New Foundation-Model Research Targets Embodied Learning, Visual Intelligence and Decisions

Open-HumanScale, WeLM-Image, and Drex remain research leads, while TailSteer’s available post lacks enough results to assess. Google’s Muir linear-attention work reports 2.7× faster training at 32K context and 10M-token extrapolation from 4K training, but lacks independent replication; an Amazon agent-memory study reports 59.9% on a small eight-user evaluation, with generalization still open. EmbeddingGemma 2 and Memorilla warrant paper review, but the supplied items contain no methods or results.