# Goodfire presents mechanistic shaping, a training method built on model internals

> Original title: We cut through the mystery around interpretability in training.

- Source: Goodfire
- Published: 2026-11-05T17:20:50.000Z
- Bad Influence: https://badinfluence.news/items/na6w2pjkzyxlckgqo1yf8jvx7
- Original: https://x.com/GoodfireAI/status/2118393112737591387

## Summary

Goodfire says it has developed mechanistic shaping, a framework that turns model internals into training feedback across domains, problems, and modalities. The company says it has used the method to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. Goodfire also warns that training against a model's internals can make it harder to monitor, and says the thread explores both the technique and this concern.

## Post

We cut through the mystery around interpretability in training.

A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities.

We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment.

But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵

[Video](https://badinfluence.news/api/img-proxy?u=https%3A%2F%2Fpbs.twimg.com%2Famplify_video_thumb%2F2118393092026028032%2Fimg%2FBrLWZKd1kn32gHQT.jpg&mode=thumb&exp=1794096000&sig=e82dd9a18a18e336)