optimization

The article presents fast algorithms for calculating the day-of-the-week from a day count, using bit manipulation and modulus techniques that outperform existing solutions, with implementations tuned for different platforms and use cases.

The author has updated their NInfer fork with rk2v4-e8 quantization for the KV cache, enabling up to 250-350K tokens context on a single RTX 4090 without system RAM spill and with optimizations for faster generation speeds.

Introduces DFlash 2, a block-diffusion drafter for speculative decoding with the Qwen3.8-27B model, demonstrating improved acceptance length and throughput in benchmarks.

This paper presents CUTLASS-based kernels that fuse SwiGLU activation with GeMM at the tile level, achieving up to 2.47× speedup on NVIDIA H100 for efficient LLM inference.

The article highlights the paper 'A Gentle Introduction to Matrix Calculus' by Jan Magnus, published in the Journal of Econometrics in 2024, as a clear and valuable resource for fields like econometrics, machine learning, statistics, and optimization.

A tweet advising owners of NVIDIA's DGX Spark to prioritize using nvfp4 for optimal performance.

The article explains how RuneScape in 2004 optimized data transfer to fit a multiplayer RPG on 56k dial-up modems, detailing network constraints, encryption, and client decompilation.

This MIT paper introduces a mathematical framework for a unified representation of deep learning architectures, enabling automated analysis and optimization. It aims to replace manual design with a common language that can generate diagrams, graphs, or PyTorch code.

The article presents a cost-effective method to reduce LLM tail latency in real-time applications by sending duplicate requests and selecting the faster response, outperforming paid priority tiers in production experiments.

This paper presents a systematic machine learning study of 6G-IoT beamforming optimization, comparing network, environmental, device, and vision feature groups for predictive power, and applying clustering methods to enhance performance.

This paper introduces an ADMM-inspired momentum framework (AIM) and a new optimizer RADAR, offering a theoretical interpretation of momentum as a residual-driven multiplier correction, with experiments showing consistent gains over strong adaptive baselines.

Introduces 'Training Under Challenge', an executable-certificate framework that uses architecture-valid procedures to construct alternative models and estimate the empirical global-optimality gap of neural network checkpoints, with theoretical guarantees and experiments on ResNet-18 distillation and quantized denoising.

This paper analyzes when random low-dimensional reparameterizations can train neural networks, deriving an orientation-resolved master formula for the random-slice residual and introducing RaMaN, a scalable framework that predicts required latent dimensions while dramatically reducing memory costs.

The paper introduces ε-MemEvo, a framework for cross-task knowledge transfer in LLM-based program evolution, storing tactic memories as natural-language summaries and using an adaptive injection gate. It reports consistent gains on 8 optimization benchmarks with less than 1% computational overhead.

A deep dive into Ruby's internal ar_table structure for small hashes, explaining the linear lookup mechanism and exploring potential optimizations.

An essay arguing that NP-hard problems are often not as intractable in practice as commonly believed, citing modern solvers and algorithmic advances that handle real-world instances efficiently.

The article explains CUDA shared memory swizzling techniques to optimize GPU memory access patterns, with code examples demonstrating performance improvements.

This paper proposes a new asymmetric robust bounded sparse smooth (aR) loss function for l1-norm penalized geometric twin support vector machine (aRSGTSVM) to handle classification and regression tasks with label and feature noise, achieving feature selection and robustness. Experiments on synthetic and UCI datasets plus China stock market index tracking demonstrate superiority.

This paper introduces PAIR, a pairwise-aware inclusion reweighting method for adaptive rollout allocation in RLVR, improving sample efficiency and accuracy over pointwise allocators by correcting biases in pairwise gradient estimation.

This paper proposes Adaptive Hybrid PSO (AHPSO), which uses a sigmoid function on swarm diversity to automatically modulate gradient influence during search. Results show it outperforms standard PSO and rivals CMA-ES on certain problem classes, but the advantage is not universal.