saintsauce/uniar-steering-eval

UniAR Steering Eval — Cross-Modal Transfer

Activation-steering generations and blind LLM-judge concept-rate scores for UniAR (a unified vision-language model: Qwen3-VL backbone, BSQ visual tokens, and an SD3 decoder). The study asks whether a steering direction extracted in one modality transfers to generation in the other — i.e. is a concept direction shared across the image and text channels, or modality-specific?

The four quadrants (input-pairs → output-modality)

A steering vector is extracted from minimal pairs in one modality, then injected during pure generation in a (possibly different) modality.

sub	vector from	steers	output
img2img	image pairs	image tokens only	image
txt2img	text pairs	image tokens only	image
txt2img-w-prompt	text pairs	whole prompt + image	image
txt2txt	text pairs	all text tokens (full)	text
img2txt	image pairs	all text tokens (full)	text

Within-modal diagonals: img2img, txt2txt. Cross-modal off-diagonals: txt2img-w-prompt (text→image), img2txt (image→text).

Sweep axes

• 7 concepts — semantic: emotion age cleanness chaos · visual: size near_far spatial_lr
• 7 layer configs — L8 early4 L16 mid4 L26 late4 all
• 10 alphas — -0.05 … +0.05 (norm-relative injection, h += α·σ·unit(v))
• 20 prompts per cell · 2 poles judged per pair (positive & negative)

Concept-rate metric

Per (baseline, steered) pair, two blind A/B judge calls (Gemini 2.5 Flash), one per pole, random A/B order: "Which is more {pole}?". pair_score = mean(success_pos, success_neg) ∈ {0, 0.5, 1}; the concept rate is the mean over 20 prompts. 0.5 = chance; < 0.5 = steered the wrong way / broke.

Layout

steered-gen/<group>/<concept>/
    img2img|txt2img|txt2img-w-prompt/          # image quadrants
        baseline512/pNN.png             # unsteered baseline (20 prompts)
        layer-setups/<config>/a<±0.0N>/pNN.png
        grids/                          # montage sheets
    txt2txt|img2txt/generations.json    # text quadrants: {prompts, baseline[20], cells{"<config>|a<±0.0N>":[20]}}
judge_results/
    results.csv                         # master long table (all 5 subs)
    raw/<sub>/<concept>_<config>.json   # per-config alpha→rate
    figures/{density,curves}_<group>_<sub>.png
vectors/<concept>/{img,txt}/*.pt        # the steering directions
prompts/<concept>/<concept>.json        # generation prompts

group = semantic {emotion, age, cleanness, chaos} · visual {size, near_far, spatial_lr}.

results.csv columns

concept, sub, config, alpha, rate, n, model, thinking — one row per unique (concept, sub, config, alpha, model).

Headline finding

At a controlled mid-network config, cross-modal ≈ within-modal for semantic concepts (image-derived vectors steer text ~as well as native text vectors; text-derived vectors steer images ~as well as native image vectors), while visual concepts do not transfer to the text channel (near chance) — the concept directions are shared across modalities for semantics but not visual ones.

Notes

• The all-layer config over-saturates (breaks images / degenerates text) at higher alpha; treat its high scores as artifacts, not signal.
• Several concepts steer one-sided (e.g. emotion→sad, cleanness→dirty steer strongly while the opposite pole collapses) — a baseline-entanglement effect.
• Generations are from UniAR; base-model licenses (Qwen3-VL, SD3) apply to any reuse of the underlying models. This dataset's own contents are CC-BY-4.0.