TL;DR
This paper uses sketches as a "structural probe" to uncover how Large Vision-Language Models (LVLMs) represent objects. The researchers found that LVLMs internalize human drawing conventions (like drawing a cup's body before its handle) within their language models and that part-level grounding emerges naturally in specific, predictable "mid-network bands" of the architecture without any explicit training.
What Problem Does This Paper Address?
Standard benchmarks for LVLMs typically measure output accuracy—did the model correctly identify the object in the photo? However, natural photographs are "noisy" probes; they contain texture, lighting, and background context that models can use as shortcuts. This makes it hard to tell if a model actually understands the underlying structure of an object or is just pattern-matching pixels.
The authors argue that sketches are a superior probe because they strip away everything except the essential structural strokes humans use for recognition. By testing models on sketches, they aim to answer:
- What visual conventions (like drawing order) have models retained?
- Can models "ground" (locate) specific parts of an object without being explicitly trained to do so?
- Where, mechanistically, does this structural knowledge live within the billions of parameters of a transformer network?
Methods
The authors introduce a training-free probing framework with three distinct "readouts":
- Temporal Probing: The model is given a sketch and a shuffled list of its parts (e.g., "handle," "body," "rim"). It is asked to reorder them based on the sequence in which a human likely drew them.
- Spatial Probing: Instead of using a dedicated segmentation model, the authors extract cross-attention weights from the frozen LVLM to see if the model's internal "gaze" naturally falls on the correct part of the sketch when prompted (e.g., "Where is the handle?").
- Mechanistic Probing: They track which specific layers and attention heads carry the grounding signal to see if there is a consistent "locus" of structural knowledge.
Datasets
- SPG (Sketch-Part-Grouping): A dataset of object sketches with part-level annotations and, crucially, the temporal order of every stroke.
- FS-COCO: A dataset of scene-level sketches used to test if grounding holds up in complex environments.
- Pascal-Part & ControlSketch: Used in a pilot study to observe how model performance degrades as images move from natural photos to increasingly abstract sketches.
Key Insights and Ideas
- Conventions Live in the Language Model: Surprisingly, the model's ability to predict drawing order doesn't depend much on the image. Even when shown a blank canvas, models still predicted the "canonical" human drawing order. This suggests that the "visual" convention is actually stored as linguistic knowledge about human behavior.
- Grounding is an Emergent Property: Frozen, off-the-shelf LVLMs can locate object parts in abstract sketches better than some models specifically trained for grounding (like CLIPSeg). This "zero-shot" grounding survives even when texture and color are completely removed.
- The "Mid-Network Band": Grounding isn't scattered randomly. It concentrates in a narrow band of layers. Interestingly, the location of this band depends on the LLM family, not the model size. For Vicuna-based models (like LLaVA), it's at fractional depth 0.35–0.45; for DeepSeek-based models, it's deeper, at 0.50–0.65.
Key Results
- Superior Grounding: The strongest probe, LLaVA-1.5-13B, achieved a mean Accuracy ([email protected]) of 0.73 on object sketches, significantly outperforming trained grounders like GroupViT (0.59).
- Predictive Layer Bands: By restricting the search for grounding information to the identified "mid-network band," the authors improved grounding accuracy across every model tested. For example, LLaVA-1.5-7B saw an improvement from 0.65 to 0.73 [email protected].
- Human Alignment: Model agreement with human drawing order correlates strongly with human consensus (Pearson $r = +0.58$). When humans agree on how to draw something (like an airplane), the model agrees too; when human conventions vary (like drawing a pig), the model's predictions become equally variable.
Limitations and Future Work
- The "Phantom Part" Problem: The current probe assumes the queried part exists. If asked to find the "wings of a coffee cup," the model will still highlight a region (usually the handle) rather than signaling that the part is missing.
- Architectural Scope: The study focused on two major LLM families (Vicuna and DeepSeek). Future work is needed to see if these "grounding bands" exist in other architectures like Mistral or Llama-3.
- Static vs. Dynamic: The temporal probe uses the final sketch. A more advanced probe might look at the model's internal state during the construction of a sketch.
Summary updated