Benchmark RadarAI BENCHMARK PROFILE
RNG-Bench
RNG-Bench evaluates multimodal LLMs in controllable non-Markov games, requiring reconstruction of past observations and acting on them, with two games: Matching Pairs and 3D Maze.
- Released
- 2026-06-17
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Addresses the gap in evaluating models' abilities to remember and act on hidden state, which is critical for real-world deployments where observations are partial.
Motivation
Deploying multimodal foundation models as closed-loop policies increasingly requires conditioning actions on observations that are no longer visible.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.