No Training, Sixfold Math Gain, MIT Claims

Four MIT researchers say an OpenAI model changed one weight in about every 6,100 inside a small open model¹. The small model's math score rose from 7.9%² to 48.1%. Nobody trained it.

Rikiya Takehi and three MIT CSAIL colleagues published the result in a November 2 blog post. They say OpenAI's GPT-6 Astra edited 131,072 weights in Alibaba's Qwen3.5-0.8B³. That is about 0.016% of its 800 million⁴ parameters. No optimizer reportedly ran, and no gradient descent ran.

We have to start with a damaging admission: we cannot check any of it. OpenAI has not confirmed the figures. Astra was released on September 3, 2026⁵, so outside teams can rerun the test with it. Treat every number below as reported and unverified.

The claim still matters. If one model can patch another, improving a model starts to look like shipping a software update. That would change who builds models, who pays for them and who gets paid.

Weight-space programming repairs a known bug

The authors call the method weight-space programming, which they define as a setup where "a general-purpose LLM reads and writes a network's weights directly." According to letsdatascience.com, the team handed Astra a Python kernel and the raw Qwen weights and asked it to fix the looping.

The bug was no secret. Qwen's own model card warns that "this model is prone to thinking loops," the MIT team notes. On MATH-500, 93%⁶ of the model's answers ran until they hit the token limit, which is the cap on how much text a model may write for one answer.

The authors write that Astra "traced the looping to one direction written by the MLPs of layers 15 and 19⁷⁸." It then "subtracted that direction from 64 output columns in each." An MLP layer is the part of each block that does most of the number crunching. The arithmetic works out: two layers times 64 columns is 128 columns, and 131,072 divided by 128 is 1,024 numbers per column³.

There are two ways a result like this can go. Repair means the model stops looping and finishes answers it could already reach. Installation means the model gains math skill it never had.

The evidence points hard at repair. i10x reports that outside checks found the truncation rate fell from 93%⁴ to about 33.1%. The authors themselves say the gain came from fixing a failure mode, and they do not claim Astra taught Qwen new math.

Who owns the editor owns the margin

The authors' full scoreboard goes past math. GSM8K, a grade-school math test, rose from 13.8%³ to 67.1%, a 4.9x gain. IFEval, an instruction-following test, rose from 21.0% to a reported 62.3% in the 50-forward-pass setting, a 3.0x gain. That IFEval jump fits the repair story too, because a model that stops rambling follows instructions better.

The sixfold headline deserves a second look. MATH-500 has 500 problems⁹, so 7.9%² is roughly 40 correct and 48.1% is roughly 240. The real move is 40.2 points from a very low floor. A model that went from one right answer to six would also be "sixfold" better.

Pushback came fast. Within a day, two outside developers published replies arguing that the biggest result relied on standard machine-learning engineering, R&D World reports. One called the write-up "massively misleading," according to letsdatascience.com.

Even so, the pattern showed up twice in one week. A day after the MIT post, Legora's CTO said Anthropic's Claude Fable hand-wrote the weights of a roughly 100,000-parameter¹⁰ legal classifier, Dealroom also reports. In about 24 hours, models writing weights for other models went from one lab post to a working company claim.

Every one of these edits needed open weights. A closed API model shows you outputs and keeps its parameters private, so nobody can patch it this way. The editor, though, was a closed frontier model. Open models become the raw material, and the closed lab sells the tool that works it.

Futurum's October 26 report shows how that split already pays. In its October AI Tools Pulse, open-weight models make up just 5%¹¹ of enterprises' metered AI consumption. Microsoft 365 Copilot takes 36% and Claude takes 29%. Futurum also finds that in-house open-weight builds report the weakest ROI of any approach.

Follow the cash and the picture is plain. Tokens are moving to open models while the money stays with the frontier labs. A patch economy would extend that pattern, because the most valuable step, finding the edit, runs on a closed model. Whoever owns the editor collects on every checkpoint it fixes.

We think the honest headline is automated debugging. A frontier model found a known defect and switched it off. That is useful, and it is cheap to ship. Whether the method scales to a 70-billion-parameter model is unclear, since skills there may spread across far more layers.

For now the evidence supports one decision: treat any edited checkpoint as a supply-chain artifact that needs its own tests. Three results would change the call: a replication on fresh problems, a gain on answers Qwen already finished, and success on a larger model.

Patched open models may matter by 2030

Futurum's survey data sketches the longer arc. In a drill-down of 200¹¹, open-weight models ran in production at 31% of enterprises in July 2026 and 42% by September. Production users put open weights at 34% of their AI tokens today, and they expect 41% within a year.

That is 7 points a year. Extend it in a straight line and the share reaches 48% in 2028 and 62% by 2030, with open weights crossing half of these users' tokens around 2029. Straight lines rarely hold for four years, so read that as a rough guide and nothing more.

Every one of those tokens runs on weights that someone can read and edit. Futurum says open weights already carry most tokens on the Vercel and OpenRouter gateways, led by Chinese models. A patch economy would sit on top of that base. A model could start as Alibaba's Qwen and carry a patch written by an OpenAI model before an enterprise ever deploys it.

Buyers already care about provenance. Country of origin is a major or deciding factor for 47% of production users, up from 35%, Futurum reports. A patched model muddies that answer. If one checkpoint carries work from two countries, a buyer's origin policy has to cover patches as well as base models.

For builders, the risk is lopsided. Hashing and diffing every checkpoint costs a script and a few minutes per release. A hidden patch that weakens refusals or plants a trigger costs a breach and a recall. And edits compound, since each patched checkpoint can become the base for the next patch.

Diff the Astra checkpoint against stock Qwen

You can test the repair story yourself on modest hardware, since the model has only 0.8 billion parameters. A community upload on Hugging Face, JonathanColetti/Qwen3.5-0.8B-Astra-WeightEdit, is labeled as the edited model. R&D World points to the team's logs on GitHub under RikiyaT/lm_understand-logs, released under CC BY 4.0. We have not verified that the upload matches the authors' edit, so that is your first check.

Step one: load stock Qwen3.5-0.8B and the edited upload, then subtract each weight tensor from its twin. A weight diff is just the list of numbers that changed. If the upload matches the post, you should find about 131,072 changes³, all in layers 15 and 19⁷.

Step two: write 50 math problems of your own³. Skip MATH-500, since an edit tuned against a public test can look better than it is. Mix easy word problems with a few hard ones. Save them in a plain text file so you can rerun the set later.

Step three: run both models on your set with the same token cap. Log two things per answer: whether it finished or hit the cap, and whether it was right. Then compute accuracy only on answers that finished.

That last number is the real test. If the edited model's accuracy on finished answers matches the stock model's, you are looking at loop repair. If it climbs, something more than repair may be going on, and that result is worth publishing.

Step four: run 20 prompts from your own domain on both models, such as support tickets or code questions. Look for regressions, like worse facts or answers that stop too early. A small edit can move behavior far from the layer it touched.

Step five: hash both checkpoints and store the hashes next to your test log. A hash is a short fingerprint that changes if even one weight changes.

Expect the stock model to fail most hard problems, and expect your first harness to break on answer parsing. Fix the parser and rerun all 70 prompts. Keep the script and run it on every edited checkpoint you pull from now on.

Practise it step by step: Build the Verification Habit