Match VISE model card tone

README.md

@@ -19,25 +19,29 @@ tags:
 - unsupervised
 ---
 
-# CVPD
-
-This is the CVPD LoRA adapter for `Qwen/Qwen3-VL-4B-Instruct`, from our BMVC 2026 paper
-[Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spots](https://github.com/mbzuai-oryx/CVPD).
-
-Vision-language models often miss details they can actually see. Ask about a small
-sign in a photo and the model gets it wrong, but crop to that sign and it answers
-correctly. The information was there, the model just didn't attend to it.
-
-CVPD trains models to fix this using nothing but the model itself. We show it an
-unlabeled image, let it write its own question, and then check two things: does
-zooming into a region change and sharpen its answer, and does blurring that region
-leave its answer alone? When both hold, we've found a blind spot: something the
-model can perceive but doesn't use. We then train the model on the full image to
-behave the way it did on the crop, and away from how it behaved on the blurred
-version.
-
-No captions, no bounding boxes, no labels, no reward models, no bigger teacher
-model. Just the model and a pile of raw images.
+# CVPD: Contrastive Counterfactual Visual Process Distillation
+
+This is the CVPD LoRA adapter for `Qwen/Qwen3-VL-4B-Instruct`, from our paper
+[Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spots](https://github.com/mbzuai-oryx/CVPD),
+accepted to BMVC 2026.
+
+CVPD is a fully self-contained framework for dense, token-level visual self-distillation.
+Where prior visual distillation builds its privileged context from outside the model, using
+segmentation systems, region proposals, or a stronger annotator, CVPD recovers that context
+from the model's own counterfactual behavior. We train on raw, unlabeled images with no
+captions, bounding boxes, labels, reward models, or teacher models, by locating **visual blind
+spots**, regions the model can perceive but fails to exploit under full-image conditioning:
+
+- **Discovery:** the model writes its own question and probe answer per image and proposes
+  candidate regions by self-grounding and by 3x3 and 2x2 partitions. A region is kept when
+  cropping to it moves and sharpens the answer distribution while blurring ("ghosting") it
+  leaves the full-image behavior unchanged.
+- **Contrastive self-distillation:** each retained region instantiates four policies from the
+  same backbone. The online student learns from a crop-conditioned positive teacher, is pushed
+  away from a ghost-conditioned negative teacher, and is anchored to a frozen reference policy.
+
+We combine these as latent transfer plus contrastive ranking (`lambda_rank=0.5`, margin
+`m=0.1`) and optimize under a KL anchor adapted online to a target divergence of `0.03`.
 
 ## Usage
 
@@ -72,13 +76,16 @@ print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special
 
 ## Training
 
-- Base model: `Qwen/Qwen3-VL-4B-Instruct`, with the vision encoder frozen.
-- LoRA: `r=32`, `alpha=64`, `dropout=0.05`, on the attention and MLP projections
-  of the language model.
-- AdamW, learning rate `2e-5`, weight decay `0.01`, gradient clipping `1.0`, bfloat16.
-- One epoch over the curated blind-spot pool.
-- Data: 15,000 raw unlabeled images, which the discovery pipeline narrows down to
-  about 2,590 training examples.
+- Base: `Qwen/Qwen3-VL-4B-Instruct`, vision encoder frozen.
+- LoRA: `r=32`, `alpha=64`, `dropout=0.05` on the attention and MLP projections of the
+  language model backbone.
+- Optimizer: AdamW, `lr=2e-5`, weight decay `0.01`, gradient clipping `1.0`, bfloat16.
+- Objective: latent transfer + contrastive ranking (`lambda_rank=0.5`, margin `m=0.1`) +
+  KL anchor, initialized at `1e-3` and adapted online to a target KL of `0.03`.
+- Teachers: momentum pair maintained as an EMA of the LoRA parameters (`alpha=0.05`).
+- Discovery: thresholds `tau_crop = tau_ghost = 0.05`, all views resized to 448x448.
+- Data: 15,000 raw, unlabeled images yielding ~2,590 curated blind-spot tuples. No captions,
+  bounding boxes, labels, answers, or reward signals.
 
 ## License