Here's a full edit example: wherever Qwen latently represents "tractor," we replace it with the "revolver" direction. Since the model can still see features like wheels and a license plate, it interprets the image not as a tractor, but as a "surreal revolver that looks like a car"!

Is the info being read by this lens actually causal, though? To answer this, we use the *inverse* of our lens to obtain latent concept vectors. Just by adding the latent vector for "iPod" and subtracting the vector for "ant," we can fool Qwen into describing an ant as an iPod.

Some of the information revealed by this lens is local (e.g., colors, objects), but other information seems to be broader. For example, Qwen3-8B seems to think of this photo as being taken in Wyoming (which isn't too far off from where I actually took the photo, in Calgary).

This method gives us probs. across all vocab tokens, so we can observe subtleties depending on word meaning. For example, "phone" has high probability in this image for both a cell phone and landline, whereas 手机 and 电话 focus on cell phone / landline tokens respectively.

To quickly run this approach over many image tokens, we collapse together the weight matrices (the OV matrices) of this set of verbalization heads. This gives us a single d_model x d_model matrix that can be applied to any hidden state at any layer, and even works at layer 0.

First, we found some heads responsible for OCR (see paper). We could see them attending to image tokens that contain text, e.g. the word "bike". But then, we found that if we just forced all of these heads to attend to, e.g., a bird wing token, Qwen3 would output "feathers." 😮

[New preprint] How do images in VLMs align with words? In this work, we found a set of attention heads responsible for OCR. But to our surprise, these heads were actually able to verbalize much more than just text. So, we used them to create a simple logit lens for image tokens!

[📄] Are LLMs mindless token-shifters, or do they build meaningful representations of language? We study how LLMs copy text in-context, and physically separate out two types of induction heads: token heads, which copy literal tokens, and concept heads, which copy word meanings.