Bias in Bios

anchor

Findings

• IC-1137Stable Diffusion represents certain concepts primarily through specific named exemplars rather than abstract category features [eval]
• IC-1139Stable Diffusion encodes social biases in its internal concept representations that are not always visually apparent [eval]
• IC-1584LLaMA-7B and GPT-J-6B fail to interpret textual emphasis markers, with marked prompting degrading performance substantially [eval]
• IC-1585LLaMA-7B and GPT-J-6B exhibit positional bias in instruction following: zero-shot performance varies significantly when the instruction is moved from after to before the context [eval]
• IC-1586In LLaMA-7B, steering all attention heads degrades JSON format accuracy below zero-shot, while steering a subset of 50-100 heads selected via multi-task profiling raises it to 96.64; performance varies dramatically across the 32 layers and individual heads [eval]
• IC-161Linear probes on Pythia-70m and Gemma-2-2b trained on the ambiguous Bias in Bios set rely on gender as a spurious feature, with gender accuracy far exceeding profession accuracy [eval]
• IC-1630Llama and Pythia models represent entity-attribute bindings via additive binding id vectors that form a continuous subspace with metric structure [eval]
• IC-1631Binding id mechanism fidelity increases with model size in both Llama and Pythia families [eval]