AWS published a walkthrough showing how to run the open-source SkyRL RL framework on SageMaker HyperPod to post-train Qwen3-VL-8B for multi-turn visual maze navigation. GRPO training raises maze solve rate from 43.75% to over 95% on a fixed 64-maze eval set. It matters to practitioners because it shows a reproducible recipe for multi-node RL post-training with colocation, checkpointing, and managed monitoring.
AWS published an engineering guide for scaling MoE reinforcement-learning post-training (RLHF/GRPO) on Amazon EKS, combining EFA networking, DeepEP expert-parallel communication, and EC2 Spot instances for rollout generation. Amazon's libfabric contributions to DeepEP enable native EFA support, delivering 40% higher aggregate rollout throughput on 48 P5en instances. The work targets the communication bottleneck that increasingly dominates sparse MoE training as expert parallelism spans more nodes.
AWS published an engineering guide for scaling MoE reinforcement-learning post-training (RLHF/GRPO) on Amazon EKS, combining EFA networking, DeepEP expert-parallel communication, and EC2 Spot instances for rollout generation. Amazon's libfabric contributions to DeepEP enable native EFA support, delivering 40% higher aggregate rollout throughput on 48 P5en instances. The work targets the communication bottleneck that increasingly dominates sparse MoE training as expert parallelism spans more nodes.
A new arXiv paper introduces Video-HopChain, a dataset of 22,550 multi-hop video questions over 13,378 videos plus a 1,000-question held-out benchmark, built to expose chained visual-evidence errors that standard RLVR data rarely tests. Fine-tuning Qwen3-VL-8B with GRPO on it, then adding a proposed Confidence-Gated Exploration method, raises the mean over eight video benchmarks from 55.4 to 59.3. Dataset, checkpoint, and training code are released, giving practitioners a recipe for multi-hop video reasoning data.
A new arXiv paper introduces Video-HopChain, a dataset of 22,550 multi-hop video questions over 13,378 videos plus a 1,000-question held-out benchmark, built to expose chained visual-evidence errors that standard RLVR data rarely tests. Fine-tuning Qwen3-VL-8B with GRPO on it, then adding a proposed Confidence-Gated Exploration method, raises the mean over eight video benchmarks from 55.4 to 59.3. Dataset, checkpoint, and training code are released, giving practitioners a recipe for multi-hop video reasoning data.
A voice agent stack team surveyed 113 TTS evaluation papers and organized them into a taxonomy of roughly 50 evaluation dimensions while choosing CI metrics. They found several dimensions have almost no literature coverage and drew practical conclusions for production evals. The findings matter to practitioners wiring TTS quality metrics into automated pipelines.
A voice agent stack team surveyed 113 TTS evaluation papers and organized them into a taxonomy of roughly 50 evaluation dimensions while choosing CI metrics. They found several dimensions have almost no literature coverage and drew practical conclusions for production evals. The findings matter to practitioners wiring TTS quality metrics into automated pipelines.
Researchers propose LoRA Scaffolded Policy Optimization (LSPO), a sampling-time mechanism that recovers the lost reinforcement-learning gradient on 'cliff' prompts where all sampled rollouts fail, making GRPO's group-normalized advantage zero. LSPO detects such prompts, fits a small LoRA adapter via a brief supervised step on ground-truth solutions, and uses it to scaffold sampling, enabling gradient flow on frontier capability prompts.
Researchers propose LoRA Scaffolded Policy Optimization (LSPO), a sampling-time mechanism that recovers the lost reinforcement-learning gradient on 'cliff' prompts where all sampled rollouts fail, making GRPO's group-normalized advantage zero. LSPO detects such prompts, fits a small LoRA adapter via a brief supervised step on ground-truth solutions, and uses it to scaffold sampling, enabling gradient flow on frontier capability prompts.
AWS announced new Ray capabilities on SageMaker HyperPod, integrating the open-source Ray framework with HyperPod's purpose-built infrastructure for foundation model training and serving. The integration uses KubeRay to manage Ray clusters on Kubernetes, enabling distributed Python workloads such as Ray Train and Ray Serve.
AWS announced new Ray capabilities on SageMaker HyperPod, integrating the open-source Ray framework with HyperPod's purpose-built infrastructure for foundation model training and serving. The integration uses KubeRay to manage Ray clusters on Kubernetes, enabling distributed Python workloads such as Ray Train and Ray Serve.