Back to Devoured - June 26, 2026
DEVOURED
We removed an LM's ability to speak German

We removed an LM's ability to speak German

AI research Threadreaderapp
Goodfire AI successfully removed a 67M-parameter model's ability to speak German by fine-tuning a single internal weight component.
What: The experiment used 'parameter decomposition' to isolate a specific sub-component of the model's weights associated with German, demonstrating targeted model editing that avoids the collateral damage seen in LoRA methods.
Why it matters: This validates a future where model behavior can be patched with surgical precision, reducing the need for costly full-model retrains.
Deep dive
  • The team targeted a single scalar factor within a decomposed weight matrix.
  • Achieved targeted functionality removal using only 4 training tokens.
  • Comparison with LoRA showed that the parameter decomposition method preserved other languages, whereas LoRA tended to degrade performance in French, Spanish, and Italian.
  • Validates the concept of interpretability-driven model editing.
Decoder
  • Parameter decomposition: A technique to divide weight matrices into smaller, sparsely-activating sub-components to increase interpretability.
  • LoRA (Low-Rank Adaptation): A method for fine-tuning large models by freezing most weights and training a smaller set of low-rank matrices, often leading to unintended side effects on unrelated capabilities.
Original article

We removed an LM's ability to speak German by fine-tuning on only 4 German tokens.

As part of a 1-day hackathon with our product Silico, we removed a 67M-parameter language model's ability to predict German text, by tuning only a scalar factor on one subcomponent of the weights.

This was an early exploration in fine-tuning with parameter decomposition, our method which divides a model's weight matrices into interpretable, sparsely-activating components.

We picked German as it seemed to be the model's strongest non-English language.

https://twitter.com/3079387847/status/2051717264286609516

We benchmarked vs LoRA fine-tuning. Our edit matched its German removal with far fewer tokens.

Strikingly, it also left other languages almost untouched.

The LoRAs often wrecked French, Spanish, Italian, and sometimes English, while our edit mostly left them alone.

In a sense this is cheating: we're indirectly exploiting the tokens from when we did the parameter decomposition and interpreted the resulting subcomponents.

But if our decomposition is good, that cost can be amortized over arbitrarily many tasks & component edits.

Plus, that interpretability lets us notice and fix problems.

E.g.: initially we tuned the top 16 German-related components, but their labels showed most were about foreign languages in general.

So we narrowed to the single component for German alone, improving precision.

This is an early demo of how parameter decomposition could enable targeted, predictable model editing.

Details on this experiment: lesswrong.com/posts/ieoWstub…

If you want to run experiments on your model too, learn more and request access to Silico: goodfire.ai/silico

Correction: a plotting error caused the bars in the plot of off-target effects to display at 0.01 nats above the true means. The corrected plot is below.

Explore all 53 articles in today's digest

Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi