New AI Benchmark GIM Tests Multi-Domain Cognitive Integration
Researchers have introduced the Grounded Integration Measure (GIM), a benchmark of 820 original problems designed to evaluate AI models on tasks requiring coordination across multiple cognitive domains. Unlike existing benchmarks that either escalate knowledge demands (GPQA, HLE) or remove knowledge entirely for abstract reasoning (ARC-AGI), GIM focuses on integration of constraint satisfaction, state tracking, epistemic vigilance, and audience calibration over broadly accessible knowledge. The benchmark comprises 615 public and 205 private problems, each authored by experts and scored using a rubric with a median of six independently judged criteria. The approach aims to avoid conflating memorization with capability or divorcing reasoning from practical contexts. The paper is available on arXiv under identifier 2605.18663.
Key facts
- GIM stands for Grounded Integration Measure
- Benchmark contains 820 original problems
- 615 problems are public, 205 are private
- Problems require coordinating multiple cognitive operations
- Operations include constraint satisfaction, state tracking, epistemic vigilance, audience calibration
- Knowledge used is broadly accessible, not specialized
- Each problem is expert-authored with rubric-decomposed scoring
- Median of 6 independently judged criteria per problem
Entities
Institutions
- arXiv