English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Research in mechanistic interpretability seeks to explain behaviors ofmachine learning models in terms of their internal components. However, mostprevious work either focuses on simple behaviors in small models, or describescomplicated behaviors in larger models with broad strokes. In this work, webridge this gap by presenting an explanation for how GPT-2 small performs anatural language task called indirect object identification (IOI). Ourexplanation encompasses 26 attention heads grouped into 7 main classes, whichwe discovered using a combination of interpretability approaches relying oncausal interventions. To our knowledge, this investigation is the largestend-to-end attempt at reverse-engineering a natural behavior "in the wild" in alanguage model. We evaluate the reliability of our explanation using threequantitative criteria--faithfulness, completeness and minimality. Though thesecriteria support our explanation, they also point to remaining gaps in ourunderstanding. Our work provides evidence that a mechanistic understanding oflarge ML models is feasible, opening opportunities to scale our understandingto both larger models and more complex tasks.

Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small