What Sketches Tell Us about LVLMs: Conventions, Grounding, and Localisation

AI summary

Overview

Research area: Multimodal interpretability — specifically, using human sketches as a controlled probe of what large vision-language models (LVLMs) internally represent, where, and how reliably. The work sits at the intersection of sketch understanding, attention-based grounding, and mechanistic analysis of transformer layers and heads.

Technical level: Intermediate. The methods are deliberately training-free and assembled from existing components (shuffled-list ordering, the Localization Heads recipe of Kang et al., classical sequence metrics), but the analysis assumes familiarity with cross-attention, layer/head indices, and grounding metrics such as pointing accuracy and mAcc@0.1.

Scope: A training-free, three-readout diagnostic of nine LVLMs on two part-annotated sketch datasets, reporting part-construction order, attention-based grounding, and the layer/head location of the grounding signal.

What This Paper Is About

Standard LVLM benchmarks report whether a model's final answer is correct, but not which visual conventions the model has absorbed, nor where in the network those conventions are stored. The authors argue that natural photographs are a poor probe because texture, lighting, and scene context let a model answer part-level questions without ever decomposing an object into parts — so they substitute sketches, which keep only the strokes a human judged sufficient for recognition.

The goal is to turn sketches from a domain that models are evaluated on into a probe used to evaluate with: strip the surface signal, ask what structural knowledge survives, and read the answer out of the frozen model's internal state rather than its output accuracy.

Key Contributions

1. A training-free sketch probe with three readouts. A single frozen LVLM is queried under two protocols to yield (i) a temporal readout of part-construction order, (ii) a spatial readout of attention-based part/object grounding, and (iii) a mechanistic readout of the layer and head coordinates that carry that grounding. No fine-tuning, no auxiliary segmenter, no rendering between iterations.

2. Evidence that LVLMs follow human drawing convention rather than imposing their own schema. Models reproduce canonical part orderings where human drawings converge and become correspondingly variable where humans disagree, with blank/withheld/mismatched-image controls that locate the convention in the language model.

3. Demonstration that part-level grounding survives abstraction without supervision. Heads selected with an existing recipe from frozen cross-attention recover part and object regions on both object (SPG) and scene (FS-COCO) sketches.

4. Localisation of the grounding signal to a narrow, family-dependent mid-network band, confirmed causally by ablation, and shown to work as a head-selection prior that matches or improves grounding mAcc@0.1 in every model and dataset tested.

Main Findings

• Part order tracks human consensus. Under a class-consensus protocol on SPG (25 categories, 100 strict-mode-matching sketches per category, parts presented as a shuffled set), agreement rises and falls with human agreement: per-category agreement correlates with human consensus at mean Pearson r = +0.58 across all nine models, significant in eight. Human consensus itself is bimodal — airplane, suitcase, and coffee_cup reach ≥90% strict-mode agreement, while angel, pig, and house fall below 40%.

• Closed-source models lead, but the open-source gap is narrow. Claude Haiku 4.5 is the best overall on the temporal probe (Edit 0.80 ± 0.20, LCS 0.89 ± 0.11, Pair 0.88 ± 0.13), ahead of GPT-5 (nano) (0.72 / 0.85 / 0.81) and Gemini 2.5 Flash-Lite (0.73 / 0.85 / 0.83). Among open models, Qwen2.5-VL-7B reaches 0.69 / 0.84 / 0.80 — within 0.05 LCS of the closed-source leader — while LLaVA-1.5-7B is lowest at 0.38 / 0.66 / 0.46.

• The ordering comes from the language model, not the image. Of eighteen paired comparisons over the 25 categories (two models, three conditions, three metrics) with prompt, parts, shuffle, and decoding fixed, none favours the real sketch, and the only significant comparison favours the no-image condition (p = 0.04). For Qwen2.5-VL-7B, real sketch and blank canvas are identical to two decimals (0.69 / 0.84 / 0.79); for LLaVA-1.5-7B the real sketch is 0.37 / 0.66 / 0.46 against 0.44 / 0.69 / 0.51 with no image. A test on sketches whose human order departs from the canonical one gives p = 0.26 under clustering by category.

• Frozen attention recovers regions without grounding supervision. The strongest probe, LLaVA-1.5-13B, reaches 0.46 pointing accuracy and 0.73 mAcc@0.1 on SPG, and 0.61 and 0.78 on FS-COCO. Against trained grounders, LLaVA-1.5-13B exceeds GroupViT on pointing accuracy but trails CLIPSeg (0.46 vs 0.56 on SPG, 0.61 vs 0.67 on FS-COCO), while on mAcc@0.1 it leads both (CLIPSeg sits at 0.40 on FS-COCO against 0.78). At stricter overlap the lead holds at mAcc@0.3 (significant in three of four comparisons) and disappears at mAcc@0.5, where no comparison is significant.

• Grounding quality is not monotonic in parameter count. Within the LLaVA family, scale helps (0.43 to 0.46 pointing on SPG, 0.53 to 0.61 on FS-COCO from 7B to 13B), but within DeepSeek-VL the larger 7B variant underperforms the 1.3B on objects (0.23 vs 0.31 pointing).

• Richer input sharpens localisation. Scene-grounding pointing accuracy on FS-COCO rises from 0.32 at one object to 0.55 at seven objects; within SPG a smaller trend holds (0.34 to 0.42 from two to four parts). This mirrors the pilot, where increasing abstraction along the natural-photo → grayscale → ControlSketch axis monotonically degrades grounding and pushes the peak grounding layer earlier while spatial entropy rises only modestly.

• The grounding band is narrow and set by the base language model. Top-3 grounding heads concentrate at fractional depth 0.35 to 0.45 for the three Vicuna-based models (LLaVA-1.5-7B/13B, ShareGPT4V-7B) and 0.50 to 0.65 for the two DeepSeek-based models. The band is conserved across object and scene sketches (Δ = +0.004; |Δ| = 0.034 in fractional depth), with within-family SD of 0.049 (Vicuna trio) and 0.040 (DeepSeek pair) against a cross-family gap of 0.175 — roughly a 3.5× ratio.

• The same heads recur across domains. On average the five models share 5.4 of their top-10 heads, against fewer than one expected by chance, with selection frequencies correlating at r = +0.74. Combined top-3 spatial entropy concentrates between 0.78 and 0.95, with overlapping SPG and FS-COCO distributions.

• The band acts as a usable prior. Restricting head selection to each family's empirical band ([0.35, 0.45] Vicuna, [0.50, 0.65] DeepSeek) matches or exceeds full-grid scoring in every (model, dataset) cell, with the largest gain on LLaVA-1.5-7B (+0.08 on SPG, +0.09 on FS-COCO mAcc@0.1).

• The selected heads are causally load-bearing. Zeroing the three most frequently selected heads in LLaVA-1.5-13B lowers the logit margin towards the correct side on a downstream localisation question by 0.118 (95% CI 0.074 to 0.163, p = 0.0001, clustered over 24 categories), against 0.001 for three random heads and 0.010 for three other heads in the same layers. Side accuracy itself barely moves.

• Temporal and spatial probes couple only at scale. Per-class LCS against per-class mAcc@0.1 over the 25 SPG categories dissociates in LLaVA-1.5-7B (r = −0.13, ρ = −0.16, not significant) but couples in LLaVA-1.5-13B (r = +0.40, ρ = +0.42, p < 0.05).

Methodology in Plain English

The design rests on the observation that a sketch removes texture, lighting, and photographic context while preserving exactly the strokes a person considered necessary for recognition. If a model can still work with that skeleton, the surviving signal is structural rather than photometric.

Temporal probe. The model is shown a sketch, an object name, and its parts as a shuffled set, then asked to return the parts in the order a human would have drawn them. The response is compared to human drawing trajectories from SPG using three complementary sequence metrics: edit distance, longest common subsequence, and pairwise-order agreement (Kendall's τ). Because the parts are textual and never re-rendered, the returned sequence reflects only the model's internal ordering — which is why the authors run controls that blank the sketch, remove it entirely, or substitute a sketch from another category.

Spatial and mechanistic probes. The model is asked a locational question — "Where is the {p} of the {c}?" for object parts, "Where is the {c}?" for whole objects on scene sketches. Cross-attention from the last input token to the image patches is extracted across all layers and heads. Following Kang et al. (2025), each head is scored by how much attention mass it places on visual tokens and by spatial concentration (lower entropy means tighter focus), reduced with a chord-elbow rule, with the top k = 3 heads forming the grounding set. Those heads' maps are smoothed (mean-subtract, ReLU, Gaussian with σ = 1), combined by element-wise maximum, and thresholded at the 80th percentile to produce a mask scored against ground-truth part masks (SPG) or object masks (FS-COCO, using FS-COCO-Seg).

Mechanistic readout. The same machinery is reused, but the object of study becomes the (layer, head) coordinates of the selected heads and the entropy of their combined map — asking whether the band that carries grounding is stable across models and across the object/scene divide.

Model sets. The temporal probe needs only generated text, so it covers nine models: six open-source (LLaVA-1.5-7B/13B, Qwen2.5-VL-3B/7B, Molmo-7B-D, InternVL2-8B) and three closed-source (GPT-5 (nano), Claude Haiku 4.5, Gemini 2.5 Flash-Lite). The attention readouts need access to attention weights, so they cover five open models spanning two visual stacks and two base language families: LLaVA-1.5-7B/13B and ShareGPT4V-7B (CLIP ViT-L/14 encoder, Vicuna language model) and DeepSeek-VL-1.3B/7B (SigLIP-based encoder, DeepSeek-LLM). Decoding is deterministic at temperature 0, with closed-source predictions averaged over three runs per item.

Why This Matters

Impact on research. The paper offers a diagnostic that a benchmark score cannot: it shows not just whether a model is right, but whether its errors are structured, graded, and attributable. It also reframes sketches from a domain-shift test set into a controlled instrument for interpretability, and reports three findings that an aggregate score would hide — that DeepSeek-VL-7B trails DeepSeek-VL-1.3B on object grounding while placing its heads in the same band; that under abstraction a model becomes wrong about where a part lies before its attention becomes diffuse; and that variability on ambiguous categories is informative rather than simply a lower score.

Real-world applications:

• Sketch-based retrieval and segmentation, where pipelines currently bolt on a dedicated segmenter could read region grounding directly from an LVLM they already run.
• Sketch editing and generation tools, which need part-level region signals to support targeted edits without an auxiliary network.
• AR and UI prototyping pipelines, where users draw rough structures and require the system to identify which region corresponds to which named component.
• Backbone selection for grounding-heavy systems, since band position follows language-model lineage rather than parameter count in the two families compared — suggesting the smaller variant in the right family often suffices.

Industry relevance. The probe is training-free and forward-pass only, so it can be applied to deployed models as a diagnostic without retraining. The finding that a narrow, causally load-bearing band can serve as a head-selection prior gives attention-steering methods a natural intervention point instead of searching the network blind.

Future Directions

• Extend beyond two language-model families. The attention readouts cover only Vicuna and DeepSeek lineages; the paper states explicitly that further families are future work, which matters because the cross-family gap (0.175) dominates the within-family variation (0.049 and 0.040).

• Add a presence check. The probes assume the queried part exists. Asked about parts drawn from other categories, the model never signals absence — producing answers such as "The bell of the airplane is located at the top of the wing" and "The stamens of the alarm clock are located at the top of the clock." The readout therefore says where a named part is, not whether it is there, and pipelines where parts may be absent need a separate check.

• Use the band for attention steering. Since the identified heads are causally load-bearing, targeted steering of those heads — rather than blind network search — is a natural next experiment.

• Reconcile temporal and spatial coupling. The two readouts couple only in the larger LLaVA variant and dissociate in the smaller one; understanding what changes at scale to produce that coupling is left open.

Target Audience

Researchers and engineers working on multimodal interpretability, attention-based grounding, or sketch-based vision systems. It is most useful to readers who already know what cross-attention over image patches looks like and who want a diagnostic they can run on a frozen model — and to practitioners choosing a backbone for a grounding-heavy pipeline, who will find the family-versus-parameter-count result directly actionable. Readers seeking new architectures or training objectives will not find them here by design: the paper's analytical weight falls on the frozen models, not on new apparatus.

Authors’ abstract

LVLM benchmarks tell us whether a model answers correctly, but not which visual conventions it has retained, nor where those conventions live. Sketches expose this hidden structure. A sketch removes texture, lighting, and photographic context while preserving the strokes humans judged sufficient for recognition. If an LVLM can operate on that abstraction, the surviving signal is structural rather than merely photometric. We introduce a training-free sketch probe with three readouts: temporal part-construction order, spatial attention-based grounding, and the mechanistic layer/head locus of grounding. We evaluate nine LVLMs on part order and five on the attention readouts, over two part-annotated sketch datasets. Three findings emerge. First, LVLM part order tracks human convention: models reproduce canonical orderings where human drawings converge, and become correspondingly variable where human orderings vary. Controls that blank, withhold, and mismatch the sketch locate this convention in the language model. Second, part-level grounding survives the abstraction: attention heads selected from the frozen model with an existing recipe recover part and object regions on object and scene sketches, with no grounding supervision. Third, the grounding signal concentrates in a narrow mid-network band whose location follows the base language model rather than parameter count in the two families we compare, and remains stable across object and scene sketches. The result is a practical diagnostic for sketch-driven LVLMs: no fine-tuning, no auxiliary segmenter, and a layer band that doubles as a head-selection prior, matching or improving grounding mAcc@0.1 in every model and dataset we tested.