Astra reads and writes neural network weights like code
On November 2, four MIT CSAIL researchers (Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab) published a post showing that OpenAI's GPT-6 Astra can read and write neural network weights without training, gradients, or specialized tools. The headline result: Astra edited 131,072 of Qwen3.5-0.8B's 800 million weights (0.016%) and, with the same sampling settings, lifted MATH-500 from 7.9% to 48.1% and IFEval from 21.0% to 62.3%. They call it weight-space programming: the weights of a network are a program that gets read and written, by a general-purpose LLM.
The concept
For decades, weights have been written only by optimizers. The authors note that LLMs already optimize prompts, harnesses, even training code. The weights were the last artifact left to symbolic algorithms, because no human can hand-edit 800 million numbers. Astra now does that, reading the weights to understand what a network computes and writing or editing them to change what it computes. The authors call the results "sparks" of a capability, not a method. Some experiments give the model zero tools, others a Python kernel or a budget of forward passes.
Reading: what is this network for?
Astra gets raw weights as text and is asked what they do.
- Four game policies (grid versions of Breakout, Pong, Pac-Man, Freeway), each an identical 17,253-weight MLP trained with PPO for a billion steps. Astra named the correct game in all four without seeing the options, reading where first-layer weights concentrate: in Breakout, the agent's sit on the bottom row, the paddle.
- Three diffusion models trained to generate circles, squares, or triangles (23,824 weights each), no tools, no forward passes. Astra identified what each generates by tracing shapes in the output layer and drew them as outlines although the models produce filled shapes, which the authors read as evidence it reads the structure rather than simulating the network.
- Three character-level transformers of 7.16 million weights, trained on Python code, Japanese romaji, or English Wikipedia, with the input and output layers removed, so nothing says which character is which and the model cannot be run. Astra, allowed Python but forbidden to run the network, named the correct language in all three and found every "copying" (induction) head with no false positives, matching ground truth. The method it arrived at is one published by Elhage et al., which the authors never mentioned.
Writing: fix it, or build it from zero
In the repair experiment, Astra got the official Kinetix PPO actor (171,664 weights) for a car-on-ramp level where PPO had hit a local minimum: zero successes in 256 episodes. Given the weights, the scene, and one recorded episode, it edited 3.5% of the weights in one try with no tools; the modified car succeeded in all 1,024 episodes. The fix, the authors say, is a small parallel path that reads the car's position and speed, backs it up past the strip, then lets the trained network take over.
From scratch, on three MinAtar games (Breakout, Seaquest, Asterix), Astra got only the rules and source code: no data, no gradients, no tools, no rollouts. It wrote the weights as text in one reply. The authors report that the results beat PPO trained for 100 million steps in all three games and score 20 times as much as 500-million-step PPO in Breakout, coming close in Asterix and Seaquest.
The largest writing task: a 1.48 million-weight network for 9x9 Go, in a Python sandbox with no internet and up to 50 forward passes. With zero passes the generated network won more than half its games against two smaller PPO networks; with 50, it beat an equal-sized PPO network trained on 20 million games. Astra wrote it like a small program: early layers mark stones and liberties, the last layer scores moves on capture and avoiding atari.
The Qwen fix
The result that will get the most discussion: a real language model getting better at math through weight edits. Qwen3.5-0.8B has 0.8 billion weights, 24 layers, and thinking on. Its model card notes a propensity for thinking loops, and the base model confirms it: 85 of its first 96 practice answers ran to the 32,768-token limit, as did 93% of its MATH-500 answers.
Astra got a Python session over the weights (no GPU), practice problems from the GSM8K and MATH training sets, and 10, 25, or 50 forward passes. It traced the loops to a single direction written by the MLPs of layers 15 and 19, and subtracted that direction from 64 output columns in each layer: 131,072 weights, 0.016% of the model.
With 50 passes, the authors report MATH-500 rising from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%, with answers hitting the token limit falling from 93% to 33%. At 10 and 25 passes, MATH-500 reached 12.3% and 14.1%, and the loops barely went away.
The negative result matters as much. On Gemma 3 270M, a model that rarely loops (8.5% of answers hit the token limit), the same treatment moved MATH-500 by no more than a point (base: 6.8%), and Astra itself reported the gain as not established. It read Gemma's failures as ordinary reasoning errors, wrong arithmetic, invalid algebra, with no component to switch off. In the authors' words, the edit removed an obstacle hiding an ability; it did not add an ability the model lacked. The number to remember: 0.016% of the weights, no retraining.
Why it matters
The authors list five directions: auditing models for hidden problems or capabilities, like backdoors, without running inference; cutting optimization cost with one-shot initialization instead of full training; steering or unlearning with edits of a few weights, avoiding the catastrophic forgetting fine-tuning often causes; a more native path to recursive self-improvement than a model launching its own training; and cross-architecture merging, grafting a circuit between models instead of averaging identical ones. They say they are building a benchmark and harness for it.
The timing: the post landed the same day Anthropic's amended S-1/A filed that its internal estimates point to imminently meeting its own RSI threshold. If models can inspect, debug, and improve weights, the line between improving a model and self-improvement gets thinner. The authors' framing is calmer: a spark, with real limits that include context length and harness inefficiency, in a field they expect to become a rigorous, reproducible engineering discipline.
Source: rikiyat.github.io
More news
Have something in mind?
Start a conversation