saintsauce/unified-vlm-steering-eval

Unified VLM steering eval

Activation steering of unified vision-language models: steered text and image generations, steering vectors and judge scores. One folder per model; prompts/ is shared. Steering is h <- h + alpha * v_hat everywhere.

model	run	text rows	image rows	concepts with images	published
uniar/	av_v2	21,560	21,560	age, chaos, cleanness, emotion, near_far, size, spatial_lr	2026-09-11

Layout per model: manifest.json (exact config), vectors/<concept>/{txt,img}.pt, generations/sweep_text.csv, generations/sweep_image.csv + generations/images/<concept>/, judge/ (when scored). Quadrants: txt2txt, img2txt (text out), img2img, txt2img (image out); <source vector>2<output>.