Direct Weight Edits Raise Qwen’s MATH-500 Score From 8% to 48.1%

Károly Zsolnai-Fehér describes experiments in which GPT-6 Astra was given access to another neural network’s weights: the numbers that encode its learned connections. Those numbers are usually difficult to interpret. Astra nevertheless identified which games small networks had been trained to play, edited a network that was failing at a driving task, and changed a language model’s behavior enough to raise its score on a math benchmark.

The math result is striking, but the explanation matters. Zsolnai-Fehér says Astra did not teach Qwen mathematics. It recognized that Qwen was not necessarily bad at math, but could get caught in repetitive “thinking loops,” revisiting the same calculation instead of moving on. In the example shown, the model repeatedly reconsiders how to interpret and calculate an expression. Astra tested the model repeatedly, observed what was going wrong, identified parts of the network responsible for the looping behavior, and edited those parts to suppress it. The result is presented as unblocking an existing capability, not adding mathematical knowledge.

The gain is large on the benchmark shown. The narrower claim is that, in this case, editing weights helped a model use an existing ability by reducing a behavior that interfered with solving the task.

The other demonstrations involve different networks and tasks. For game identification, Astra inferred that small networks had been trained to play simplified versions of Breakout, Pong, and Pac-Man by inspecting their weights. That shows that information about a network’s training can sometimes be recovered from its parameters; it is not the same as making the network better at playing. In a driving simulation, Astra edited a network whose car got stuck on a ramp. The displayed result shows the car driving over the ramp after the edit, with a 100% success label. These are separate experiments, not three capabilities combined into one procedure.

The distinction between “can do all three” and “can do all three at once” is central to Zsolnai-Fehér’s account. Reading, fixing, and improving were demonstrated in separate experiments, mostly on small systems with custom setups.

Nor is the ability to identify a game unique to Astra. A graphic credited to Goswami (2026) and Willison (2026) shows game-identification examples involving DeepSeek V4.1 Flash and Qwen 3.8 27B, including Breakout, Pong, Pac-Man, Freeway, and Asterix. Zsolnai-Fehér says these independent experiments use smaller, free, open models, including Qwen models that can be run at home. Game identification, in other words, is not exclusive to Astra.

The results also vary across tasks and models. A graphic reports that Astra’s attempt to improve Gemma 3 moved its MATH-500 score from 6.8% to, at best, 7.7%. It also shows Astra identifying a sycophantic model from weights alone in zero of four cases. These results temper the successful demonstrations: the source presents task-specific capabilities with uneven outcomes, not automatic improvement whenever one model edits another’s weights.

Astra also wrote a game-playing neural network from scratch by setting weights directly. The displayed code shows an architecture with layers of 600, 32, and 6 units, with 815 of its 20,486 weights set. Zsolnai-Fehér contrasts this with reinforcement learning, which typically uses many rounds of trial and error. In the comparison shown, PPO uses 500 million training steps, while Astra’s approach uses zero.

In the games and setup shown, Astra’s one-shot Breakout score is far above the two PPO comparisons. Zsolnai-Fehér says the researchers are building a benchmark for this kind of “weight-space programming,” which could provide a way to test the approach more systematically.