AI training system reports faster updates on agent workloads
A 2026 preprint reports faster AI-agent training updates, lower memory use and stronger scaling for psRL across four benchmark workloads.
A 2026 preprint reports faster AI-agent training updates, lower memory use and stronger scaling for psRL across four benchmark workloads.
This preprint develops sound and complete logics for reachability in polyhedral spaces with reversible dynamics, while leaving non-invertible cases open.
A preprint finds weak links between automated metrics and human ratings, while an LLM judge scores AI-generated and human-generated stories differently.
A preprint reports SRPO benchmark scores for math and agents alongside substantially lower training compute than GRPO and scaled SFT.
An arXiv preprint reports higher model-based text-to-CAD scores with experience memory on a hard CADFusion test, but results varied elsewhere.
A preprint tests ShardMeter, a lightweight model for forecasting distributed AI training runtime, throughput, bottlenecks and cost.
An arXiv preprint tests a decentralized market for AI agents, finding its best benchmark scores with subcontracting but weak direct cost estimates.
Amazon SageMaker Feature Store now supports two new APIs: BatchWriteRecord writes up to 25 records across multiple feature groups in a single call, and ListRecords enumerates record identifiers within a feature group. In this post, we walk through each API...
A preprint tests a Dirichlet-mixture model for missing compositional data and mild outliers in simulations and American time-use responses.
GLM-5.3-Flash and Qwen3.8-Flash-Next converge on hybrid linear attention, sparse 2048-token indexers, and four residual streams.
A preprint reports that a task-routed video-language model averaged 62.9 across five COIN tasks, with lower compute and higher throughput in one profile.
A mathematical preprint proves conditional solution results for equations with time-dependent infinite delay and gives a variation-of-constants formula.