Analysis

GPT-6 Astra Edited Qwen's Model Weights and Fixed a Loop, Not Its Math

GPT-6 Astra took Qwen from 7.9% to 48.1% on MATH-500 by editing 0.016% of its model weights. MIT's logs show it fixed a loop and added no new math.

MesmerTools•November 11, 2026•4 min

In November 2026 a team at MIT CSAIL gave GPT-6 Astra a Python session over Qwen3.5-0.8B, a tiny open model, and told it to make the model better at math with no training. Astra changed 0.016% of the weights, and the score on the MATH-500 test went from 7.9% to 48.1%. An outside re-run got a similar jump.

That reads like one model teaching another algebra. MIT itself says this mostly fixed a loop, so I read the published logs to see what the edit was. They back MIT up. Qwen could already solve a lot of these problems, but it kept getting stuck repeating itself before it reached an answer. Astra got it moving again, and my read is that no new math went in.

A dark grid of weight cells with two thin columns glowing amber and a text cursor at the top of one, under the headline Models now edit models

What 7.9% looks like: one line, 1,503 times

Take problem 807: the point (0,3) in polar coordinates. Qwen gets r = 3 right away, starts worrying about tan(π/2), and then settles on one sentence.

Verbatim from MIT's logs: the unedited model never leaves this line, and the edited one boxes the right answer on the same problem.

That's the normal case. MIT published 200 of these answers in full, so I counted, and 186 hit the token limit. One of them notices, writing "Wait, I'm just repeating." It writes that 175 more times. Here are all 200, before and after the edit.

A cut-off answer almost never scores, so the score is just the amber dots. Qwen's own model card warns about these loops and ships with thinking off.

What changed in the model weights: five times past zero

So Astra went after the loop, with a head start, since the task prompt had already named repetition. It traced the looping to a direction in what layers 15 and 19 each write, then picked the 64 neurons per layer that push along it hardest. The obvious move is to take a direction like that to zero. Astra kept going. The control at the end is why I believe it.

The real columns, measured from the diff: each one's lean along the loop comes out five times longer on the other side, while a larger edit at right angles barely moves the score.

Going past zero sounds reckless until you see that zero isn't special. Removal is just one notch on a dial, and on the same problems doubling the strength doubled the score gain. I read the loop as a pressure you push against until it stops winning.

Measured on test texts in MIT's logs: removal costs a looping token about 2.4% with no kink there, and the final 64-column edit costs it about 10% against about 1% for an ordinary token.

But finishing only helps if the model already knew

On fresh practice problems the edit rescued 579 answers and lost 5. It let the wrong answers out too: answers finishing wrong went from 6 to 187, nearly all of which used to hit the wall.

So rescues make a rough ruler: of the answers that newly finish, how many are right?

Of the answers the edit freed, about two-thirds are right on MATH-500, chance level on GPQA's four-option test and almost none on AIME.

Gemma 3 270M, another tiny model, is the clean test because it rarely loops. None of Astra's edits moved its math score by more than about a point.

So check the token limit before you believe the jump

My read as of November 2026 is that an edit like this frees what a model already does, and better algebra still has to be trained. That's still a real fix if you run a small model that loops. I'd just check its answers harder afterward, because a wall is obvious and a confident wrong answer isn't.

When the next "LLM made a model smarter" result comes along, the first thing I'd look up is how many answers stopped hitting the limit. If that moved with the score, it's this story again.

More from the Blog