MIB: A Mechanistic Interpretability Benchmark
Northeastern University · Stanford University · Ai2 + UMD · Computer Science Departmen, Technion-Israel Institute of Technology · ML Alignment & Theory Scholars · Brown University · Technion · University of Amsterdam · Technion - Israel Institute of Technology, Technion · MIT · University of Cambridge · ETH Zürich · University of Zagreb · Technion, Technion
PDF 由论文原始站点提供，PaperCompass 不保存论文文件。
摘要
How can we know whether new mechanistic interpretability methods achieve real improvements?In pursuit of lasting evaluation standards, we propose MIB, a Mechanistic Interpretability Benchmark, with two tracks spanning four tasks and five models. MIB favors methods that precisely and concisely recover relevant causal pathways or causal variables in neural language models. The circuit localization track compares methods that locate the model components---and connections between them---most important for performing a task (e.g., attribution patching or information flow routes). The causal variable track compares methods that featurize a hidden vector, e.g., sparse autoencoders (SAE) or distributed alignment search (DAS), and align those features to a task-relevant causal variable. Using MIB, we find that attribution and mask optimization methods perform best on circuit localization. For causal variable localization, we find that the supervised DAS method performs best, while SAEs features are not better than neurons, i.e., non-featurized hidden vectors. These findings illustrate that MIB enables meaningful comparisons, and increases our confidence that there has been real progress in the field.