Causal mediation analysis / Vig et al. 2020 (causal mediation analysis)

Findings

• IC-1078LLaMA models (7B through 65B) exhibit gender bias in language generation, coreference resolution, and sentence likelihood, with stereotypical associations driving predictions [builds-on]
• 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 [builds-on]
• IC-1630Llama and Pythia models represent entity-attribute bindings via additive binding id vectors that form a continuous subspace with metric structure [primary]
• IC-1631Binding id mechanism fidelity increases with model size in both Llama and Pythia families [primary]
• IC-1632Tulu-13B uses a direct binding mechanism rather than binding ids for multiple-choice question tasks [primary]
• IC-395Most mainstream LLMs exhibit positive ADCE across five tasks, indicating reliance on deep structure for problem-solving, with ADCE strongly correlated with accuracy (r² > 0.7) [primary]
• IC-745Function vectors for simple list-oriented tasks can be algebraically combined via addition and subtraction to produce new vectors that trigger composed tasks, sometimes outperforming 10-shot ICL. [primary]
• IC-982In Stable Diffusion's UNET, visual attribute knowledge is distributed across multiple components with attribute-specific patterns, concentrated more in the up-block, and cross-attention layers are not the primary causal states [primary]
• IC-983In Stable Diffusion's CLIP text-encoder, knowledge about all visual attributes is localized to a single causal state: the first self-attention layer at the last subject token [primary]