Interpretable Sparse Fine-Tuning Based on TopKLoRA Enables Faithful Circuit Discovery

The researchers propose TopKLoRA, an interpretable-by-design sparse fine-tuning method, and use it to localize and ablate the circuit necessary and sufficient for implementing a backdoor in a sleeper agent backdoor model. The method activates only the k largest adapter latent variables for each token; unlike SAEs, it is trained solely with the downstream loss, and the optimization pressure drives it to learn latents that are functionally causally relevant. The authors fine-tune gemma-2-2b, Qwen2.5-1.5B, and Qwen2.5-7B on a poisoned instruction-following dataset to build backdoor models; the resulting sparse circuits are up to an order of magnitude smaller than dense LoRA baselines and allow surgical ablation of the malicious capability without sacrificing instruction-following performance. The authors also propose a technique for constructing localized circuit models using TopKLoRA latents and gradient-routing to validate circuit search results. Code is open-sourced at https://github.com/interpretable-finetuning/TopKLoRA。