A new interpretability framework can trace an AI model's refusal behavior back to specific training pairs — and explain the mechanism in between.
Researchers introduced Symbolic Mechanistic Data Attribution (SMDA), a framework that fits a regression model over sparse autoencoder features to represent a target behavior as a readable policy, then mathematically breaks down how each supervised fine-tuning example shifts that policy. Tested on Meta's Llama-3.2-3B-Instruct with 200 training pairs, the tool surfaced gaps in the model's safety behavior for categories like religious stereotyping. It also identified training pairs whose main influence landed on unintended features — a subtle form of misfiring that existing methods tend to miss.
Most data attribution tools answer a narrow question: which training example lit up this circuit? SMDA asks the harder question of how training data shapes what a model decides to do. That distinction matters as regulators and auditors push for explainability in AI safety systems — knowing a model refuses harmful requests is less useful than knowing why, and whether the mechanism is stable.
The researchers position SMDA as sitting between black-box influence functions, which are scalable but opaque, and manual circuit analysis, which is precise but slow. Whether it stays useful as models grow past 3 billion parameters is the open question — small-model interpretability results have a poor track record of traveling upward.