MIT researchers say GPT-6 Astra edited Qwen open weights to boost math

Four MIT CSAIL researchers say OpenAI's GPT-6 Astra raised a small open-source model's math score about sixfold by changing its weights directly, with no training. In a blog post published November 2, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab reported that Astra edited 131,072 weights, or 0.016%, of Alibaba's Qwen3.5-0.8B. That edit took the model's MATH-500 score from 7.9% to 48.1%. GSM8K rose from 13.8% to 67.1%, and IFEval instruction-following went from 21.0% to 62.3%.rikiyat.github.io

The researchers call the method "weight-space programming." Their claim is that frontier language models can read neural network weights and patch them the way a programmer patches code. The authors say the Qwen result came from fixing a failure mode, not from teaching the model new math.

Fixing a loop, not adding ability

Astra worked in a Python environment and could run the model a limited number of times on practice problems. It traced Qwen's repetitive "thinking loops" to one direction written by the MLPs in layers 15 and 19, then subtracted that direction from 64 output columns in each layer, according to the post. The share of MATH-500 answers that hit the 32,768-token limit fell from 93% to 33%. Qwen's own model card warns that the model is prone to these loops.rikiyat.github.io

The strongest run took 158 conversational turns, 80 Python executions and 20.3 hours, including 4.7 hours lost to a service outage, VentureBeat reported. With smaller budgets of 10 or 25 test runs, the loops "barely went away."venturebeat.com

The same approach did not work on Gemma 3 270M, which rarely loops. Its MATH-500 score moved by less than a point. "Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked," the authors wrote.rikiyat.github.io

Early checks and open questions

The work is a blog post, not a peer-reviewed paper, but the team released full experiment logs. One community re-run on Hugging Face applied the released edit and scored 50.2% on MATH-500 with thinking mode on. The same check found that the unedited base model already scores 45.8% when thinking mode is off. Takehi said on X that a separate rebuild from the logs scored 51.0%. Neither check has reproduced the GSM8K or IFEval gains.x.com+1

In other experiments, Astra identified game-playing networks from their raw weights and repaired a failing driving policy by editing 3.5% of its weights. It also wrote a Breakout network from scratch that outscored a conventionally trained baseline, VentureBeat reported.venturebeat.com

Hype and caution

The reactions split. PauseAI CEO Maxime Fournes called it "a terribly reckless research direction." Other commenters argued it showed only that a model can remove an obvious flaw. The authors say their next step is a benchmark and harness meant to turn what "currently seems like black magic into a rigorous, reproducible engineering discipline," The Register reported.theregister.com