Mechanisms are Transferable: Data-Efficient Low-Resource Adaptation via Circuit-Targeted Supervised Fine-Tuning
This addresses the challenge of data-efficient adaptation for low-resource languages, though it appears incremental as it builds on existing methods like CD-T and focuses on specific cross-lingual tasks.
The paper tackles the problem of adapting large language models to low-resource languages with scarce labeled data, proposing Circuit-Targeted Supervised Fine-Tuning (CT-SFT) to improve cross-lingual accuracy over full fine-tuning while updating only a small subset of parameters, as demonstrated on datasets like NusaX-Senti and XNLI.
Adapting LLMs to low-resource languages is difficult: labeled data is scarce, full-model fine-tuning is unstable, and continued cross-lingual tuning can cause catastrophic forgetting. We propose Circuit-Targeted Supervised Fine-Tuning (CT-SFT): a counterfactual-free adaptation of CD-T (Contextual Decomposition Transformer) that uses a label-balanced mean baseline and task-directional relevance scoring to identify a sparse set of task-relevant attention heads in a proxy-language checkpoint, then transfer learns to a target language by updating only those heads (plus LayerNorm) via head-level gradient masking. Across NusaX-Senti and XNLI, CT-SFT improves cross-lingual accuracy over continued full fine-tuning while updating only a small subset of model parameters. We find an editing-preserving trade-off: harder transfers favor editing circuit heads, while easier transfers often favor near-zero (i.e., low-relevance heads) updates, preserving the source mechanism. CT-SFT also substantially reduces catastrophic forgetting, preserving proxy/source-language competence during transfer.