Benchmark RadarAI BENCHMARK PROFILE
SelectBench
SelectBench evaluates selective evidence adoption in retrieval-augmented language models, focusing on rejecting deceptive content. The benchmark includes a 325-example test set and rule- or judge-based reward scoring.
- Released
- 2026-07-22
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Retrieval-augmented models often face mixed contexts with misleading content. A standardized evaluation for selective evidence adoption helps measure safety and reliability in real-world deployments.
Motivation
Retrieval-augmented large language models frequently face contexts that interleave useful evidence with misleading statements or instruction-like content.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.