Researchers at MIT's Computer Science and Artificial Intelligence Laboratory say a general-purpose language model can now work out what a neural network does by reading its weights, and can change those weights by hand to repair or build networks. In a post published on November 2, 2026, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab call the idea Weight-Space Programming. They used OpenAI's GPT-6 Astra as the editor and did not let it write training loops or use gradient descent, so every change came from the model reasoning about the numbers.
The reading experiments came first. The team trained four small policies on its own grid versions of Breakout, Pong, Pac-Man and Freeway, giving each exactly the same 17,253-weight architecture, inputs and actions. Handed the raw weights as text, and without being told which games were candidates, Astra named the right game for all four. It also identified which of three diffusion models drew circles, squares or triangles, and, given only the middle blocks of three 7.2-million-parameter character-level transformers, said which had been trained on Python, English or Japanese written in romaji, picking out every copying head along the way.
Writing proved harder but still produced striking numbers. Given a car-control policy that always stalled on a ramp in the Kinetix simulator, Astra explained the failure and rewrote 3.5% of the weights in a single reply; the researchers report the edited policy succeeded in all 1,024 test episodes. Asked to write networks from scratch for three MinAtar games, using only the rules, source code and network layout, it beat agents trained with PPO for 100 million steps in every game. In Breakout its network scored 517.4 against 25.3 for a 500-million-step PPO agent, though it trailed that stronger baseline in Seaquest and, narrowly, in Asterix.
The headline result involves a real language model. The team gave Astra a Python session over the weights of Qwen3.5-0.8B, practice problems from GSM8K and MATH, and a capped number of forward passes. Astra traced the model's habit of looping until it hit its token limit, a weakness Qwen's own model card mentions, to one direction written by MLPs in layers 15 and 19, and removed it from 131,072 weights. With a budget of 50 forward passes, MATH-500 rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1% and the IFEval instruction-following score from 21.0% to 62.3%. With 10 or 25 passes the gains were far smaller.
The authors are open about where it broke down. On Gemma 3 270M, a smaller model that rarely loops, Astra's edits moved MATH-500 by about a point at most, which the researchers read as evidence that the method removed an obstacle rather than adding mathematical skill. In a test of whether Astra could tell an honest fine-tuned Gemma from a sycophantic twin, it failed in all four attempts from the weights alone and succeeded only once it was shown attention patterns. VentureBeat, which covered the work, reports that the strongest Qwen run took 158 conversational turns, 80 Python executions and 20.3 hours, including several hours lost to a service outage.
The team has published full logs of the prompts, Astra's replies, code and evaluations on GitHub, and says a Hugging Face researcher rebuilt and re-evaluated the edited Qwen model from those logs. It is also building a benchmark and harness for this kind of task. The work is a blog post rather than a peer-reviewed paper, and VentureBeat notes it follows earlier targeted weight-editing methods such as ROME and MEMIT, which were built specifically for editing facts.
Reactions have split along familiar lines. VentureBeat and the outside researchers it quoted saw a possible new way to maintain, debug and improve models, while PauseAI's Maxime Fournes called the direction reckless because it could speed up capability gains. The authors themselves list recursive self-improvement among possible future uses, but nothing in the experiments shows a model improving itself, and the gains so far are confined to small networks and one tiny language model under the team's own evaluation settings.