Reading the parameters, not the outputs
Four researchers at MIT CSAIL — Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab — published results on 2 November from a setup they call Weight-Space Programming: a frontier model, OpenAI’s GPT-6 Astra, is handed another network’s raw parameters and asked to interpret or rewrite them, with no gradient descent anywhere in the loop.
The interpretation experiments come first. Given the 17,253 weights of four reinforcement-learning policies, the model named the game each one played — Breakout, Pong, Pac-Man and Freeway — for all four. Given a 23,824-weight diffusion model, it identified which shape the model generated. From only the middle four transformer blocks of 7.2M-parameter language models, it identified whether the model had been trained on Python, romaji or English, and located the copying heads.
Editing, and writing from zero
The editing results are the sharper ones. On a Qwen3.5-0.8B model, the authors report that the model edited 131,072 weights — 0.016% of the parameters — and, allowed 50 forward passes to check its work, moved MATH-500 from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%. The share of MATH-500 answers that ran into the 32,768-token output limit fell from 93% to 33.1%, which suggests the edit mostly stopped the model looping.
In a physics control task, editing 3.5% of a 171,664-weight network produced a policy that succeeded in all 1,024 test episodes, where the original network had succeeded in none of 256. Starting instead from all-zero weights, one-shot networks scored 517.4 on Breakout against 15.8 for a policy-gradient agent trained on 100 million steps and 25.3 at 500 million. On 9x9 Go, a 1.48M-weight network written with zero forward passes won more than half its games against two smaller trained networks; with 50 forward passes it beat an equal-sized network trained on 20 million games.
Where it failed
The write-up reports its negative results too. Asked to pick out a sycophantic Gemma 3 270M model from weights alone, the model could not; it succeeded only once allowed to inspect attention patterns. And on Gemma 3 270M, which the authors say rarely loops, the editing method produced no meaningful gain — MATH-500 moved from 6.8% to at most 7.7%. That pattern is consistent with the token-limit figure: the large gains came where there was looping to remove.
Every number here is the authors’ own, measured with a model built by OpenAI, and none has been independently reproduced.
What to watch
If the method holds up outside this write-up, it cuts in two directions at once. It is a new interpretability tool — a model that can read another model’s parameters and say what they do. It is also a capability result: the same access lets the model change behaviour at a scale of a hundred thousand weights without training. The useful next step is a replication on models the authors did not choose.
