Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces SCALPEL, a framework for selectively ablating capabilities in LLMs by representing them as low-rank parameter subspaces and using LoRA adapters to reduce the model's ability to distinguish correct from incorrect answers on specific tasks. This approach allows for fine-grained capability removal without affecting other capabilities, addressing the limitations of coarse-grained methods that assume direct mapping between capabilities and modules. Experiments on diverse tasks demonstrate SCALPEL's effectiveness in removing target capabilities while preserving general language modeling abilities, revealing the low-rank structure of capabilities and enabling targeted parameter-space interventions.
LLMs' capabilities can be surgically removed via low-rank parameter edits, offering a more precise alternative to coarse-grained module ablation.
Large language models excel across diverse domains, yet their deployment in healthcare, legal systems, and autonomous decision-making remains limited by incomplete understanding of their internal mechanisms. As these models integrate into high-stakes systems, understanding how they encode capabilities has become fundamental to interpretability research. Traditional approaches identify important modules through gradient attribution or activation analysis, assuming specific capabilities map to specific components. However, this oversimplifies neural computation: modules may contribute to multiple capabilities simultaneously, while single capabilities may distribute across multiple modules. These coarse-grained analyses fail to capture fine-grained, distributed capability encoding. We present SCALPEL (Selective Capability Ablation via Low-rank Parameter Editing for Large language models), a framework representing capabilities as low-rank parameter subspaces rather than discrete modules. Our key insight is that capabilities can be characterized by low-rank modifications distributed across layers and modules, enabling precise capability removal without affecting others. By training LoRA adapters to reduce distinguishing correct from incorrect answers while preserving general language modeling quality, SCALPEL identifies low-rank representations responsible for particular capabilities while remaining disentangled from others. Experiments across diverse capability and linguistic tasks from BLiMP demonstrate that SCALPEL successfully removes target capabilities while preserving general capabilities, providing fine-grained insights into capability distribution across parameter space. Results reveal that capabilities exhibit low-rank structure and can be selectively ablated through targeted parameter-space interventions, offering nuanced understanding of capability encoding in LLMs.