From worlds to grounding

Unlike other world models, it supports multi-turn free-form video interaction by separating persistent scene memory from local action commands.

It clusters tasks by shared capabilities, co-evolving simulators and reusable modules for 34-point gains on 28 benchmarks and 26 points on 18 unseen tasks.

FVQ uses symmetric parallel full-feature codebooks to avoid high-bit saturation, slightly surpassing continuous QwenViT.

Unlike most control repos, it fully reproduces SuperMimic on Unitree G1, matching 0.82 LAFAN1 in 1.66 hours on one RTX 5090.

Unlike black-box AI text detectors, it exposes named feature sums and token-level evidence while preserving accuracy above 0.7 interpretability on COLING and RAID.

Unlike other grounding repos, it exactly recovers greedy coordinate-token boxes in 5.6 passes, running 1.5× faster than vLLM on identical weights.

Unlike most attribution datasets, it provides 254,561 de-duplicated open passages labeled across 18 LLMs with provenance and leakage flags.

Unlike translation-only multilingual benchmarks, it fully localizes tasks, tools, and environments across 23 languages.

Matched masks across explicit/implicit wording and target counts isolate instruction-grounding failures across four domains.