• Long video understanding doesn't need to watch every frame. OmniAgent models perception as a decision the model makes for itself; its 7B agent hits 50.5% on LVBench, beating Qwen2.5-VL-72B at 10x the size.

• The bottleneck in multimodal-as-policy is memory, not decisions. RNG-Bench isolates "rebuild what's no longer visible and act on it" in two games, and finds frontier models mostly fail by forgetting earlier observations.

• Uniform diffusion language models finally have a serious open baseline. Sumi is the first at-scale (7B, 1.5T tokens) fully open uniform diffusion model trained from scratch, with weights, recipe, and data mix all released.

• Every reasoning step from an AI scientist now leaves auditable evidence. Xcientist externalizes literature, ideas, plans, and ablations into contract-governed artifacts, and names "claim drift" as a failure you can't catch by looking at outputs alone.

• User simulators are shifting the target from "match the line" to "pass as a person." Turing-RL swaps similarity matching for a Turing-test discriminator reward, and beats matching baselines in chat and forum settings.

Also Notable

• To Use a Video Model as a World Model, First Check Whether It Understands Physics — Physics-IQ pulls physical understanding out of generation quality and quantifies it on its own.
• Materials Foundation Models Adapt to New Systems via Sparsity-Promoting Fine-Tuning — robust, interpretable calibration for machine-learning interatomic potentials transferring to new domains, accepted at ICLR.
• Home Assistants Ignore How Instructions Get More Elliptical as a Conversation Goes On — PEC-Home handles this progressive, context-accumulating ellipsis in instruction interpretation, accepted at ACL.
• Idioms Transfer Poorly Across Languages Because of Non-Compositionality and Weak Surface Grounding — G-IdiomAlign uses English glosses as anchors for a cross-lingual alignment benchmark, accepted at ACL.

Read the full edition →