The strongest work treats context handling as an engineering control. qMLX cuts repeated long-context prefill on one Mac Studio, while ContextOps checks payload structure before inference.
Robot learning is concentrating on practical bottlenecks inside existing policy pipelines. Failed rollouts become supervision, latent actions are cleaned of visual confounders, and action trajectories gain explicit…
The day’s strongest work treats large language model (LLM) coding as a controlled engineering process. ReProAgent and TestAgent tie repository context to runtime feedback, while DualVeri pairs machine-checked proofs…
Robot learning work in this period concentrates on making existing policies more dependable under perturbations and sparse feedback. Harness VLA adds planning and retries around a frozen controller.
Agent performance is increasingly determined by the executable control layer around the model. TTHE improves fixed large language models (LLMs) by editing their harnesses from unlabeled traces, while…
The day’s evidence is practical: coding agents need safer execution boundaries, better task inputs, and lower-cost repeatability.
The day’s robot papers put physical detail inside policy pipelines. Vision-language-action (VLA) models add 3D structure, reusable demonstrations, cached action chunks, and explicit handoff checks.
The day’s research treats coding agents as systems that need external checks during work. Aria shows verifier-gated proof search at unusual scale; SWE-Review adds repository-aware review; TraceProbe measures how a run…
The day is dominated by robot-manipulation work that makes policy internals more explicit: future states, latent actions, camera pose, and subtask memory.
The day’s strongest evidence treats coding agents as repository actors that train, act, and fail inside real workflows.
Robot research this week puts vision-language-action (VLA) policies inside real execution constraints. The strongest evidence comes from models that predict action-relevant change, manage long rollouts, and keep serving…
This week’s evidence treats coding agents as production systems. The strongest work measured review burden, token spend, and privilege control after agents leave single-task demos.