Benchmark RadarAI BENCHMARK PROFILE
SABER-Math
SABER-Math evaluates information retrieval for mathematical queries, with about 283K problems and tasks for reranking based on fine-grained relevance. Scoring uses preference tournament ratings.
- Released
- 2026-06-29
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Existing IR benchmarks fail to capture mathematical relevance, and MTEB doesn't predict math performance. SABER-Math provides a math-specific benchmark to guide retriever selection.
Motivation
As agentic AI systems tackle more complex mathematical tasks, they increasingly rely on information retrieval (IR) to search problem databases, theorem libraries, and educational resources.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.