AI / ML

NVIDIA has open-sourced Model-Optimizer, a unified library designed to compress and accelerate deep learning models for downstream deployment. The tool supports techniques such as quantization, pruning, distillation, and speculative decoding to optimize inference speed. It accepts Hugging Face, PyTorch, or ONNX models as input and targets frameworks like TensorRT-LLM and vLLM for execution.

• Unified library handles quantization, pruning, distillation, and speculative decoding in one place.
• Targets TensorRT-LLM, TensorRT, and vLLM to accelerate real-time inference workloads.
• Accepts standard inputs from Hugging Face, PyTorch, and ONNX formats directly.
• Reduces model size and latency, easing deployment constraints for large language models.
• Open-source documentation and roadmap available for integration planning.

This paper addresses uneven hallucination reduction in multimodal large language models (MLLMs) during reinforcement learning. The authors identify two failure points: hard queries causing advantage collapse and confident errors becoming gradient-invisible. They propose Dual-Entropy Enhanced Policy Optimization to correct these issues by stabilizing advantage estimates and ensuring weak gradients for high-confidence wrong predictions.

• Hard queries with high semantic entropy can collapse group-relative advantages to zero, increasing hallucination risk.
• Standard categorical policies fail to update confident-but-wrong tokens because their gradient norms vanish as distributions sharpen.
• Dual-Entropy optimization explicitly targets these weak points to improve correction chains from reward to parameter updates.
• This approach aims to make reasoning sharper without sacrificing accuracy on multimodal inputs.

Amazon EMR on EKS now supports interactive Apache Spark sessions via Spark Connect, allowing developers to run persistent Spark contexts from local IDEs like Jupyter or VS Code. This client-server architecture decouples the application client from the Spark driver, enabling seamless blending of local Python execution with remote Spark operations on EKS clusters. The feature integrates with Amazon SageMaker Unified Studio notebooks to streamline data engineering workflows.

• Developers can debug Spark apps interactively from local IDEs while Spark runs on EKS.
• Persistent Spark contexts span across cells and scripts for better workflow continuity.
• Spark Connect's client-server model separates local tooling from remote compute resources.
• Compatible with Amazon SageMaker Unified Studio and standard Python environments.

TW3Cast achieves third place on the GIFT-Eval time-series forecasting benchmark, outperforming most entries by relying on a frozen selection table rather than agentic reasoning or large language models. The system uses public foundation models like Chronos-2, TiRex, and Toto, which are lightly fine-tuned on cleaned and enriched training data. Selection for specific dataset, frequency, and horizon configurations is determined by a pre-computed table derived solely from the training split.

• Ranks 3rd of 130 on GIFT-Eval without using agents or LLMs for reasoning.
• Selection logic is a frozen table computed once on the training split.
• Experts are public models (Chronos-2, TiRex, Toto) with light fine-tuning.
• Approach avoids complex agentic workflows in favor of deterministic routing.
• Training data is cleaned and enriched before fine-tuning the foundation models.