ARTIFICIAL INTELLIGENCE · EXPLAINERS

From training models to editing their weights

How AI models learn, how fine-tuning and distillation change them, and a new idea: editing their weights directly.

By Ebinesh A · · 2 min read

An AI model’s weights are the numbers that shape its responses. During initial training, the model processes huge amounts of data and repeatedly adjusts those numbers to reduce errors. This is how it learns patterns in language and other data.

Fine-tuning: keep training for a purpose

Once a model has been pretrained, developers can continue training it on examples for a particular task or domain. This is fine-tuning: the model learns from new examples, and its weights change to better fit the intended use.

Distillation: pass behavior to a smaller model

In 2015, Geoffrey Hinton, Oriol Vinyals, and Jeff Dean described a practical way for a smaller “student” model to learn from a larger “teacher.” Rather than starting from scratch, the student imitates the teacher’s outputs. The aim: preserve useful behavior in a model that costs less to run.

Weight-space programming: edit the model itself

A newer research direction asks an LLM to inspect another model’s weights and directly edit a small part to address a specific behavior. That skips the usual gradient-based training loop. The diagram shows how it compares with fine-tuning and distillation.

Hover to loop
Follow the orange cells: learning from examples, learning from a teacher, or applying a direct edit. The grids illustrate the mechanisms, not the number of weights changed.

In a recent MIT CSAIL experiment, researchers asked GPT-6 Astra to inspect Qwen3.5-0.8B and target a repetitive reasoning loop. They reported changing about 0.016% of its weights and raising its MATH-500 score from 7.9% to 48.1%.

This was not conventional fine-tuning: the model used a Python environment and a limited number of forward passes on practice problems, without gradient descent. The goal was to patch a narrow behavior directly.

Hover to loop
An edit is a candidate repair that must be evaluated. The bars show the researchers’ reported result for Qwen3.5-0.8B, not a guarantee for other models.

Final thought

Fine-tuning learns from examples; distillation transfers behavior from a teacher; weight-space programming tries a direct repair. This is still an early research result, and direct edits can fail when a model lacks the underlying ability. But the idea is compelling: for some specific failures, perhaps we can debug a model by changing the small part of its weights behind the problem.