optimization

The paper introduces BFTR, an algorithmic framework for telecom tariff recommendation that prioritizes customer budgets and prevents overcharging. Experiments on a Nigerian MTN-inspired dataset demonstrate the framework's effectiveness with zero surcharge and high utility.

SALT is a tool that uses CELF to retrieve important sentences from long documents, reducing KV cache usage in language models for improved efficiency. The author is seeking help to optimize the retrieval budget percentage based on input queries.

Introduces iterative tensor network transformations (ITNTs) for efficient element-wise evaluation of nonlinear functions on tensor trains, enabling data processing and optimization on large-scale datasets with applications in reactive flow fields and Max-SAT problems.

This paper introduces BayesPrompt, a Bayesian approach to optimize prompts for large language models, ensuring they are both efficient and human-readable, unlike traditional methods that produce unintelligible pseudoprompts.

LEGO-RL presents a framework to bridge native coding-agent harnesses with scalable policy-gradient reinforcement learning, improving performance on benchmarks like SWE-bench.

The author achieved a 12x speedup for DeepSeek V4 Flash on a Mac Studio M3 Ultra by optimizing kernels and implementing effective caching strategies, reducing chat turn latency from 6-20 seconds to 1.6 seconds.

A user reports that the Qwen 3.8 27B model achieves 50-60 tokens per second on dual 5060 TI cards, showing unexpected speed improvements over previous versions like Qwen 3.6.

The paper explores using complex-valued states inspired by quantum theory in sequence models, showing it can reduce optimization steps in Mamba and attention-based models with better sample efficiency.

This empirical study explores the interaction between fractional optimizers and fractal activation functions in neural network training, showing selective pairings and highlighting controlled fractional memory as a promising direction.

This paper proposes a task-driven three-layer distributed scheduling method for emergency Earth observation in large low-Earth-orbit constellations, achieving higher emergency coverage and lower routine-plan disruption compared to existing methods.

Tweet highlights running the Qwen 3.8 27B model locally on an RTX 5090 system with 32GB VRAM, achieving 115 tokens/sec, and notes the official BF16 checkpoint is 55.6GB.

The author details performance improvements to the garbage collector for their toy programming language Plush, focusing on optimizing the copying algorithm to reduce collection time.

The article describes various tricks and strategies for writing fast compilers, focusing on minimizing code execution, reducing memory usage, and optimizing common paths to achieve compilation speeds of over 500,000 lines of code per second.

Andrew reflects on the burnout caused by vetting AI-generated code for a game optimization, comparing the time and mental cost to manual coding. He uses an interactive visualization to illustrate the trade-offs between AI-assisted development and traditional methods.

This paper introduces Smart Routes, a platform for developing and comparing algorithms for vehicle routing problems with realistic constraints, showing that deep learning and heuristic methods can match exact solutions in quality with less time for larger problem sizes.

The paper proposes an event-driven framework using Transformer and deep reinforcement learning for dynamic multi-depot vehicle routing, comparing it with heuristic and optimization methods.

QUASAR is a quantization-aware training method that uses loss-aware reconstruction to lower the loss floor, improving low-bit model performance in large language models with significant accuracy gains at 2-4 bits.

This paper explores using large language models and AI agents for autonomous chip design, modeling it as an AI-organization and discussing action spaces for black-box optimization in chip design scenarios.

This paper introduces the Travelling Thief Problem with Drone (TTP-D), which jointly optimizes ground routing, drone synchronization, and item selection using mixed-integer programming, metaheuristics, and attention-based deep reinforcement learning.

The paper proposes improvements to the optimization problem in combination loss analysis using modern techniques and AlphaEvolve, yielding an improved upper bound on the matrix multiplication exponent.