麻省理工学院（MIT）的研究人员展示了一种新颖的方法，其中一个AI模型GPT-6 Astra直接编辑了另一个AI模型Qwen的权重，而无需传统训练。这一过程被比作神经网络的基因编辑，涉及修改Qwen权重的一小部分（0.016%）。干预显著提高了Qwen在MATH-500基准测试中的表现，其分数从7.9%提升到48.1%。然而，研究人员指出，该技术仅在消除性能瓶颈方面有效，而非传授新技能。 AI
影响 通过直接编辑权重，绕过传统训练周期，展示了AI模型改进的潜在新范式。
排序理由 研究论文详细介绍了模型权重编辑的新型AI技术。
在 Mastodon — mastodon.social 阅读 →
AI 生成摘要 · Google Gemini · 来自 6 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
<p><img alt="" class="attachment-full size-full wp-post-image" height="768" src="https://the-decoder.com/wp-content/uploads/2026/11/Weight-Space-Programming.png" style="height: auto; margin-bottom: 10px;" width="1376" /></p> <p> MIT researchers show that OpenAI's GPT-6 Astra can …
<p>On 2 November 2026 a team at MIT CSAIL published <strong>Sparks of Weight-Space Programming</strong>, and within a day my feed was full of the same claim: GPT-6 Astra opened up a small open model, Qwen3.5-0.8B, changed 0.016% of its weights by hand, and its MATH-500 score went…
An AI rewrote another AI's brain — no training. MIT had GPT-6 Astra edit Qwen's raw weights directly (just 0.016%), and its math score jumped 7.9% → 48.1%. Authors' own results, not yet replicated. ▶️ https:// youtu.be/qnNRWkpktKg Is weight-editing the future of fixing models — o…
an AI just did brain surgery on another AI lol mit researchers gave gpt-6 astra the raw weights of a tiny qwen model and said: make it better at math. no training allowed. it found what made the model loop forever and edited 131,072 weights. that's 0.016% of the model. math-500 w…
'Like gene editing for neural networks': MIT researchers used GPT-6 Astra to edit Qwen's open weights directly, boosting math performance. Via @venturebeat #AI #ArtificialIntelligence 💻 🤖 🧠 'Like gene editing for neural ...
MIT researchers had GPT-6 Astra rewrite a smaller AI’s brain directly, and its maths score jumped sixfold MIT researchers say GPT-6 Astra can read and edit neural network weights directly, lifting Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 with no training. # OpenAI # Interpreta…