Cross-Architecture Model Diffing with Crosscoders

"Cross-Architecture Model Diffing with Crosscoders: Unsupervised Discovery of Differences Between LLMs" is interpretability research. The deeper interest, for the operator, is that it's the first method we've seen that can surface — without being told what to look for — where two LLMs differ in their internal representations.

Today, when a team decides to upgrade a model in production, the question 'what actually changed' is answered with held-out evals, anecdotal spot-checks, and faith. "Cross-Architecture Model Diffing with Crosscoders" is a principled answer to that question at the level of features the model represents, not just outputs the model produces. That's a different category of evidence than anything an operator currently has access to.

At the technical level, crosscoders extend sparse-autoencoder methodology to operate across two models simultaneously — learning a shared feature space and surfacing where the two models' internal representations diverge. The 'unsupervised' word in the title is the load-bearing one: the operator doesn't have to specify which capabilities or behaviors to compare. The technique surfaces them.

That matters because the failure modes of a model swap are almost always the ones nobody thought to check. Held-out eval suites are designed against the failure modes the team has already imagined. "Cross-Architecture Model Diffing with Crosscoders" is the first method we've read that addresses the unimagined ones.

Take the typical model-upgrade decision in a real company. Engineering runs a regression suite, product runs a vibe check on a few prompts, leadership signs off. Six weeks later something subtle is off — tone has shifted, edge-case handling has changed, refusal patterns are different — and nobody can tell whether the model changed or the workflow drifted around it.

That conversation has played out in nearly every production AI deployment we've watched in the past eighteen months. The team doesn't have an instrument that can answer the question. "Cross-Architecture Model Diffing with Crosscoders" is the first paper we've read that points at where the instrument eventually comes from.