Meta's InteractiveWorld Benchmark Shows MLLMs Still Far from Reliable Physical Assistants
meta · hf · 2026-10-27
Meta and Duke University present InteractiveWorld, the first verifiable closed-loop environment where an MLLM assistant guides a simulated human agent through physical tasks. Its InteractiveBench benchmark has 120 multi-turn tasks scored by verifiable state-based rewards. Experiments with 12 frontier MLLMs show even the strongest, Claude Fable 5.1, reaches only 20% and 25% on Long-horizon Guidance and Planning.
More from Embodied
- Humanoid robots will change the world via boring factory tasks, not backflips — nikola_mr64990 · 2026-10-27
- iPhone Duo foldable displays cracking within hours or days, Bloomberg verifies complaints from 6+ customers — firstadopter · 2026-10-27
- Noetix Robotics, backed by CATL, reportedly files for Hong Kong IPO — pstAsiatech · 2026-10-27
- Ex-Microsoft exec Mikhail Parakhin slams Tesla FSD for repeating the same mistakes — MParakhin · 2026-10-27
- 8.3M words transcribed across nearly 5,000 conversations on MIRA smart glasses — AnhPhuNguyen1 · 2026-10-27
- Surface Laptop Ultra review: impressive local AI but bandwidth limited — draginol · 2026-10-27