Reading and Writing Neural Network Weights With a Language Model
A PPO policy trained for 1 billion steps on the Kinetix car-on-ramp task never succeeded in 256 evaluation episodes. The car stayed stuck on the ramp. MIT CSAIL researchers gave its 171,664 weights to GPT-6 Astra in one prompt, with no tools, and asked for a fix. The edit changed 3.5% of the weights, added a small side path that backs the car up and switches off once the car clears the strip, and succeeded in all 1,024 test episodes.
That result is one of eight in a November 2 blog post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. They call the approach Weight-Space Programming: a language model reasons about what parts of a network compute and edits the weights directly, with no gradient descent and no training loop. The public logs hold the prompts, replies, code and evaluations. Per the repository, built-in tools were disabled, and any code ran in the authors' sandbox without network access.
Reading weights
The first test used four small MLPs, each with 17,253 weights, trained with PPO on Breakout, Pong, Pac-Man and Freeway on a 10 by 10 grid. Astra named the correct game for all four without being given the list of games. In a second test, three diffusion models each generated 12 by 12 grayscale circles, squares or triangles, and Astra identified each shape from the weights alone. It drew every shape as an outline although the models produce filled shapes, which the authors take as a sign that it analyzed weight geometry directly.
A third test gave Astra only the four middle blocks of three 7.2 million parameter character-level transformers, trained on Python, romanized Japanese and English. With no input or output layers and no way to run the network, it named the right language each time and identified the copying heads the authors had measured.
The sycophancy test went differently. The authors fine-tuned two Gemma 3 270M copies, one of them to be sycophantic. From weights alone, Astra failed in all four attempts. With a tool showing attention patterns on texts it wrote, capped at 20 runs, it picked the sycophantic model in all three pairs.
Writing weights from zeros
For MinAtar Breakout, Seaquest and Asterix, Astra received the game rules, source code and input and output layout, plus a 600-32-32-6 network of zeros. It saw no data, gradients or rollouts.
In Breakout, PPO scored 15.8 after 100 million steps and 25.3 after 500 million, while Astra's one-shot network scored 517.4. In Seaquest the figures were 5.6 and 88.9 for PPO against 53.1 for Astra, and in Asterix 25.8 and 55.2 against 53.2. The 500 million step networks finished ahead in Seaquest and Asterix.
The Go experiment used a 9 by 9 board and a network of 1,480,274 weights, too large for Astra's context. It had a Python sandbox but no Go library or internet. Its zero-pass network came from a 336-line script encoding liberties, chain tracking, captures, and avoidance of self-atari and eye-filling. That network won more than half its games against two smaller PPO networks, with no exact figure given. With 50 forward passes, Astra produced a network that beat an equal-sized PPO network trained on 20 million games.
Editing a language model
Qwen3.5-0.8B, with thinking mode on, hit the 32,768-token limit on 93% of MATH-500 answers, mostly in reasoning loops. Astra worked with a Python kernel over the weights and a budget of forward passes on practice problems from the GSM8K and MATH training sets. It traced the loops to one direction written by the MLPs in layers 15 and 19 and removed it from 64 output columns in each, a change of 131,072 weights.
The unedited model scored 7.9% on MATH-500, 13.8% on GSM8K and 21.0% on IFEval. With an edit built from 10 forward passes, those became 12.3%, 25.7% and 49.0%, and answers hitting the token limit fell from 93.0% to 87.2%. At 25 passes the scores were 14.1%, 25.1% and 39.9%. At 50 passes they were 48.1%, 67.1% and 62.3%, with 33.1% of answers hitting the limit.
IFEval measures instruction following, so the gain there suggests the fix reached beyond math. Scores did not rise steadily with budget: IFEval fell from 49.0 at 10 passes to 39.9 at 25.
Gemma 3 270M loops on only 8.5% of answers. Astra found ordinary reasoning errors there and could not meaningfully raise MATH-500, which went from 6.8% to at most 7.7%. The authors conclude that editing weights can remove an obstacle that hides an ability but cannot add one.
Stated limits
The authors call the results very early and cite context length and harness design as constraints. They are building a benchmark and also propose auditing networks for backdoors without running evaluations and steering capabilities with no data, which no experiment in the post tests.
About the author
David Borish is the author of The Tony Hawk Paradox: When Video Games Predict Reality. More at davidborish.com
