Technical Summary: CARE-MH

Problem Statement

Large language models (LLMs) are increasingly deployed to provide non-clinical mental health support, necessitating reliable evaluations of safety, empathy, and therapeutic appropriateness. However, the current landscape of mental health benchmarks suffers from significant limitations that hinder scientific progress. Existing benchmarks are difficult to reproduce and compare due to inconsistent evaluation designs, under-specified prompting configurations, and benchmark-specific metric definitions. Furthermore, the rapid evolution of LLMs and the deprecation of proprietary evaluator models introduce instability, making it challenging to determine whether performance differences arise from model capabilities or variations in evaluation pipelines. These issues prevent unified analysis across datasets, evaluator models, and evolving LLM generations.

Methodology: The CARE-MH Framework

To address these challenges, the authors present CARE-MH (Comparable And Reproducible Evaluation of non-clinical Mental Health LLMs), a unified framework designed to treat mental health LLM evaluation as a controlled, configurable experimental pipeline.

Design Principles

CARE-MH is built on three core principles:

1. Explicit Parameterization: Major evaluation choices (prompt templates, evaluator instructions, generation settings, metric definitions) are treated as explicit, configurable variables rather than hidden implementation details.
2. Factorized Evaluation: The evaluation process is decomposed into distinct stages to isolate and analyze different sources of variability independently.
3. Semantic Preservation: Benchmark-specific evaluation criteria are preserved while organizing related metrics under a shared taxonomy to enable cross-benchmark comparison.

Core Components

The framework formalizes evaluation into five configurable components:

• Benchmark Datasets: Collections of non-clinical mental health prompts and conversation excerpts.
• Metric Definitions: Explicit criteria for evaluating System Under Test (SUT) responses, including instructions, output formats, and scoring scales.
• Systems Under Test (SUT): The LLMs being evaluated.
• Evaluators: LLMs used to assess SUT responses based on specific metrics.
• Configurations: Separate specifications for prompting (formatting inputs/instructions) and generation (decoding hyperparameters like temperature and max tokens).

Unified Metric Taxonomy

The authors consolidate overlapping but inconsistently defined metrics from prior benchmarks into a unified taxonomy comprising six major categories:

1. Therapeutic Communication: Empathy, active listening, non-judgment, and encouragement.
2. Content Quality & Problem Fit: Relevance, specificity, informativeness, and holistic coverage.
3. Actionability: Guidance, structure, and next steps.
4. Safety, Ethics & Scope: Harm avoidance, toxicity, ethical boundaries, and unlicensed medical advice.
5. Trustworthiness & Correctness: Factual consistency and composite reliability.
6. Evaluation Artifact: Rationales and evidence spans supporting ratings.

Experimental Evaluation

The authors applied CARE-MH to reproduce and analyze three representative state-of-the-art English benchmarks: CounselBench, MentalBench, and MentalChat16K. The study involved:

• Reproducing original benchmark results where configurations were accessible.
• Substituting deprecated models (e.g., GPT-4 Turbo, Gemini Pro) with newer generations (e.g., GPT-4.1, Gemini-2.5-Pro) to assess stability.
• Conducting cross-benchmark evaluations to test consistency.
• Applying a unified evaluation design with standardized metrics across all datasets.

Key Findings

1. Reproducibility Depends on Model Stability

Reproducibility varies significantly based on the stability of both the SUTs and the evaluators.

• CounselBench demonstrated strong reproducibility, with absolute differences between original and reproduced results generally below 0.05, even when using newer evaluators like GPT-5 and Llama-3.3.
• MentalBench showed high fidelity when the original evaluator (GPT-4o) was reproduced but exhibited large deviations (exceeding 1.0 in some metrics) when newer or substituted SUTs were used, suggesting behavioral differences in model generations.
• MentalChat16K could not be fully reproduced due to the deprecation of its proprietary evaluators. Replacing these with newer LLM generations resulted in substantial score shifts across nearly all metrics, particularly in subjective dimensions like empathy and active listening, where newer evaluators assigned consistently lower scores.

2. Cross-Benchmark Disagreement Stems from Metric Definitions

While overall model rankings were largely consistent across benchmarks, significant inconsistencies arose when comparing semantically similar but differently defined metrics.

• Same Metric Consistency: When the exact same metric (e.g., "Empathy") was applied across benchmarks, results were highly consistent.
• Different Definition Inconsistency: When comparing analogous metrics with different operational definitions (e.g., CounselBench "Specificity" vs. MentalBench "Relevance" vs. MentalChat16K "Active Listening"), ranking shifts occurred. This indicates that cross-benchmark disagreement is driven primarily by differences in metric definitions and evaluator instructions rather than inherent model variability.

3. Unified Evaluation Improves Comparability

Applying the unified CARE-MH taxonomy with standardized prompts and metrics across all datasets improved comparability. Both closed-source (Gemini-2.5-Pro) and open-source (Llama-3.3) evaluators produced consistent global trends, identifying advanced models (e.g., Gemini-2.5-Pro, GPT-4.1) as top performers and weaker models (e.g., LLaMA2) as bottom performers. This unified approach allows for direct comparison across dimensions previously fragmented across different benchmarks.

Significance and Contributions

The paper makes the following contributions to the field of mental health AI evaluation:

• Framework Development: Introduces CARE-MH, the first unified framework for comparable and reproducible evaluation of non-clinical mental health LLMs.
• Taxonomy Creation: Develops a unified taxonomy of mental health evaluation metrics that consolidates definitions from existing benchmarks.
• Systematic Analysis: Conducts the first systematic study of reproducibility and cross-benchmark consistency in this domain.
• Empirical Evidence: Reveals that evaluator/model stability and metric definitions are the primary drivers of benchmark disagreement and reproducibility variation.
• Demonstration of Utility: Shows that unified evaluation with standardized metrics significantly improves cross-benchmark comparability and reproducibility across evolving LLM generations.

The authors conclude that future mental health benchmarks must release standardized evaluation configurations alongside datasets. This includes explicit metric definitions, evaluator prompts, and generation settings to ensure long-term reproducibility and reliable comparison as LLM technology continues to evolve. The work emphasizes that without such standardization, claims regarding model safety and therapeutic quality may remain inconsistent or misleading.