• A 1T Relay-Tree Refit Takes 150 Seconds at a 3% Change Rate. The paper measured 87.5 minutes for a complete 1T checkpoint transfer between two AWS regions. NeMo-DCR replaces full transfers with bit-exact delta sync while handling shard placement, version consistency, and retries.

• ALIVE Makes Inserted Objects Participate in Video Actions. Without VLM guidance, it improves the Overall score over the strongest evaluated baseline by 43.9% on ALIVE-interaction and 4.4% on the general video object-insertion benchmark.

• HERMES Turns Repository State Into Local Agent Interfaces. Its dependency-activated development primitives beat matched baselines by 12.4% on average across four benchmarks and show the cost potential of combining large and small models.

Also Notable

• Static Judges Become Stale as Policies Expose New Failures; VeriFine Co-Evolves Policies, Curricula, and Verifiers. When progress plateaus and verification becomes a bottleneck, it selectively requests human guidance on informative failure cases. Coactive calibration resolves disagreements, while driving and robot-navigation experiments show continued gains in policy and judge capability.
• AI Can Improve Vineyard Patrol Allocation Yet Still Produce Operationally Incomplete Plans. A 140-hectare field study found unreliable internal 2024 estimates caused by synthetic oversampling before train/test splitting. It evaluated forecasts against independent 2025 scouting. A separate sampling plan omitted instructions for replacing missing vines.
• WASD Matches Teacher and Student Distributions by Token Meaning, Not Only Vocabulary Position. It uses Sinkhorn divergence to control cost, needs no additional network, and covers instruction following, mathematical reasoning, and code generation.
• Medical Knowledge-Graph Reasoning Can Balance Evidence Budgets, Source Quality, and Citation Integrity. BAR raises citation precision from 59.8% to 77.9% while consuming only 62%–65% of its budget cap. It was evaluated across eight diseases and three prediction horizons on MIMIC-III and MIMIC-IV.
• Models Invent Group Correlations and Causal Stories From Random Data More Readily Than Humans. Sparse autoencoder analysis links this illusion to frequency-sensitive features across three tasks adapted from classic psychology experiments.
• ThinkFuse Selectively Triggers Auxiliary Fusion at Points Flagged as Unstable by Local and Trajectory-Level Uncertainty Signals. The training-free method requires fewer fusion triggers, generates fewer tokens, and remains effective with a smaller primary model.
• OpenWAM Makes World-Action Model Designs Comparable on a Shared Foundation. Starting from Wan2.2-5B, it builds a common causal robot-video foundation through pretraining on over 10,000 hours of video. It supports joint, video-then-action, action-then-video, and decoupled generation.
• EmbodiedSmith Lets Generated Tasks Demand Missing Assets and Physical Conditions From Their Scenes. Its recursive improvement loop expands long-horizon, dexterous-hand, and fluid simulation data through one pipeline for assets, scenes, and tasks.
• Continuous Retraining Can Drift When Models See Labels Only for Accepted Cases. The paper provides a bounded correction with a worst-case objective. Historical accepted-case data can progressively tighten confidence intervals under a sensitivity assumption.
• FoG Uses Deep Knowledge-Graph Feedback to Guide Earlier Path Selection. Reverse feedback and a compact memory subgraph preserve evidence that local pruning might discard while reducing model calls and token use.
• SquidAgent Treats Parallelism as an Explicit Cost Calculation. Splitting work pays only when critical-path savings exceed the cost of rebuilding context and aligning results. Session forking and shared conventions reduce both overheads.

Read the full edition →