1T Cross-Region Relay-Tree Refit in 150 Seconds at 3% Change Rate

Featured

01 Training: Model Distribution May Bottleneck Agentic RL

Agentic RL separates training from inference, so every policy update must travel to the rollout cluster. Transferring one full checkpoint for a 1T model between two AWS regions takes 87.5 minutes.

NeMo-DCR exploits a useful property of BF16 training: only about 1% of weights change their stored value per step. A fixed mapping projects updates from training shards into unified checkpoint coordinates. Residual transformations handle the rest, while the inference runtime's native loader performs final placement.

Compressible changes travel as XOR masks; other changes directly overwrite their targets. In-place updates make retries safe after partial failures. Joint commits ensure each new delta starts from the correct version. At a 3% change rate, cross-region refit takes 150 seconds through a relay tree. The paper reports 12–40× speedups over a transport-only full-checkpoint reference for 30B-to-1T models at 3% and 5% change rates. Large-scale Agentic RL now depends on version consistency, cross-cluster transfer, and failure recovery as well as training algorithms.

Key takeaways:

• Include per-round weight synchronization latency in throughput budgets for large-scale Agentic RL.
• Delta transfer needs correct shard placement, bit-exact updates, and safe retries to work in production.
• For cross-region training and inference, measure actual weight-change rates and test relay-tree transfers.

Source: NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter Scale

02 Video Gen: Insertion Is Easy; Interaction Is Hard

Adding an object to a video is not enough. The object must also be picked up, moved, or manipulated naturally.

ALIVE trains on 35,800 paired samples. Each action scene appears once with the target object and once without it. This teaches the editor how objects participate in actions while preserving the source video. A vision-language model generates interaction instructions automatically, removing the need for extra user descriptions. Those instructions raise the interaction score by 0.95 points.

Without VLM guidance, ALIVE 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. ALIVE's advantage is concentrated on integrating objects into actions, not video editing as a whole.

Key takeaways:

• Test whether inserted objects participate in actions, not merely whether they appear.
• Without VLM guidance, the Overall-score gain is 43.9% on ALIVE-interaction and 4.4% on the general video object-insertion benchmark.
• Automatic interaction instructions can reduce prompting work in interactive video products.

Source: ALIVE: Interaction-Aligned Object Insertion for First-Frame-Guided Video Editing

03 Code Intelligence: Better Agents Need Better Repository Interfaces

Long coding tasks force models to revisit code, configuration, tests, and dependencies. The missing piece may be an agent-friendly repository interface, not a larger context window.

HERMES packages repository objects as Dev-Primitives. Each component has a local model that understands its implementation and dependencies. The system activates these components only when required. Its diagnostic process maps execution evidence back to components that require revision.

The paper reports an average 12.4% gain over matched baselines across four benchmarks. When paired with strong activation and diagnosis models, Qwen3-8B Dev-Primitives stay within 4.5% of an all-GPT-5.6-Sol configuration across all four benchmarks. They also cut inference cost by 26.2% on Terminal-Bench 4.0. Cost definitions, scheduling overhead, and cross-repository generalization still need confirmation from the full paper.

Key takeaways:

• Repository interfaces may deserve more investment than simply expanding the context window.
• Strong models can handle activation and diagnosis while smaller models work on local components.
• Evaluate task success, total inference cost, and scheduling overhead together.

Source: Harness Engineering for Software Engineering via Modular Executable Dev-Primitives