Subspace Validity Suite (SVS)
A diagnostic toolkit for validating claimed "visual directions" in Vision-Language Models.
Do your visual subspace directions actually capture visual content, or are they just network geometry?
SVS provides 6 tests that any claimed visual direction must pass before being attributed mechanistic significance.
Quick Start
from svs_toolkit import SubspaceValiditySuite
svs = SubspaceValiditySuite()
# Run all tests
report = svs.full_report(
directions=your_visual_directions, # (k, d) numpy array
h_visual=visual_hidden_states, # list of (d,) arrays
h_gibberish=gibberish_hidden_states, # list of (d,) arrays
h_factual=factual_hidden_states, # optional
h_math=math_hidden_states, # optional
)
svs.print_report(report)
The 6 Tests
| # | Test | What it checks | PASS condition |
|---|---|---|---|
| 1 | Gibberish Specificity | Do directions respond more to visual than gibberish? | Gib/Vis < 0.8, TOST rejects equivalence |
| 2 | Cross-Type Discrimination | Can projection magnitude tell visual from non-visual? | AUROC > 0.65 |
| 3 | Projection Magnitude | Do directions capture more variance than random? | Ratio > 1.5x |
| 4 | Anisotropy Orthogonality | Are directions different from general PCA? | Mean cosine < 0.3 |
| 5 | Direction Consistency | Are directions stable across calibration splits? | Split alignment > 0.8 |
| 6 | Pairwise Discrimination | Do projections encode visual content specifically? | Only visual pairs separable |
Key Finding
We tested 11 methods across 3 VLM architectures, 2 text-only backbones, and 2 non-VLM transformer families. Every method fails the SVS.
PCA-based "visual directions" capture generic network geometry — the dominant modes of the weight matrices — not visual content. Gibberish activates them identically to visual descriptions (Gib/Vis ≈ 1.00, TOST equivalence p < 0.0001, N = 4,000).
Installation
pip install numpy scipy scikit-learn
No additional dependencies required. Single file: svs_toolkit.py.
Generating Test Inputs
SVS needs hidden states from a text-only backbone (e.g., Vicuna-7B):
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained("lmsys/vicuna-7b-v1.5",
torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("lmsys/vicuna-7b-v1.5")
def get_hidden(prompt, layer=16):
inp = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
out = model(**inp, output_hidden_states=True)
return out.hidden_states[layer][0, -1, :].cpu().float().numpy()
# Visual prompts
h_visual = [get_hidden(p) for p in [
"A kitchen with a table and chairs.",
"A beach with surfers and umbrellas.",
# ... 20+ prompts
]]
# Gibberish prompts
h_gibberish = [get_hidden(p) for p in [
"Xkq plm wvt zzz brrn.",
"Qwzyx nkl jjj hhh ttttt.",
# ... 20+ prompts
]]
Citation
@inproceedings{svs2027,
title={The Subspace Validity Suite: Do Visual Directions in
Vision-Language Models Survive Basic Sanity Checks?},
author={Anonymous},
booktitle={WACV},
year={2027}
}
License
MIT