Resume
← Blog

2026.11.05 · 4 min · ai · llm · news

Weight-space programming: an AI patched another AI's weights.

Researchers at MIT CSAIL had GPT-6 Astra edit 0.016% of a small model's weights, with no training, and its MATH-500 score reportedly rose from 7.9% to 48.1%. What it actually changed, why it didn't teach the model maths, and what the critics say. As of 2026.11.

An AI edited 131,072 of another model's 0.8 billion weights, 0.016%, and the authors report its MATH-500 score rose from 7.9% to 48.1%. Nobody trained it, and nobody taught it maths. The term for this is weight-space programming, and it's a few days old.

Watch the 60-second version on Instagram.

What the term means

The best-supported meaning: a general-purpose LLM reads and writes a network's weights directly, instead of an optimizer changing them through training. Takehi, Teoh, Zhang and Khattab at MIT CSAIL coined it on 2026.11.02, in a blog post using OpenAI's GPT-6 Astra. There's no separate paper.

Because the term is so new, it isn't settled. One repository uses it more broadly, for setting a network's weights without training through a compiler. And weight-space learning (hypernetworks, weights treated as data) is an older, separate field.

The contrast with training fits in two lines. This is an illustration, not the real patch (the exact diffs are in the authors' logs):

edit.py
# training: the optimizer nudges weights
loss.backward(); optimizer.step()
 
# weight-space programming: a reasoned edit
W = model.layers[15].mlp.down_proj.weight
W[:, cols] -= loop_direction_part  # no gradients

What Astra changed

The patient was Qwen3.5-0.8B. Astra traced its thinking loops to one direction written by the MLPs of layers 15 and 19, and subtracted that direction from 64 output columns in each. Only two tensors differ from the base model.

QWEN3.5-0.8B · 0.8B WEIGHTS131,072 · 0.016%LAYER 15 · MLP OUT64 COLSLAYER 19 · MLP OUT64 COLSGOLD NOT TO SCALE · ALL OTHER WEIGHTS UNCHANGEDFIG. 01 / 0.016% OF THE WEIGHTS

The gold in the figure is drawn far wider than scale. At true scale, 0.016% of that bar would be thinner than a hairline.

It also wasn't done by looking at the weights alone. Astra had Python over the weights and 50 forward passes of up to 512 problems each. No gradients, no GPU. With only 10 or 25 passes, MATH-500 reached 12.3% and 14.1%, so the test runs did real work.

Rambling was the bug

Before the edit, 93.0% of answers hit the 32,768-token limit: the model looped until it ran out of room. After it, 33.1% did. The scores the authors report moved with it: MATH-500 from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, IFEval from 21.0% to 62.3%.

0%25%50%75%100%HIT TOKEN LIMIT / QWEN 0.8B93.033.1MATH-500 / QWEN 0.8B7.948.1MATH-500 / GEMMA 3 270M6.87.7BEFOREAFTERFIG. 02 / FEWER CUT-OFFS, BETTER ANSWERS

Then the control. The same treatment on Gemma 3 270M barely moved it: 6.8% to 7.7% at best. The authors' conclusion is that editing could remove an obstacle that hid an ability, but couldn't add an ability the model lacked.

One gap to keep in mind: the post doesn't say how cut-off answers were scored, so you can't assign the gain precisely to fewer truncations. A rerun hosted on Hugging Face got 51.0% on MATH-500 for the edited model, but its base-model check was only a 25-problem smoke test.

What people get wrong

  • "An AI read the weights and taught a model maths." It removed a looping fault that hid ability the model already had, using code and test runs.
  • "It's sixfold smarter." One benchmark went from 7.9% to 48.1%. The Gemma control says the edit doesn't add skills.
  • "This replaces training." It doesn't. Goedecke argues the best way to change a model is still to train it.
  • "It's proven." It's one blog post, not peer reviewed, tested with one LLM, and there's no benchmark for this kind of work yet. Self-improvement and avoiding catastrophic forgetting are expectations, not results.

The pushback

Honnibal points out that the raw-floats-only experiments covered networks of just 17K to 172K weights, and every real language model experiment used Python, though he grants that earlier AI would have failed these tasks. Goedecke says a competent researcher would do the same, and that the fix resembles steering.

There is a no-tools result, on tiny networks: without tools or play-testing, Astra wrote a Breakout network from scratch that scored 517.4, against 25.3 for PPO at 500M steps. It didn't beat PPO in Seaquest or Asterix.

On a client project I'd treat a hand-patched model like a hot-patched binary: fine for diagnosing what's wrong, but I'd ship the version that comes out of a training run I can reproduce.

The short version

  1. Weight-space programming: an LLM edits a network's weights directly, no optimizer. Coined 2026.11.02 at MIT CSAIL.
  2. GPT-6 Astra changed 131,072 of Qwen3.5-0.8B's weights (0.016%), in two tensors, using Python and 50 test runs.
  3. Answers hitting the token limit fell from 93.0% to 33.1%; MATH-500 rose from 7.9% to 48.1%, as the authors report.
  4. On Gemma 3 270M it barely helped: editing removed an obstacle, it didn't add a skill.
  5. One blog post, not peer reviewed, and critics call it standard ML engineering. It doesn't replace training.

Sources: Takehi et al., Sparks of Weight-Space Programming · Authors' logs and diffs · Edited model card and rerun · dennj, Weight-Space-Programming · Honnibal · Goedecke · R&D World · Hacker News · The Register · Survey of Weight Space Learning