Also Worth Noting - 2026-07-09

Also Worth Noting

02 [Training] Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning Async RL pipelines for LLMs lose training stability because model updates run on rollouts generated by an already-outdated policy. Single-rollout updates with a staleness-aware correction close that gap: each rollout triggers an immediate update, and a correction term accounts for the policy drift that accumulated since the rollout was generated. This sidesteps the group-sampling assumption baked into GRPO, which breaks down when rollouts arrive out of order. Teams running long-horizon agentic RL pipelines should treat policy staleness as a first-class engineering constraint, not an acceptable background cost. link

03 [Inference] Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity Fixed state size is the hard ceiling that keeps linear attention models behind transformers on long-context recall, and simply enlarging the state raises FLOPs proportionally. Sparse Delta Memory breaks that ceiling by updating only a sparse subset of state slots per token, extending Gated DeltaNet with a sparse addressing scheme that scales hidden state capacity by orders of magnitude while keeping per-token compute flat. Recall performance improves without the FLOP penalty that previously made large-state linear RNNs impractical. For teams evaluating linear RNN alternatives to transformers, SDM removes the most common reason to dismiss them on long-context tasks. link

04 [Agent] Automating the Design of Embodied Agent Architectures Researcher intuition currently decides where perception feeds memory, how observations get processed, and how model calls connect in embodied agents, encoding bias into every architectural choice before a single experiment runs. Agent Architecture Search automates that design space using AgentCanvas, a typed-graph runtime that treats agent modules as composable nodes and searches their connections through simulator rollouts. Configurations found by AAS outperform hand-designed baselines on standard embodied tasks, with no domain expert required to specify the wiring. Teams building perception-heavy agents should treat architecture search as a viable alternative to intuition-driven module composition. link

05 [Application] Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning Most AI models for chemistry, biology, and materials science fit correlations between structure and property without ever encoding the physical rules that govern them. This approach instead interprets structural evidence through explicit constraints including symmetry, bonding geometry, and energetics, producing explanations that scientists can audit against domain knowledge rather than predictions they must accept on faith. The model operates across biology, chemistry, and materials science under a unified reasoning framework, which means failure modes are diagnosable rather than opaque. For practitioners in computational materials or drug discovery, auditability is the differentiating feature worth evaluating. link

06 [Eval] RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies Sim-only robot benchmark numbers are unreliable predictors of deployment success, and RoboDojo quantifies exactly why: the sim-to-real performance gap varies dramatically by task type, meaning a method that looks strong in simulation can fail in specific real-world conditions that sim never surfaces. RoboDojo pairs simulation and physical evaluation on the same task taxonomy for generalist manipulation policies, the first benchmark to do both under a unified framework. The gap is not uniform, and that non-uniformity is the finding. Teams reporting manipulation results on sim-only benchmarks should treat RoboDojo's paired evaluation as the new baseline for credible capability claims. link