Technical Summary: K-Bench

Problem Statement

Large language models (LLMs) are increasingly utilized for mental health support, yet their safety in high-risk, evolving conversations remains poorly characterized. Existing benchmarks often rely on isolated prompts or single-turn classifications, failing to capture the gradual disclosure of risk, co-occurring psychosocial stressors (e.g., suicide, self-harm, domestic violence, substance misuse), and the necessity for sustained, multi-turn therapeutic engagement. Furthermore, many current evaluation frameworks make operational test materials publicly available, creating a risk of "benchmark gaming" where models are optimized for specific known scenarios rather than general clinical safety. There is a critical need for a clinically calibrated, protected benchmark that evaluates models across a broad configuration space while maintaining the integrity of independent prospective testing.

Methodology

The authors developed K-Bench, a clinician-calibrated benchmark designed to evaluate LLMs in high-risk mental health conversations. The methodology consists of five core components:

1. Vignette Framework: The benchmark utilizes a fixed cohort of 200 multi-turn vignettes (up to 20 turns) involving suicide, self-harm, domestic violence, substance misuse, and no-risk presentations. These vignettes are derived from lived-experience narratives and expanded via a 122-variable factorial design. This design systematically varies mental health presentations, cognitive coherence, risk history, protective factors, and disclosure styles, allowing risks to emerge indirectly, escalate, or co-occur.
2. Synthetic Patient Agents: To ensure scalability and consistency, the benchmark employs synthetic patient agents. The realism of these agents was validated by comparing their embedding-space distributions (using nearest-neighbour cosine distance, maximum mean discrepancy, and Fréchet distance) against a de-identified corpus of 50 real human–AI conversations. GPT-4o was selected as the fixed patient agent for the public evaluation due to its comparable realism and lower cost.
3. Clinician-Calibrated Rubric: A rating instrument comprising seven dimensions and 47 rubric components was developed through an iterative process involving a stakeholder panel of clinicians, lived-experience experts, and AI researchers. The dimensions include:
   • D1: Clinical judgement and risk awareness.
   • D2: Risk exploration.
   • D3–D7: Ethical reasoning, supportive conversation, psychological knowledge, cultural competence, and autonomy/boundaries.
     Six clinicians rated 151 transcripts, generating 16,157 consensus item-level ground-truth cells.
4. Automated Judge: A frozen GPT-4o judge was calibrated against the clinician consensus. The judge was trained to apply the rubric consistently across all 151 transcripts, achieving 94.2% exact agreement with clinician consensus across 6,751 eligible item comparisons.
5. Evaluation Protocol: The benchmark evaluates 125 model–configuration entries representing 33 base models from 14 providers. Evaluations cover default and therapeutic prompts, various reasoning settings, and inference costs. Crucially, the operational test materials (prompts, disclosure schedules, scoring examples, and transcripts) are protected to prevent direct optimization by model developers.

Key Contributions

• Scale and Breadth: K-Bench evaluates a significantly larger set of configurations (125 entries, 33 base models, 14 providers) than previous benchmarks like VERA-MH or SIM-VAIL.
• Clinical Depth: Unlike single-turn tests, K-Bench assesses risk recognition, exploration, and management across sustained interactions where risks may escalate or co-occur.
• Protected Architecture: By withholding operational materials while publishing methods and aggregate results, K-Bench adheres to AEF-1 standards for independent third-party evaluation, ensuring the leaderboard remains a valid measure of general capability rather than test-set memorization.
• Validated Synthetic Data: The paper provides empirical evidence that synthetic patient interactions closely approximate real human–AI conversations, establishing a scalable basis for future evaluation.

Results

• Model Performance: Leading models achieved overall scores above 95, with the top performer (OpenAI GPT-5.5) scoring 98.96 overall and 95.51 on the combined-risk metric. However, significant variation existed in risk exploration (D2), which had a mean score of 79.86 and a range of 22.1–93.1, distinguishing models that gather necessary clinical information from those that merely generate supportive language.
• Risk Sensitivity: Performance generally remained high across severity levels, though mean scores decreased slightly for imminent-risk cases and vignettes with multiple concurrent risks. The benchmark successfully identified configurations that became less reliable under complex clinical presentations.
• Prompt and Reasoning Effects:
  • Therapeutic Prompting: Effects were configuration-specific. While therapeutic prompts improved combined-risk scores for weaker models (e.g., IBM Granite 4.0 H Micro saw a +17.91 point gain), they offered mixed or negligible benefits for top-tier models.
  • Reasoning Settings: Elevated reasoning did not produce a uniform improvement in safety. In matched comparisons, elevated reasoning improved scores in some configurations but reduced them in others, indicating that increased deliberation does not automatically translate to safer clinical behavior.
• Cost vs. Performance: Strong risk performance was available across a broad price range, with several lower-cost configurations approaching the performance of significantly more expensive systems.

Significance and Claims

The paper claims that K-Bench establishes a new standard for evaluating conversational AI in mental health by combining broad clinical coverage, validated synthetic conversations, and a protected test architecture. The authors assert that the benchmark demonstrates that contemporary models can perform remarkably well in high-risk scenarios when allowed to engage in sustained interaction, yet it also exposes persistent failures in risk exploration among weaker systems.

The significance of K-Bench lies in its ability to provide a clinically grounded, reproducible system for continuous evaluation at a scale that clinician-only assessment cannot sustain. By separating risk exploration from supportive conversation, the benchmark reveals safety-relevant weaknesses that might be masked by high overall scores. The authors position the public leaderboard not as a final verdict on model safety, but as an ongoing accountability tool to track genuine progress, detect configuration-specific regressions, and highlight the specific clinical work required for safer deployment. The paper concludes that while leading models are suitable for low-risk support and information gathering, clinical deployment requires clear pathways for human intervention, as general-purpose chatbots cannot verify identity, contact services, or manage the consequences of intervention.