MedDialBench: Evaluating LLM Diagnostic Robustness Against Adversarial Patient Behaviors

MedDialBench is a controlled factorial benchmark for evaluating large language model diagnostic robustness under parametric adversarial patient behaviors. It is an anonymous submission to the NeurIPS 2026 Datasets and Benchmarks Track, with a companion paper under double-blind review. The dataset was last updated on May 6, 2026.

Use Cases

• Benchmarking LLM diagnostic accuracy based on adversarial patient conversation scenarios.
• Evaluating model robustness to parametric variations in patient behavior described in the benchmark.
• Training and testing clinical dialogue agents on controlled, factorial test conditions.

Strengths

• Designed as a controlled factorial benchmark for systematic evaluation.
• Focuses on a specific, high-stakes application area: diagnostic robustness in medical dialogue.
• Associated with a peer-reviewed submission to a major conference (NeurIPS 2026).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• The dataset is currently anonymized, and full provenance details are withheld pending publication.

Provenance

Source

Anonymous submission by anon-meddial-2026.

Collection Method

Likely contains synthetically generated or curated medical dialogue scenarios designed for benchmark evaluation.

Freshness

Last updated 2026-05-06 09:59:44; freshness should be verified.