. The paper further claims compliance with AEF-1 Minimum Operating Conditions for Independent Third-Party AI Evaluation, and uses that framing to justify withholding essentially all operational material.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is K-Bench and what does it measure in LLMs?
K-Bench is a clinician-calibrated benchmark that evaluates large language models in simulated multi-turn mental health conversations involving suicide, self-harm, domestic violence, substance misuse and no-risk presentations. Each of 125 model configurations (33 base models, 14 providers) is run on the same 200 vignettes in conversations of up to 20 turns and scored by a frozen GPT-4o judge against a 47-item, 7-dimension clinical rubric covering risk recognition, risk exploration, ethical reasoning, supportive conversation, psychological knowledge, cultural competence, and autonomy and boundaries.
How reliable is the automated judge used to score K-Bench?
The frozen GPT-4o judge matched clinician consensus exactly on 94.2% of 6,751 eligible item comparisons, versus 89.4% exact agreement between clinicians themselves on 22,816 paired comparisons. That figure is measured on the same transcripts used to calibrate the judge, a reserved holdout is mentioned but no holdout result is reported, and no chance-corrected statistic such as kappa is given, so it should be read as a consistency check rather than proof of clinical validity.
Do therapeutic prompts or more reasoning make LLMs safer in mental health conversations?
Not reliably. Across 61 matched default-versus-therapeutic prompt pairs the mean change in combined-risk score was +0.11 points (improved in 24 pairs, reduced in 37); in a stricter paired analysis the mean effect was +0.78 points for no/low-reasoning baselines (95% CI -0.54 to +2.61) and -0.25 for elevated reasoning (95% CI -1.17 to +0.65). Elevated reasoning improved 9 of 30 matched pairs and reduced 21, with a mean of -0.58 points. Both effects depend on the specific model and reasoning setting.
Can K-Bench results be reproduced or audited independently?
No. The vignettes, patient-agent prompts, disclosure schedules, judge prompts, transcripts and raw clinician ratings are all withheld by design to prevent benchmark contamination, and the code for generation, scoring and evaluation is not public. Only aggregate leaderboard results, configuration metadata, clinical constructs and methods are published at k-bench.ai, so external verification is limited to re-running new models through the authors' own hosted pipeline.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 could not be visually inspected; all statements about them rely on the figure captions and the surrounding text, including the claim that bootstrap intervals in Figure 2 and Figure 5 do not overlap for particular entries.
- Model identifiers in Supplementary Table 1 (for example openai/gpt-5.5, anthropic/claude-fable-5, stealth/ox-alpha) could not be checked against any external source and are reported exactly as the paper lists them.
- The full rubric wording, patient prompts, disclosure schedules and judge prompts are withheld by the authors, so construct validity and scoring correctness could not be examined directly.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.