Causal Tracing

anchor

Findings

• IC-1078LLaMA models (7B through 65B) exhibit gender bias in language generation, coreference resolution, and sentence likelihood, with stereotypical associations driving predictions [supporting]
• IC-1079Causal tracing reveals that mid-upper MLP layers (layers 18–25 in 7B) are the primary mediators of stereotypical gender bias in LLaMA, while the last layers show negative coefficients that counter the bias [primary]
• IC-636Syntactic phenomena (determiner-noun and subject-verb agreement) localize to the same topmost-layer MLP neurons as factual information in BERT, GPT-2, and Llama-2 [supporting]
• IC-639The causal tracing pattern of MLP at early layers and attention at late layers is not stable across factual and syntactic phenomena in GPT-2 XL [primary]