Meta and Oxford's OverSeer lifts multi-agent world-model consistency on Minecraft to 60.2% from 0–12.5% baselines
armand_ruiz · x · 2026-10-29
Meta AI and Oxford introduce OverSeer, a multi-agent world model that augments each agent's egocentric stream with six fixed-pose auxiliary camera views acting as a shared world-centric reference, compressed via a δ-VAE encoding residuals against a static reference frame (864→108 tokens per frame, 8x reduction).
- On MAC-Bench (Minecraft), accuracy jumps from 0.0%/6.2%/12.5% baselines to 60.2%/37.9%/40.5% across three consistency tasks
- Transfers to Doom reunite tasks at 62.0% and 67.0%; validated with real multi-camera captures, not tied to game engines
- Six auxiliary views cost only 37.5% of the token budget of two full agent streams
- Frame-level metrics miss failures: a baseline landed within 3.1 FID of OverSeer while answering 0% of episodes correctly
Dataset on Hugging Face; project page public.
More from Models
- Mistral Large 4 launch-day interview: open weights, own data centers, excels at cyber — huggingface · 2026-10-29
- Mistral returns with Large 4: a 1T-param 49B-active multimodal MoE, 1M context — Xianbao_QIAN · 2026-10-29
- Gamma 5 takes #1 on Design Arena for slides, beating Claude Opus 5.5 — haroonc · 2026-10-29
- Mistral Open-Sources Trillion-Parameter Coding Model Le Chonk, Still Trails DeepSeek and GLM — The Decoder · 2026-10-29
- Mamba co-author credits SmolLM3's open recipe and checkpoints for enabling research — srush_nlp · 2026-10-29
- ThursdAI Recap: Gemini 4 Carbon, Genie 4, Anthropic IPO Rumors and More — WolframRvnwlf · 2026-10-29