Frontier Capability, Efficient Inference & Long-Horizon Agent Evaluation

Research continues to test multimodal grounding, persistent agents, and inference efficiency. JevAgentHarness reports benefits from confidence-thresholded delegation, while naive delegation reduced task success by 19 points; Velox reports a 4.36× attention speedup on a stated world-model benchmark. These are paper-reported results, and broader independent evaluation is needed; other recent signals, including GracefulRL and long-context inference, also remain preliminary.