Researchers expected artificial intelligence to struggle with the opaque internals of another neural network. Instead, experiments from MIT CSAIL showed a model reading another network’s weights and rewriting them to fix failures without any retraining, Živě.cz reported.
The sharpest contrast came with Qwen3.5-0.8B, a small language model prone to loops that completed only 7.9% of the MATH-500 math test. After another model, GPT-6 Astra from OpenAI, edited 131,072 weights—just 0.016% of the network—the success rate jumped to 48.1% across 50 test runs.
Reading Inside the Black Box
Neural networks are black boxes built from billions of numbers called weights, which determine how strongly one artificial neuron influences another. These weights emerge from training, so no human wrote them or knows exactly why they hold their values. Until now, people had to sort through that confusion slowly, finding connections one by one over years. Now a language model itself found the spot breaking a network and changed it to work as intended.
From Loops to Answers in Two Layers
Astra traced Qwen3.5-0.8B’s failures to a specific pattern: the 15th and 19th layers were adding a direction that pulled the model into repetition until it exhausted the 32,768-token limit. Astra deleted that direction from both layers, and the math performance soared.
Matthew Honnibal, creator of the spaCy natural-language library, argued that for Qwen, Astra did not read raw numbers but wrote Python code and ran the model. Other critics noted the approach resembles abliteration, an existing technique for removing unwanted behaviors from models. The study’s authors counter that this time the model itself identified where to intervene from the task description alone.
A Car That Learned to Back Up
In a separate test, a neural network controlling a car in a 2D simulation failed all 256 attempts to cross a ramp behind a red line. Astra received the network’s 171,664 weights, a scene description, and a recording of one failed run, with no tools allowed. It recognized the car lacked momentum and inserted a shortcut across neurons in each layer to track position and speed: back up briefly, then accelerate. After changing 3.5% of the weights, the car succeeded in all 1024 test runs.
Why Gemma 3 Stayed Broken
The same procedure failed to improve Google’s Gemma 3 270M, which does not loop but simply cannot do math. Astra found no faulty component to repair, and MATH-500 scores improved only marginally, from 6.8% to 7.7%. The authors concluded: “Weight editing could remove an obstacle that hid an ability, but could not add an ability the model didn’t have.”
The Scale Barrier
Astra has so far read raw weights only in small networks. From 17,253 numbers in a game-playing network, it correctly identified which of four games the network played every time. It also recognized whether image generators drew circles, triangles, or squares by reading weight geometry rather than output. The largest network Astra edited had roughly 0.8 billion weights, while large models contain orders of magnitude more, and the authors acknowledge limited context length currently holds them back.
Layer Tweaks Fix Repetition Loops
Which layers caused Qwen to loop?
The 15th and 19th layers added a direction that pulled the model into repetition. Astra removed that direction from both layers to restore performance.
How did the weight changes compare between the car and language model tests?
Astra changed 3.5% of weights in the car network to achieve 1024 successful runs, versus just 0.016% of weights—131,072 numbers—in Qwen3.5-0.8B to raise math scores from 7.9% to 48.1%.
What token limit did the looping model hit?
Qwen3.5-0.8B exhausted the 32,768-token limit on its responses while stuck in loops, failing to complete answers in 93% of math tasks.
The open question remains whether any AI can perform this kind of repair on the massive networks that power today’s most capable systems.
Keep reading