AWS published a walkthrough building an AI product tagging system using Amazon SageMaker serverless model customization. It fine-tunes Qwen3-8B with supervised fine-tuning (SFT) and then reinforcement learning with verifiable rewards (RLVR) via GRPO on an Amazon sales dataset, raising the overall tagging score from 0.354 to 0.6941. The approach targets high-volume retail catalog enrichment where a stable schema and programmatically scorable outputs favor a small custom model over a frontier LLM.
A new arXiv paper presents DPara, a parallel speculative decoding framework that eliminates the probabilistic fallback to serial drafting in existing PSD methods. It precomputes draft representations for every acceptance boundary during verification, then uses a lightweight autoregressive head to emit next-round draft tokens almost instantly. Tested on Qwen3-8B and Qwen3-14B, it achieves average speedups of 3.21x and 3.52x over autoregressive decoding.
A new arXiv paper presents DPara, a parallel speculative decoding framework that eliminates the probabilistic fallback to serial drafting in existing PSD methods. It precomputes draft representations for every acceptance boundary during verification, then uses a lightweight autoregressive head to emit next-round draft tokens almost instantly. Tested on Qwen3-8B and Qwen3-14B, it achieves average speedups of 3.21x and 3.52x over autoregressive decoding.
AWS engineers fine-tuned a Qwen3.6-27B search agent using Amazon SageMaker AI multi-turn reinforcement learning (MTRL), optimizing full agent trajectories with nDCG@10 as the trajectory-level reward. The fine-tuned model improved retrieval quality on three of four held-out benchmarks and sharply cut failure rates, matching small-model cost with frontier-model-like reliability. The writeup serves as a practical recipe for enterprises training agents on proprietary tools.
AWS engineers fine-tuned a Qwen3.6-27B search agent using Amazon SageMaker AI multi-turn reinforcement learning (MTRL), optimizing full agent trajectories with nDCG@10 as the trajectory-level reward. The fine-tuned model improved retrieval quality on three of four held-out benchmarks and sharply cut failure rates, matching small-model cost with frontier-model-like reliability. The writeup serves as a practical recipe for enterprises training agents on proprietary tools.
Amazon SageMaker Python SDK v3 now exposes generative AI inference recommendations directly in notebook workflows, helping developers benchmark endpoints and evaluate instance configurations. The recommendations are also available through the SageMaker AI UI, streamlining deployment optimization for LLMs.
Amazon SageMaker Python SDK v3 now exposes generative AI inference recommendations directly in notebook workflows, helping developers benchmark endpoints and evaluate instance configurations. The recommendations are also available through the SageMaker AI UI, streamlining deployment optimization for LLMs.
A new arXiv paper proposes Rubric Dropout, a method to mitigate reward hacking in rubric-as-reward reinforcement learning. The authors train Qwen3-8B with GRPO on medical and science rubrics and find that the training judge and a stronger gold judge diverge, indicating exploitation of the rubric proxy. Rubric Dropout randomly omits criteria during training to reduce overfitting to the fixed rubric.
A new arXiv paper proposes Rubric Dropout, a method to mitigate reward hacking in rubric-as-reward reinforcement learning. The authors train Qwen3-8B with GRPO on medical and science rubrics and find that the training judge and a stronger gold judge diverge, indicating exploitation of the rubric proxy. Rubric Dropout randomly omits criteria during training to reduce overfitting to the fixed rubric.
An AWS Machine Learning blog post lays out an eight-step 'customization spectrum' on Amazon Bedrock, ranging from prompt engineering and RAG to fine-tuning, continued pre-training, and fully custom models via Amazon Nova Forge. It gives teams escalation criteria, data requirements, and cost tradeoffs so they avoid both premature fine-tuning and stalled prompt engineering. Most workloads, per AWS, never need to go past grounding with RAG.
An AWS Machine Learning blog post lays out an eight-step 'customization spectrum' on Amazon Bedrock, ranging from prompt engineering and RAG to fine-tuning, continued pre-training, and fully custom models via Amazon Nova Forge. It gives teams escalation criteria, data requirements, and cost tradeoffs so they avoid both premature fine-tuning and stalled prompt engineering. Most workloads, per AWS, never need to go past grounding with RAG.