CLIP Progressive Steering Pipeline

Compare 6 embedding-steering methods side-by-side on CLIP ViT-B/16 image retrieval.

Dataset
1 10
0.1 0.8
0.1 0.8

Steering Attributes

Use + to add attributes, to remove, and drag sliders to set weights. Leave all empty to let the LLM auto-generate on the next search.

Positive (steer toward)

Negative (steer away from)

Run a search to see feedback

1. Baseline CLIP

Pure cosine similarity — no steering

2. LLM Linear Steering

q′ = q + α·pos − β·neg

3. Contrastive Subspace

Centroid-based steering direction

4. Energy-Based

Gradient-descent optimisation in embedding space

5. Per-Concept Weighted

Normalised per-attribute weight steering

6. SAE PRF Steering

Pseudo-relevance feedback in sparse autoencoder latent space