MIT had GPT-6 Astra edit Qwen's weights by hand. The 6x math gain is a loop repair
Astra edited 0.016% of Qwen3.5-0.8B's weights and lifted MATH-500 from 7.9% to 48.1%. The number holds, but the authors' own tables show what it actually fixed.
Vlad Makarovreviewed and published
On 2 November, four MIT CSAIL researchers published a blog post claiming a frontier language model can read and write another network's weights in plain language, with no gradient descent and no permission to write training loops. The result that travelled was one edit to Qwen3.5-0.8B, taking MATH-500 from 7.9% to 48.1% while touching 0.016% of its weights. The number is real. It also means less than the sixfold framing implies: 576 of the 579 problems the edit rescued were answers truncated at the token limit, and accuracy on questions the model already finished did not change.
A blog post with a lab notebook attached
The artifact is the post itself, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Its header says "Not published yet" and "No DOI yet", so there is no paper, no peer review, no independent replication. VentureBeat's Carl Franzen covered it the same afternoon.
What the authors did ship is more than the usual announcement. The complete logs contain every prompt, every reply with its reasoning summary, every Python call and its output, and the evaluation behind each score. The edited checkpoint is downloadable, and the post walks through all 17 edits in the successful run, including the attempts that failed along the way. That is a checkable surface most preprints do not offer. It is still one lab's work on one small model, graded by the lab.
Qwen's model card had already named the bug
Before any clever claim, the target was documented as broken in a specific way. Qwen's own model card states that "when using the recommended sampling parameters, Qwen3.5-0.8B is more prone to entering thinking loops compared to other Qwen3.5 models, which may prevent it from terminating generation properly."
The post's evidence is blunt: 85 of the model's first 96 practice answers ran to the 32,768-token limit. It was not failing at arithmetic. It was failing to stop.
Astra traced the looping to one direction written by the MLP blocks of layers 15 and 19, then subtracted that direction from 64 output columns in each, which is 131,072 weights. The technique resembles the refusal-direction work the post cites, and ROME and MEMIT had already shown that targeted weight edits can change what a transformer says. What is new is that a general model found the direction by reasoning over weights instead of a human running probes.
The ladder underneath the 48.1
Astra got practice problems from the GSM8K and MATH training sets, a Python kernel over the weights, and a budget of forward passes. The authors ran the search three times, at 10, 25 and 50 passes, and the published table is the most useful thing in the post, because the headline number only appears at the widest budget.
| Qwen3.5-0.8B | MATH-500 | GSM8K | IFEval | Hit token limit |
|---|---|---|---|---|
| Base | 7.9% | 13.8% | 21.0% | 93.0% |
| Edited, 10 forward passes | 12.3% | 25.7% | 49.0% | 87.2% |
| Edited, 25 forward passes | 14.1% | 25.1% | 39.9% | 86.4% |
| Edited, 50 forward passes | 48.1% | 67.1% | 62.3% | 33.1% |
Read the last column first. At 10 and 25 passes the loop barely moved and math stayed near 12 to 14 points. At 50 the truncation rate fell from 93% to 33% and everything else moved with it. The authors report a combined gain of 49.8 points, standard error 1.5, with 579 rescues against five regressions, and state that 576 of those rescues were outputs previously cut off.
That is a good result and a narrow one. The model could already work out much of what it was scored on; the edit let it finish. Nor is it clear how much of the 48.1 depends on the 50-pass search rather than on the edit: the successful run is logged as 158 turns, and the published fix is the last step of 17 logged edits, most of them probes.
The control the authors ran against themselves
The most persuasive passage in the post limits the claim. Having argued that the gain came from removing loops, the authors tried the same procedure on Gemma 3 270M, a model that rarely loops: 8.5% of its answers hit the token limit, against 93% for Qwen. Astra found ordinary reasoning errors, wrong arithmetic and invalid algebraic steps, with no single component to switch off. Its edits moved MATH-500 from 6.8% to between 7.0 and 7.7.
"Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked," they write. That sentence summarises the experiment better than the announcement thread, which led with the sixfold gain, and VentureBeat reaches the same place, noting that the experiment did not establish that Astra had inserted six times as much mathematical knowledge into Qwen.
Where the evidence is genuinely interesting
Two smaller experiments are harder to explain away. Four PPO-trained policies for Breakout, Pong, Pac-Man and Freeway shared an identical MLP of 500 to 32 to 32 to 5, 17,253 weights, identical inputs and actions, a billion steps each. Handed the weights as text with no list of candidate games, Astra named all four and reasoned about where each input channel's first-layer weights concentrate: the paddle row, the brick band above it, the differentiated left and right outputs. That is a reading of structure.
In the fourth, a driving policy that never got the car over a strip was patched in one response covering 3.5% of its weights, taking success from 0% to 100%.
The from-scratch claims are thinner. Astra wrote MinAtar and Go networks without testing them first, and the post says they outplay networks trained with PPO for 100 million steps. True and carefully scoped. Against the 500-million-step baselines it splits: Breakout 517.4 against 25.3, but trailing in Seaquest and slightly in Asterix. The sycophancy test failed outright from weights alone and worked only once Astra could inspect attention patterns, which the post reports without dressing up.
What the edit cost, and what nobody has measured
Direct editing is not free. The Qwen run consumed 50 forward passes over thousands of practice problems, 80 code executions and most of a working day, and the authors claim no saving against fine-tuning. VentureBeat reports that the researchers "have not yet established" any saving in time or compute against conventional training. A patch that takes twenty hours to find, on a model small enough to fine-tune on one GPU, is a research result, not a cheaper workflow.
The 7.9% baseline is the other loose thread. That is unusually low for a 0.8B model at temperature 0.6 with a 32,768-token budget, and the reason is in the table: 93% of its answers never finished. Any intervention that fixes termination looks transformative on a benchmark scored by final answer. The fair test is a model that terminates normally and still gets the math wrong, and on the one such model tested here, it did not work.
What would settle the larger claim is not more enthusiasm on X: a checkpoint the authors did not choose, a model without a documented pathology, an edit found under a budget a normal team could afford, and a cost comparison against LoRA on the same task. Until then the defensible statement is narrow. A frontier model can locate a specific failure in a small network's weights, describe it in language a human can follow, and write a patch that measurably repairs it. That is a real capability, and it is not the same as understanding weights in general, still less optimising them.
The gene-editing line, which Yuchen Jin used on X and the coverage adopted, works precisely because gene editing is powerful, narrow, and often fails to do what was intended.
