Vision-Language Models (VLMs) are increasingly capable, yet they often stumble on tasks that seem trivial to humans—like pointing to the same object in two different photos. We’ve known for a while that these models often "know" more than they let on, but exactly where that information goes missing has been a mystery.
In a new paper, Zhaoning Zhang from the University of Toronto pulls back the curtain on this "utilization gap." By tracing signals through models like Qwen3-VL and InternVL3.5, the research identifies a specific, tiny circuit responsible for cross-image comparison and demonstrates a surgical repair that fixes it.
TL;DR
VLMs often successfully compute matches between images internally but fail to report them in their final answers. This failure is localized to a tiny pathway of just 16–32 attention heads; repairing only these heads (0.02% of parameters) dramatically improves semantic matching and proves that even unconstrained fine-tuning naturally gravitates toward this specific circuit.
What Problem Does This Paper Address?
When you ask a VLM if "Object A" in Image 1 is the same as "Object B" in Image 2, it often fails. Previous theories suggested this might be because the objects lack names (the "nameability" hypothesis) or because the vision encoder isn't powerful enough.
However, this paper argues the problem is one of utilization. The information is often present in the model's internal hidden states but gets lost or ignored before it reaches the final output. The authors set out to find exactly where this "signal break" occurs and whether it can be fixed without retraining the entire model.
Methods
The author uses a "mechanistic interpretability" approach, treating the VLM as a circuit to be mapped:
- Probing: They used linear probes to measure how much correspondence information exists at different stages: the visual tokens, the internal "comparison" stage, and the final answer position.
- Causal Interventions: To prove specific parts of the model are responsible, they used "mass-preserving knockouts." By shuffling or redistributing the attention of specific heads while keeping the total "amount" of attention the same, they could see which heads were actually doing the work.
- Attention Transplantation: They took the attention patterns from a fine-tuned "repaired" model and plugged them into a broken base model to see if the performance gain followed the attention.
- Targeted LoRA: They trained a tiny fraction of the model (only the Query and Key projections of the identified heads) to see if a surgical repair was sufficient.
Datasets
The study evaluated models across a spectrum of difficulty:
- Syn-squiggle & Syn-known: Synthetic tasks matching identical shapes (unnameable "squiggles" vs. known shapes like circles).
- HPatches: Matching the same point across different viewpoints of the same scene.
- SPair-71k: Semantic correspondence (e.g., matching the "wing" of one airplane to the "wing" of a different airplane).
- BLINK: A benchmark used to test if the repairs transferred to general visual perception tasks.
Key Insights and Ideas
- The "Seen but Unsaid" Phenomenon: For exact copies, the model's internal states are highly accurate (up to 25 percentage points better than the output). The model "sees" the match, but the output layer simply doesn't "read" it.
- The Comparison Pathway: Cross-image comparison isn't a diffuse process spread across the whole model. It runs through a "small, identifiable pathway" of 16–32 attention heads.
- Nameability is a Red Herring: The gap between internal knowledge and output is the same for nameable shapes (like "circle") and unnameable squiggles. The failure isn't about language; it's about the internal routing of visual evidence.
- Convergence of Training: Perhaps the most striking finding is that when you train a large "all-layer" adapter to fix the model, it naturally chooses to route its improvements through the same 16–32 heads the authors identified manually.
Key Results
The findings suggest that VLMs are much closer to "solving" vision than their raw accuracy scores suggest:
- Surgical Repair: Training just the Query-Key projections of the identified heads (0.02% of parameters) raised SPair-71k accuracy by +28.3 pp for Qwen3-VL-8B and +25.1 pp for InternVL3.5-8B.
- Causal Proof: Knocking out these specific heads removed 17–22 pp of accuracy on synthetic tasks, while knocking out random heads had almost no effect (at most 7 pp).
- Selectivity: The authors found that in repaired models, the "selectivity" of these heads (how much more they attend to the correct match vs. distractors) perfectly tracks the model's final accuracy.
Limitations and Future Work
While the paper provides a clear map of the failure, some questions remain:
- The "Readout" Mystery: While the paper identifies where the signal is lost for exact copies (at the output readout), it doesn't yet provide a surgical fix for that specific failure mode, as it appears to be a more complex "decoding" issue.
- Scale: Most mechanistic results focused on the 8B parameter scale. While measurements were taken up to 32B, the full repair-and-transplant analysis for larger models is a natural next step.
- Generalization: The synthetic results rely on one stimulus family. Future work could explore if these "comparison heads" are specialized for different types of objects or if they are general-purpose matching engines.
By locating these specific "failure points," this research moves us away from treating VLMs as black boxes and toward a future where we can "repair" specific cognitive deficits in AI with surgical precision.
Summary updated