AI BENCHMARK PROFILE

SeGaBench

General AICoding & Software Engineering

SeGaBench is an executable benchmark containing 120 cases (100 synthetic, 20 source-backed) to test whether LLMs can recover semantic optimization opportunities that compilers miss, with hidden enabling semantics, oracle artifacts, and validators.

Released: 2026-08-04
Readiness: Paper only
Primary field: General AI

Why it matters

Compilers miss profitable transformations when enabling semantics are absent from the program representation. SeGaBench evaluates whether LLMs can recover such semantics and produce validated, performance-improving artifacts.

Motivation

Optimizing compilers miss profitable transformations when their enabling semantics are absent from the analyzed program representation.

Primary resources

Paper ↗Hugging Face ↗

Benchmark Radar records only publicly supported details and links back to primary sources for verification.