optimization

Φ-Bench is a benchmark designed to evaluate large language models on their ability to engineer and optimize the LLM infrastructure stack, covering tasks from kernel optimization to end-to-end system design.

The article details optimization techniques for FlashAttention-4, including an S/P ping-pong method for decode to overlap operations, achieving up to 16% performance gain on NVIDIA Blackwell GPUs.

The Atlas model now runs in real-time after intense inference optimization, making world generation and navigation more interactive on both B200 and MI355X hardware.

The article explains that Python's dict and set data structures, often assumed to be O(1), can exhibit quadratic-time performance in scenarios involving hash collisions and memory hierarchy effects.

This article advises setting response budgets for spreadsheet AI agents to ensure efficient and complete data retrieval, focusing on bounding outputs and allowing contextual follow-ups.

The article shows a 247-byte JavaScript implementation of Minesweeper that runs in a browser, detailing the coding techniques used to achieve such compact size.

The article discusses alternatives to the official llama.cpp for optimized performance on Strix Halo hardware, with specific forks achieving significantly higher throughput for AI inference tasks.

The article recommends using GPT-6 Astra by breaking tasks into steps, starting with medium settings to replicate a demo and then using max settings to optimize results, leveraging subagents for efficiency.

Unsloth AI announces optimizations for GLM-5.3-Flash, enabling 1.6–3.4× faster local GGUF inference with multi-token prediction and hardware requirements for running models locally.

LeanGRPO eliminates redundant recomputation in diffusion RL methods by introducing recompute-free training schedules, achieving up to 1.83x speedup while preserving optimization objectives.

The paper proposes Tail-Likelihood Reinforcement Learning (TailRL), an optimization method that focuses on the upper tails of reward distributions to improve policy performance in generative tasks, demonstrated across various applications.

This paper adapts two plan deordering techniques from classical planning to hierarchical task network planning, showing a substantial reduction in ordering constraints while maintaining plan validity.

The paper proposes a graph-constrained agentic AI framework that uses large language models to adapt retail supply chain operations to evolving requirements, demonstrating improved correctness and end-to-end success over direct LLM reformulation in evaluations with retail partners.

The article argues that the `initial-scale=1.0` parameter in the viewport meta tag is unnecessary for modern web browsers and can be omitted to reduce HTML file size.

FlashKAN proposes a method to accelerate Kolmogorov-Arnold Networks by replacing Cox-de Boor recursion with truncated power form for B-spline evaluation, providing a fused GPU implementation and an open-source package.

Introduces Belief-Calibrated Optimization (BCO), a method that maintains a persistent world model to improve LLM agent performance by explicitly tracking and revising beliefs about environment responses to edits.

TaRA is a training-aware initialization method for Low-Rank Adaptation (LoRA) that improves gradient fidelity, leading to better fine-tuning performance for large language models across various tasks.

The author describes how a volatile request identifier in prompts disrupted prefix caching, causing increased costs and latency for high-volume users, and asks for validation methods for caching improvements.

QTEA is a sub-2-bit post-training quantization framework for large language models that achieves ternary weight compression with improved accuracy and hardware efficiency through sparse residual salient weights and column-wise optimization.

The paper proposes REAL-Q, a novel end-to-end aligned post-training quantization method for large language models that uses dynamic gradient descent to mitigate quantization errors, achieving significant improvements in KL divergence reduction over state-of-the-art methods.