Yesterday's hot paper, "Sparks of Weight-Space Programming," found that frontier models can natively read and write neural network weights as text descriptions of floating point numbers. A key finding is that Qwen 3.8 27B running on a laptop can perform this task. The paper coins the term "Weight-Space Programming" for when LLMs analyze and modify model weights without tools or code execution.
Researchers showed GPT-6 Astra can interpret raw weights to identify a network's trained task and, without environment feedback, write small game-playing networks that outperform networks trained with PPO for 100 million steps. Furthermore, Astra can fix reasoning loops in Qwen 3.5-0.8B by editing only 0.016% of weights, raising MATH-500 scores from 7.9% to 48.1% and generalizing to instruction following. The study involved training neural networks to play Atari games like Breakout, Pong, Pac-Man, and Freeway, then prompting LLMs to identify the game from weights alone.
While Matthew Honnibal called the paper misleading for conflating topics, and Sean Goedecke deconstructed the Qwen example as standard ML researcher steps, the authors demonstrate LLMs can understand model behaviors purely from weights. The accompanying GitHub repository provides prompts and weights for the experiments.
Source: Hacker News · Summarized by HeadlinesBriefing