MIT controversy over GPT-6 Astra directly rewriting Qwen weights
The story so far
The MIT CSAIL team proposed 'Weight-Space Programming', allowing GPT-6 Astra to directly read and modify the weights of Qwen3.5-0.8B without using gradient descent, backpropagation, or traditional training loops. The team reported that by adjusting only about 0.016% (131072 parameters), MATH-500 improved from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%. Subsequent discussions noted that this result is not achievable simply by 'looking at weights': Astra read and wrote weights in Qwen experiments through a Python environment and could perform up to 50 rounds of forward inference using the GSM8K and MATH training sets; the setting of truly only relying on original weights, without tools, and not running the network appeared only in small network experiments with 1.7 million to 17.2 million weights. spaCy author Matthew Honnibal questioned the study for serious misdirection, stating it mixes reading weights, writing weights, and fixing large models, with a presentation that blurs the boundaries of different capabilities. The team stated that current capabilities are still limited by context length and tool framework efficiency, and they are building benchmarks and test frameworks for Weight-Space Programming.
AI-written from the reports · Updated 14 days ago
Developments
3 developments
- 5 Nov, 15:10 · 1 reportsMIT's masterpiece criticized: Astra drastically modifies Qwen weights? Just kidding!QbitAI:spaCy author attacks MIT Astra research: modifying Qwen weights accused of being a smoke screen
- 3 Nov, 20:21 · 1 reportsAstra can read and write weights to improve QwenHacker News · 300+:The 'Astra can read and write weights to improve Qwen' post is misleading
- 3 Nov, 02:21 · 4 reportsAstra can read and write weights to improve QwenQbitAI:MIT experiment: GPT-6 Astra directly modifies Qwen3.5 weights, mathematical ability improves by about 6 times
Coverage timeline
Follow the coverage to see every side of the story.
- QbitAIspaCy author attacks MIT Astra research: modifying Qwen weights accused of being a smoke screen
spaCy author Matthew Honnibal questions MIT Astra research, claiming its experiment modifying Qwen3.5-0.8B weights to improve math scores is misleading, mixing reading weight, writing weight, and repairing large models; Honnibal points out that Astra repairs large models by using Python programming and multiple forward tests, not directly reading weights, criticizing its presentation for blurring the boundaries of different capabilities.
- Latent Space[AINews] Behind Latent Space: Kill My SaaS, Jev & AI in 2026 | AIEi Shanghai
Latent Space reports on the latest trends in the AI field, including cutting-edge large model direct weight editing technology (such as GPT-6 Astra editing Qwen3.5-0.8B weights to increase MATH-500 from 7.9% to 48.1%), as well as the latest evaluation results of Tencent Hy4 achieving 75.2 points on APEX-Agents and Kimi K3.1 Pro ranking 14th in Agent Arena.
- Hacker News · 300+The 'Astra can read and write weights to improve Qwen' post is misleading
Astra directly reads weights in 17K-172K weight network experiments, but in the 800M weight Qwen3.5 mathematical ability improvement experiment, it requires combining Python programming and model execution, not just reading weights; Astra modifying 0.016% weights increased Qwen3.5-0.8B's MATH-500 from 7.9% to 48.1% and tripled IFEval.
- ITHome · AIFeaturedMIT CSAIL proposes weight space programming: GPT-6 Astra modifies Qwen3.5-0.8B's weights by about 0.016%, boosting MATH-500 score to 6 times
MIT CSAIL researchers propose 'weight space programming', allowing GPT-6 Astra to directly read and modify the neural network weights of Qwen3.5-0.8B, adjusting only about 0.016% (131072 parameters) to increase MATH-500 score from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%.
- QbitAIFeaturedMIT experiment: GPT-6 Astra directly modifies Qwen3.5 weights, mathematical ability improves by about 6 times
MIT's experiment allowed GPT-6 Astra to directly read and edit the weights of Qwen3.5-0.8B, changing only 0.016% of the parameters, without training and gradient descent, MATH-500 score improved from 7.9% to 48.1%, and instruction execution ability also increased 3 times.
- Hacker News · 300+FeaturedAstra can read and write weights to improve Qwen
The research team proposed Weight-Space Programming, discovering that GPT-6 Astra can directly read and write neural network weights: with only 17,253 weights, it can determine which game the network is trained on, and from scratch, write a small game network that defeats PPO trained for 100 million steps without gradients or training.