The method automates manual model comparison tasks
I'm big believer in just doing the obvious thing. Turns out you can diff two models by just asking an agent to do it!
New research update from the Google DeepMind Language Model Interpretability team.
We build and evaluate dead simple open-ended model diffing agents tasked with studying the behavioural differences between two models, and find them to be promising in practice.
Users praise DeepMind's simple AI agents for model behavior diffing because straightforward methods prove effective at surfacing differences worth investigating.
Our diffing agents are conceptually much simpler than most other methods researchers use to study model differences - they are simple LLM scaffolds with access to tools providing model rollouts from the pair of models on prompts.
More details in the blogpost: https://www.lesswrong.com/posts/qi4mNbZYAFDYwfRba/building-and-evaluating-model-diffing-agents
Work with @JoshAEngels and @NeelNanda5
New research update from the Google DeepMind Language Model Interpretability team.
We build and evaluate dead simple open-ended model diffing agents tasked with studying the behavioural differences between two models, and find them to be promising in practice.
@NeelNanda5
We think more mature versions of tools of this form may play a useful role in understanding and improving model behaviour. For instance, they may help answer the question: what is the behavioural effect of training models with subtly different constitutions?
We validate our diffing agents work through evaluations on a number of settings where there is either a known lack of difference or known true difference. We also compare them directly with single model auditing agents tasked with finding strange behaviours.
Our diffing agents find interesting real differences between various production Gemini models.
@bilalchughtai_ Nice work! seems very useful in surfacing behavior differences to investigate deeper! I notice your lesswrong values "novelty" - I think @mbodhisattwa 's work https://arxiv.org/pdf/2507.00310 might be useful quantifying surprise in behavior differences
@celestepoasts @NeelNanda5 omg
@bilalchughtai_ i did it the hard way https://kwbench.github.io/insights
staggering how different these models are
@NeelNanda5 nice!
@NeelNanda5 could steering a single feature instead of finetuning give you a cleaner organism here? Or would that just produce a different set of salient changes?
@NeelNanda5 Sometimes the obvious method is just the best method. especially in AI world where folks assume the obvious way did not work because everyone else tried it and failed. very cool work was relevant to ongoing discussion thanks
@bilalchughtai_ diffing sidesteps the hardest part of interp cuz the agent has a contrast to anchor on. 'these two differ here' is way easier to ground than 'this is what the model does'
@NeelNanda5 sometimes the best solution is the one you almost skipped bc it seemed too simple
@NeelNanda5 the obvious thing is always the one nobody does until someone proves it works
