跳到正文
Turing Post (Ksenia Se)· @TheTuringPost · X·· 20 天前AI 评分58
AI 导读

MIT CSAIL 团队让 GPT-6 Astra 直接编辑 Qwen3.5-0.8B 的权重,在未重新训练的情况下把 MATH-500 成绩从 7.9% 提升到 48.1%。

正文 · 原文

.@MIT_CSAIL’s team had GPT-6 Astra edit two layers of Qwen, and its MATH-500 score went from 7.9% to 48.1%. The announcement of Sparks of Weight-Space Programming passed a million views on X within hours!

It may sound like sci-fi with "model on model surgery" but the real surprise was how the demonstration was staged.

1. The Breakout paddle:

The team fed GPT-6 Astra 17,253 weights from a small game-playing network, plus the grid and controls. It guessed Breakout - paddle at the bottom and 3 brick rows at the top - without seeing a single frame.

2. Editing Qwen’s weights:

They asked Astra to make Qwen3.5-0.8B better at math. No retraining. Test questions stayed private.
Qwen kept repeating steps, and Astra weakened a pattern associated with repetition in layers 15 and 19, touching just 0.016% of the model.

Before the edit, 93% of answers hit the length limit; afterward, 33.1% did. Much of that sixfold gain came from Qwen finally finishing problems it could already solve.

However, on Gemma 3 270M, which generally finished its answers but made arithmetic and algebra mistakes, Astra’s edits produced no clear gain.

---

So one model suggested where to look, and another figured out which of Qwen’s 800 million weights to change. That's model cooperation the media hasn’t hyped yet!

This paper is especially compelling because of the orchestration: describing a problem in plain English, giving an off-the-shelf model the tools to investigate, and letting it work out which weights to change in another network.

The team even tried a shorter prompt without the repetition hint, and Qwen matched the hinted run at 25 test batches: 14.1%. But the table has no 50-batch result for that shorter prompt. That’s the row I’d love to see.

I wrote more about this in this week’s newsletter, along with the other AI research and news that caught my eye → https://www.turingpost.com/p/what-is-weight-space-programming-and-why-it-suddenly-blew-up-the-internet

来源:Turing Post (Ksenia Se) · x.com