TopKLoRA
An interpretable parameter-efficient fine-tuning method that activates only the k largest adapter latent variables at each token, turning fine-tuning changes into an analyzable sparse circuit used for backdoor localization and ablation.
Timeline (last 90 days)
TopKLoRA is a LoRA variant that activates only the k largest adapter latent variables at each token and sets the rest to zero.
TopKLoRA introduces no additional reconstruction loss; sparsity acts directly on the fine-tuning parameters themselves, and the object of analysis is exactly the circuit the model actually executes.
TopKLoRA extends the preliminary proposal of Masiak et al. (2025).
For each model organism, the research team located and ablated a backdoor implementation circuit that is both necessary and sufficient.
Compared with the dense LoRA baseline, the sparse circuits produced by TopKLoRA are up to an order of magnitude smaller.
After removing the malicious circuit, instruction-following performance showed no noticeable loss.
The paper proposes a technique that uses TopKLoRA latent variables and gradient routing to construct localized circuit model organisms, forming a closed-loop validation of construction first and search second.
Current validation is concentrated on a specific sleeper-agent backdoor scenario and three small-to-medium-scale models; whether the method generalizes to larger models and more attack types still requires follow-up work.
The TopKLoRA code is open-sourced on GitHub in the repository interpretable-finetuning/TopKLoRA.
All Facts (9)
TopKLoRA is a LoRA variant that activates only the k largest adapter latent variables at each token and sets the rest to zero.
50%UnverifiedTopKLoRA introduces no additional reconstruction loss; sparsity acts directly on the fine-tuning parameters themselves, and the object of analysis is exactly the circuit the model actually executes.
50%UnverifiedTopKLoRA extends the preliminary proposal of Masiak et al. (2025).
50%UnverifiedFor each model organism, the research team located and ablated a backdoor implementation circuit that is both necessary and sufficient.
50%UnverifiedCompared with the dense LoRA baseline, the sparse circuits produced by TopKLoRA are up to an order of magnitude smaller.
50%UnverifiedAfter removing the malicious circuit, instruction-following performance showed no noticeable loss.
50%UnverifiedThe paper proposes a technique that uses TopKLoRA latent variables and gradient routing to construct localized circuit model organisms, forming a closed-loop validation of construction first and search second.
50%UnverifiedCurrent validation is concentrated on a specific sleeper-agent backdoor scenario and three small-to-medium-scale models; whether the method generalizes to larger models and more attack types still requires follow-up work.
50%UnverifiedThe TopKLoRA code is open-sourced on GitHub in the repository interpretable-finetuning/TopKLoRA.
50%