saintsauce/uniar-steering-eval

UniAR Steering Eval — text quadrants (samples only)

Activation-steering text generations from UniAR (Qwen3-VL backbone, BSQ visual tokens, SD3 decoder). Raw generated samples only — the LLM-judge scores have been removed.

Quadrants (2, text output)

A steering direction is extracted from minimal pairs in one modality, then injected during pure text generation.

sub	vector from	steers	output
txt2txt	text pairs	whole user message + generation	text
img2txt	image pairs	whole user message + generation	text

Image-output quadrants (img2img, txt2img) remain under steered-gen/ unchanged.

Sweep

• 7 concepts — semantic: emotion age cleanness chaos · visual: spatial_lr size near_far
• 20 subjects / concept — subject-only, adjectives stripped; the three visual concepts share the same 20 scenes
• prompt: Describe a/an <subject> as a single frozen snapshot.
• 7 layer configs: L8=[8] L16=[16] L26=[26] early4=[5-8] mid4=[15-18] late4=[26-29] all=[0-35]
• alpha: -0.2 .. 0.2 step 0.04 (10 steered + baseline 0.0)
• vector: n100-normalized diff-of-means, norm-relative injection; greedy decoding
• steer scope: whole user message + every generated token (system prompt, role markers, and special tokens are NOT steered)

Files

• steered-gen/{semantic,visual}/<concept>/<quad>/generations.json — {prompts, concept, sub, cells: {"<config>|a<±0.0X>": [20 gens]}, baseline: [20]} (70 cells + baseline)
• prompts/<concept>/<concept>_text.json — the 20 subject prompts
• vectors/ — steering vectors