A distributed MAS training framework uses consensus ADMM to enable role-local learning with decentralized communication, preserving centralized optimization benefits while keeping data local; it improves robustness to compromised agents and achieves similar reasoning gains as centralized training.
Google says GKE Dynamic Slicing can form TPU slices on demand by connecting or dividing pre-provisioned TPU cubes. The feature avoids rebuilding nodepools and can cut job startup time by up to 5× and recovery time by up to 4.5×.
TEX introduces a collaborative, execution-based cross-validation method for test-time scaling of language models, achieving state-of-the-art performance and showing benefits from collaborative execution even with random selections.
LuisCore provides runtime infrastructure for multi-agent inference, while OpenAI Swarm is an educational framework for agents that pass work to one another. The text presents them as complementary: LuisCore adds federation, identity, and audit to the handoff pattern.
Translated title. Original: シフト自動作成を3方式で作り直して分かった、貪欲法が破り続けていたルール
The article documents a progression from using an LLM to generate monthly shifts, through a greedy fill method, to a MIP solver, highlighting validation and correction as essential since LLM outputs do not guarantee rule satisfaction. It also notes practical implementation details, performance observations, and tooling choices to run the solver in Node.js.
FedAbA introduces a basis-consistent federated LoRA aggregation method for heterogeneous ranks, aligning client local bases to a global reference, reconstructing updates before aggregation, and refactoring the result with SVD and gauge fixing. Experiments show improved performance over baselines, especially with strong rank heterogeneity.
Block-split turbo LoRA for MiniMax H3 hybrids uses blocks 0-24 with a token refiner; blocks 25-49 are removed. It credits original authors, references specific source models, and outlines pairing with Ref2VA variants while advising not to stack with a different turbo variant. Start at 1.0 with 8 steps.
SageMaker Inference adds MIG-based GPU partitioning to let multiple models share a single GPU with hardware isolation, enabling smaller, right-sized partitions and potentially higher utilization without sacrificing isolation or predictability. It simplifies deployment by handling GPU configuration and placement on managed endpoints, reducing the need for manual Kubernetes setups.
Re-enables two split-scan cumsum tests on Navi3 after fixing a race in Triton's decoupled-lookback scan; replaces a prior workaround. Validation on gfx1100 shows improved results with Triton 3.9.0+gitfc8d31a5; confirms tests pass on the targeted hardware.
The paper presents a method that shares computation across model depths and uses a gate to decide when more layers are needed. On seven decision benchmarks, it matches or outperforms independently trained models at each depth while improving the balance between accuracy and efficiency.