Care RAG Eval

care-rag-eval

CARE-RAG - Clinical Assessment and Reasoning in RAG — Potluri et al. (2025) (arXiv:2511.15994, 2025)

What this evaluates

Evaluates whether LLMs using retrieval-augmented generation can faithfully apply clinical guidelines (specifically Written Exposure Therapy) to answer questions. It probes context fidelity, reasoning complexity, and question type to measure if models actually ground their inferences in retrieved evidence rather than relying on parametric knowledge or guessing.

Datasets

• CARE-RAG WET Guidelines — total ?; splits: test (-1)

Metrics

• Accuracy Score — range: [0, 1]
  • Fraction of model answers that exactly match the gold-standard correct answer across multiple-choice and yes/no questions.
• Cosine Similarity — range: [0, 1]
  • Semantic similarity between the model's open-ended output and the extracted RAG context, typically computed via embedding dot product normalized by norms.
• Inference Score (primary) — range: [0, 1]
  • Confidence score (0–1) from a judge LLM assessing whether the model’s generated reasoning is logically supported by the retrieved context.

Input / output format

Input: Clinical questions (multiple-choice, yes/no, or open-ended) paired with retrieved context passages from Written Exposure Therapy guidelines (which may be relevant, noisy, or misleading).

Output: Model-generated responses: selected option for MCQ/Yes-No, or free-text explanation/answer for open-ended questions.

Scoring recipe

def compute_metrics(predictions, golds, retrieved_contexts, judge_llm):
    # Accuracy for MCQ/Yes-No
    acc = sum(1 for p, g in zip(predictions, golds) if p == g) / len(golds)
    # Cosine Similarity for open-ended
    cos_sim = cosine_similarity(embed(predictions), embed(retrieved_contexts))
    # Inference Score via judge LLM
    inf_scores = [judge_llm.confidence(p, ctx) for p, ctx in zip(predictions, retrieved_contexts)]
    return acc, cos_sim, sum(inf_scores)/len(inf_scores)

Common pitfalls

• High accuracy on multiple-choice questions does not guarantee the model actually used the retrieved context; it may rely on pre-training or pattern matching.
• Inference scores depend on a judge LLM, which can introduce evaluation bias or fail to detect subtle clinical reasoning errors.
• Cosine similarity measures semantic overlap with context but does not verify clinical correctness or logical entailment.

Evidence (verbatim from paper)

The entailment (inference) score is computed by measuring the logical consistency between a model’s generated reasoning and the retrieved context (values range from 0 to 1, with higher scores indicating stronger support). The accuracy score is calculated as the fraction of model answers that exactly match the gold-standard correct answer.

Citation

@misc{potluri2025care-rag,
  title={CARE-RAG - Clinical Assessment and Reasoning in RAG},
  author={Potluri et al. (2025)},
  year={2025},
  note={arXiv:2511.15994}
}

• arXiv: 2511.15994