For years, the deal with neural networks has been simple. You train them. Millions of small changes done by an optimizer until numbers inside stabilize to something useful. People don’t manually write these numbers down since that would be impossible. However, according to the recent blogpost on the CSAIL website from MIT, an LLM is able to do this. They claim that a state-of-the-art model can analyze a raw network and determine its functioning before rewriting it using the same process without any optimization. What does it mean? They managed to fix a small language model by changing only 0.016% of its weights.
Also read: WiCi lets you connect an external GPU to your laptop over Wi-Fi: Here’s how
This phrase was coined by the MIT team to refer to the ability of an LLM to read and write the weights of a neural network using reasoning and not using optimisers such as Adam. The weights are basically the figures that contain what the neural network has learned, similar to a mixing desk that contains millions of faders. While training modifies each of the faders slightly, this technique analyses the desk and modifies only the faders that matter.
Well, certainly on small networks. They trained four identically sized networks on four different games. They then fed the text representation of the 17,253 weights in each of them to GPT-6 Astra without any mention of the name of the game. In all cases, it could tell the correct game perfectly well. This was because of how Breakout’s network identified itself.
Also read: AI bubble crack: Why major tech companies are under crazy pressure
Through identifying and eliminating the loops. Qwen3.5-0.8B has a tendency of getting into a loop and that is highlighted in the model card. Loops were identified in one of the directions created by the MLPs in layer 15 and 19 and eliminated from among 131,072 weights or 0.016% of the model. The score for MATH-500 shot up from 7.9% to 48.1%, while IFEval went up threefold to 62.3%. All things get destroyed by loops, not just math.
However, there is always an exception to the rule. No training doesn’t mean no feedback. Astra was able to do up to 50 forward passes on training tasks. The gains were insignificant at 10 and 25 passes.
For smaller games, yes. Based on the rules of three MinAtar games, Astra came up with the weights in a single response, without any training and without running tests. It scored 517.4 in Breakout compared to PPO trained for 500 million steps which scored 25.3. The other two games Seaquest and Asterix were more competitive, with PPO trained for 500 million steps still winning though. For Go 9×9, a 1.5 million weight Astra network outperformed an equal-size PPO network trained on 20 million games.
There’s enough evidence for that, and that is exactly what the authors claim. Astra did not find a fawning Gemma 3 270M version based on weights and achieved success after noticing attention patterns. Astra was not able to improve Gemma 3 270M skills in math either as the results were always close to 7 percent. Editing helped to uncover the hidden capability through the removal of Qwen loops. It does not allow adding any new capability though.
Training takes time and costs money. The authors believe it could be useful in auditing backdoors in ML models, deleting abilities without using any data, and transferring circuits from one model to another. This is just projection and not reality. For Indian laboratories and startups that do not have enough GPUs, a simple fix works better than retraining.
Also read: Why Alphabet shares crashed: How Google AI Mode is eating into search ad revenue