betaFeedback

WorksComputer Science

CompositionEvidence: Controlling Geometric Evidence for Cinematic Composition Reasoning with Frozen Vision-Language Models

Abstract

Cinematic composition questions depend on in-frame geometry that vision-language models (VLMs) use inconsistently. External perception tools can expose this geometry, but naive prompt injection can overwrite useful visual predictions. On ShotBench composition, pre-question geometry reduces Qwen2.5-VL-7B from 49.5% to 46.3%; corrupted measurements lower it to 33.6%. CompositionEvidence derives nine reproducible features from frozen detection, depth, and line experts and supplies their shared 24-dimensional representation through two interfaces. A training-free channel appends gated neutral evidence after the question, reaching 56.2% (+6.7, exact McNemar p=0.0018) and 48.4% under corrupted evidence. A minimal-training channel maps the same representation to four soft tokens with a 7.37M-parameter projector while leaving the VLM frozen; it reaches 61.0% (+11.5). Across nine VLMs, including five commercial models queried through official provider APIs, zero-shot accuracy ranges from 43.8% to 63.0%, and the text channel changes it by +6.7 to -6.9 points; the highest baseline is among the regressions. The effect of explicit geometry depends on measurement reliability and on how the receiving model integrates it.

Introduction

Cinematic composition carries meaning through geometry, such as where a subject sits, how visual mass is distributed across the frame, and whether a shot is mirror-alike. Current frontier VLMs are broadly capable, but they use this kind of fine-grained evidence unevenly. In our nine-model ShotBench [20] evaluation, zero-shot accuracy spans 43.8%–63.0%, and the same geometric evidence produces gains on some models and regressions on others. This paper studies how the evidence is delivered to a VLM, and every controlled comparison keeps the receiving model fixed.

Composition questions require information that general-purpose vision encoders are not trained to preserve. Such encoders are optimised for semantic alignment, while composition decisions turn on small geometric quantities such as the offset of a centroid from a reference line, the difference in left/right visual mass, mirror similarity, and depth layering. Frozen detection, monocular-depth, and line-segment experts can measure these quantities and pass them to the VLM.

Supplying the measurements does not by itself improve the model. When we place the geometry before the question, Qwen2.5-VL-7B drops from 49.5% to 46.3% and 165 of 479 answers change. Replacing the measurements with internally consistent but incorrect values lowers accuracy to 33.6%, which is above the 25% chance rate and 15.9 points below zero-shot. The injected block displaces part of the model’s visual prior, so its position and its reliability need as much attention as the extraction procedure itself.

CompositionEvidence2 supplies the same nine formalised geometric features through two interfaces. Channel A is training-free and compatible with closed APIs. It lets the VLM read the image, question, and options before appending a controlled evidence block whose uncertain lines are suppressed by training-fitted thresholds. This ordering and filtering raise ShotBench composition accuracy to 56.2% (+6.7, p=0.0018p=0.0018). Channel B uses a two-layer projector to map the 24-dimensional feature vector to four soft tokens while the VLM remains frozen. Training only 7.37M parameters yields 61.0% (+11.5), outperforming the published ShotVL-3B score of 57.4% and coming within 1.2 points of fully fine-tuned ShotVL-7B v1.1. Both channels reuse identical cached features, which isolates the interface choice.

We make three contributions.

  • We formalise nine composition features computed from frozen detection, depth, and line experts, and deliver the resulting 24-dimensional vector through two interfaces, confidence-gated post-question text and a learned four-token projection.

  • We isolate the effect of the interface with paired transition counts and corrupted-evidence controls, and show how pre-question injection replaces correct baseline decisions.

  • We measure where each interface helps and where it hurts, using interface ablations, the seven non-composition ShotBench dimensions, a second benchmark, and nine VLMs.

Related Work

Cinematic understanding. Early computational film analysis treated shot scale and camera motion as separate classification tasks [9, 25]. ShotBench [20] unifies cinematic understanding into eight dimensions with 3,608 expert QA pairs and introduces ShotQA and the ShotVL specialists. ShotVL’s evidence-free chain-of-thought SFT offers little improvement over plain SFT. RefineShot [32] restructures parts of ShotBench, but its composition subset is identical to the original. We therefore use CineTechBench [29], which has a different label taxonomy, for cross-benchmark evaluation.

Structured evidence for multimodal reasoning. CamReasoner [33] combines structured camera-motion reasoning with reinforcement learning, while Geometry-Guided Camera Motion [5] injects 3D foundation cues into a frozen VideoLLM, and explicit spatial cues have also been shown to improve multimodal spatial reasoning [37]. We study how a VLM responds to imperfect external measurements, and our paired and corrupted-evidence controls separate the effect of the injection interface from the information the perception tools supply.

Spatial reasoning and external tools. Diagnostic datasets from CLEVR [12] and GQA [10] to SpatialSense [34], VSR [19], and What’sUp [13] repeatedly expose a gap between semantic recognition and relational geometry. SpatialVLM narrows it through large-scale spatial supervision [3]; we instead retain a general frozen VLM and supply compact, auditable measurements at inference. Tool-augmented systems such as VisProg and ViperGPT compose specialist vision modules through generated programs [8, 27], and Toolformer learns when to invoke textual APIs [26]. We study the failure mode that follows. Even when a tool has returned a useful value, the receiving model may over-trust it.

Efficient conditioning and calibration. Frozen-backbone interfaces commonly learn a bridge from visual features to language-model inputs [1, 17]; prompt and prefix tuning adapt frozen models with learned input vectors [11, 14, 18], and PIN shows that a small positional insert can unlock localisation in a frozen VLM [4]. In Channel B the tokens encode a deterministic function of explicit composition measurements, whereas prompt tuning learns free task parameters. Channel A is closer to selective prediction [6], although it gates individual evidence statements and always answers. It also differs from probability calibration [7], since what we calibrate is the influence of an external evidence source.

Composition analysis and generation. Photographic composition guidance [36], composition-preserving aesthetic assessment [23], and aesthetic cropping [16, 31] support shooting or view recommendation, and generation-side control asks how to produce a layout. We ask whether a language model can interpret the composition of an existing cinematic frame from explicit geometric evidence.

CompositionEvidence

Task and Interfaces

Given a film still II and a multiple-choice composition question qq, we seek to improve the VLM prediction without changing its weights (Channel A), or by training only a small adapter (Channel B). Figure 1 traces both the inference path and the train-only construction of its artifacts. The official image, question, and options remain the VLM task input. In parallel, GroundingDINO, Depth Anything V2, and LSD produce boxes, depth, and line segments; deterministic synthesis collects them as BB, pp, DD, LL, MM. Nine pure functions (five subject, two appearance, and two scene features) then form the shared vector z∈R24\mathbf{z} \in\mathbb{R}^{24} used by both channels.

Figure 1. CompositionEvidence architecture. Frozen detection, depth, and line experts feed deterministic synthesis and nine geometric functions, which yield the shared vector z\mathbf{z}. Channel A gates and appends selected measurements as post-question text; Channel B maps z\mathbf{z} into four evidence tokens. Only the fitted thresholds and the projector reach inference.

Channel A exposes only the measurements that pass their confidence gates, renders them without answer-label language, and places the block after the MCQ before tokenisation. Channel B maps the same z\mathbf{z} to four soft tokens that replace four placeholder embeddings, leaving the vision encoder, tokeniser, and decoder unchanged. The lower strip of Figure 1 runs only at training time. There, 1,328 of the 9,626 ShotQA samples fit the mass and mirror thresholds with Youden JJ, and the full composition pool supplies the two-epoch language-model loss that learns the projector. Only the fitted thresholds and the projector weights cross into inference.

Evidence Representation

Let ϕ(I)\phi(I) denote the frozen perception-and-synthesis stack and z=ϕ(I)∈R24\mathbf{z}=\phi(I)\in\mathbb{R}^{24} its continuous values and presence flags. For Channel A, renderer rjr_j verbalises feature group jj, while a gate gj(z;τj)∈{0,1}g_j(\mathbf{z};\tau_j)\in\{0,1\} decides whether that line is exposed:

EA(z)=∥[gj(z;τj)rj(z)]j=19∥,y^A=fθ(I,q,EA).(1)E_A(\mathbf{z})=\left\lVert\left[g_j(\mathbf{z};\tau_j)r_j(\mathbf{z})\right]_{j=1}^{9}\right\rVert,\qquad\hat{y}_A=f_\theta(I,q,E_A). \tag*{(1)}

The concatenation omits a term when its gate is zero. Thresholds τj\tau_j and percentile annotations are functions only of ShotQA training statistics. For Channel B, a projector PψP_\psi produces four evidence embeddings,

h=GELU⁡(W1z+b1),(2)\mathbf{h} = \operatorname{GELU}(W_1\mathbf{z} + \mathbf{b}_1), \tag*{(2)}
TB(z)=reshape⁡4×d(W2h+b2),y^B=fθ(I,q,TB),(3)T_B(\mathbf{z}) = \operatorname{reshape}_{4\times d}(W_2\mathbf{h} + \mathbf{b}_2), \qquad\hat{y}_B = f_\theta(I,q,T_B), \tag*{(3)}

where θ\theta is frozen and only ψ={W1,b1,W2,b2}\psi= \{W_1,\mathbf{b}_1,W_2,\mathbf{b}_2\} is learned. Both channels hold the evidence source and VLM fixed; they differ in how the measurements enter the model. All controlled comparisons reuse the same cached features, so the ablations compare interfaces without changing the detector or base VLM. Channel A updates no parameters and remains compatible with API models, whereas Channel B trains only ψ\psi on ShotQA. No test image changes the feature definitions.

Geometric Evidence Extraction

The perception layer contains three frozen experts and one deterministic synthesis module. GroundingDINO [22], queried first for people and then with a generic-subject fallback, produces boxes; Depth Anything V2 [35] estimates relative depth; and the OpenCV LSD detector [28] extracts line segments. We synthesise pseudo-saliency from confidence-weighted box masks and depth closeness. A fixed rule selects the dominant subject from area, centre prior, and confidence, weighting person classes by 1.2×1.2\times and discarding any box that covers more than 90% of the frame.

Table 1 lists the nine features computed from these outputs. Each feature is a pure function of the synthesised observations and is covered by synthetic-image unit tests. F4 uses SSIM [30] between a frame and its mirror image. F6 uses equal-width depth bins; an earlier quantile design produced layer shares that were constant by construction.

IDQuantityChannel A exposes it when
F1Subject centroid (x,yx, y)a dominant subject is detected
F2Distance to midpoint and thirdsF1 is exposed
F3Subject area ratioF1 is exposed
F4Mirror similarity (SSIM)horizontal SSIM outside [0.144, 0.262]
F5Left/right visual mass b∈[−1,1]b \in[-1, 1]∣b∣≥0.143|b| \ge0.143
F6Relative depth layersa depth map is available
F7Leading lines and horizonline extraction succeeds
F8Head spacea face is detected
F9Edge proximity and full spanwithin 0.05 of an edge, or spanning a side

Table 1. The nine geometric features and the Channel-A gates. Channel B always receives the corresponding continuous values and the presence flags.

Channel A: Post-Question Gated Text

Channel A renders measurements rather than answers. Any expression sharing a word root with the benchmark’s label vocabulary is forbidden (for ShotBench: center/centre, symmetry-, balanc-, heavy, short side), and generation stops if the renderer detects a violation. On a benchmark with different labels, the ban is re-derived from that taxonomy. We also add training-distribution percentiles to quantities such as mirror SSIM and mass difference, because their values on film frames are far below intuitive natural-image scales.

Two choices calibrate how the VLM uses this text. First, we append the block after the image question, options, and answering instruction, so the model forms a task-specific prior before it reads the external measurements. A similar ordering sensitivity has been reported for long text contexts in language models [21]; here it appears in a short multimodal prompt, and we isolate it with paired and corrupted-evidence controls. Second, we expose a measurement only when its signal is reliable, using the gates in Table 1. Neutral facts such as subject position, depth, and lines are always exposed. We fit the numeric thresholds on ShotQA training data with the Youden JJ statistic and keep them fixed for every reported run. We added the gating mechanism after a first pass of ShotBench diagnostics, and § 5.3 discusses this design-selection effect.

Channel B: Soft-Token Projection

Channel B uses the same evidence without textual rendering. A two-layer MLP maps the 24 continuous features and presence flags through 24→512→4×d24 \to512 \to4 \times d, where d=3,584d=3{,}584 for Qwen2.5-VL-7B, and the four outputs replace placeholder embeddings through an embedding-layer hook. The VLM stays frozen; the projector contains (24×512+512)+(512×4d+4d)=7,367,168(24 \times512 + 512) + (512 \times4d + 4d) = 7{,}367{,}168 trainable parameters. A seed-42 shuffle splits the 9,626 ShotQA composition samples into 9,145/481 train/validation items. We train two epochs with AdamW (learning rate 10−410^{-4}, weight decay 0.01), effective batch size 8, gradient clipping at 1.0, and cosine scheduling; the projector is FP32, the frozen VLM BF16, and validation loss selects the checkpoint. Training takes a few hours on one RTX 4090D.

Experiments

Our evaluation asks three questions. RQ1: Can calibrated geometry improve composition reasoning without updating a VLM, and with only minimal training? RQ2: Why does naive injection fail, and which design choices repair it? RQ3: Is the effect specific to one model or benchmark?

Setup

We evaluate all 479 questions in the ShotBench composition subset, using its official four-option prompt. The labels are Center, Symmetrical, Balanced, Left heavy, Right heavy, Short side, and a few combinations. Our Qwen2.5-VL-7B [2] zero-shot reproduction is 49.5%, within two correct answers of the official 49.9%. Inference uses greedy decoding and a fixed seed; predictions are logged per sample and compared with paired exact McNemar tests [24].

The gates are calibrated on 1,328 ShotQA training images (at most 300 per class). Calibration moves the mirror-SSIM threshold from the intuitive 0.7 to 0.262 and the visual-mass threshold from 0.25 to 0.143, and puts pseudo-saliency direction precision at only 67.8%, which motivates selective rendering. Besides the primary model we test InternVL3-8B [38], LLaVA-OneVision-7B [15], ShotVL-7B v1.1 [20], and five commercial VLMs queried through their official provider APIs in July 2026.

Determinism and paired evaluation. Every local condition evaluates all 479 test items once with greedy decoding, seed 42, and at most 64 generated tokens, using Python 3.12.3, PyTorch 2.12.1, Transformers 4.57.1, and one 24GB GPU. All systems answer the same items, so we report exact McNemar tests on wrong-to-right versus right-to-wrong transitions instead of treating the accuracies as independent samples. We retain per-sample predictions, features, rendered blocks, thresholds, and checkpoint hashes, so every aggregate in the paper can be reconstructed.

RQ1: Results on Qwen2.5-VL-7B

Both calibrated channels improve the same zero-shot model, while reasoning or raw evidence alone does not (Table 2). Channel A gains 6.7 points without model-specific training. Channel B gains 11.5 points, exceeds ShotVL-3B (57.4%), and comes within 1.2 points of fully fine-tuned ShotVL-7B v1.1 while training roughly three orders of magnitude fewer parameters. The rule classifier reaches 37.8% and chain-of-thought without evidence falls below zero-shot, so the gains require the VLM prediction and the added geometry together. Current commercial baselines appear in Figure 4.

ConditionAcc. (%)Δ vs. C0Exact McNemar
Rule classifier (no VLM)37.8−11.7—
C0 zero-shot (official prompt)49.5——
Chain-of-thought, no evidence47.4−2.1—
Naive evidence (pre-question)46.3−3.1p=0.276p=0.276 (75↑/90↓)
Post-question, ungated53.4+3.9—
Channel A: post-question + gates56.2+6.7p=0.0018p=0.0018 (66↑/34↓)
Channel B: four soft tokens (7.37M)61.0+11.5p<10−5p<10^{-5} (100↑/45↓)

Table 2. ShotBench composition results with Qwen2.5-VL-7B (n=479n=479). Both calibrated channels improve the same zero-shot model; reasoning or raw evidence alone does not. ShotBench reference scores: GPT-4o 55.2, ShotVL-3B 57.4, fine-tuned ShotVL-7B v1.1 62.2.

Figure 4. The training-free recipe applied unchanged to nine VLMs on ShotBench composition. The five commercial models are queried through official provider APIs.

The paired counts in Table 2 show which predictions change. Channel A nets 32 corrected items and Channel B 55, while naive injection changes the most answers (165) and breaks more correct predictions than it repairs.

The gains concentrate on the classes that geometry resolves directly. Under Channel A, Left heavy rises from 38.3% to 75.3% and Right heavy from 48.1% to 73.4%. Balanced falls from 74.0% to 57.5%, which is the main residual cost of noisy visual-mass evidence, and Short side stays at zero. Figure 2 shows the redistribution across the full label space.

Figure 2. Confusion matrices for zero-shot (left), Channel A (middle), and Channel B (right). Calibrated evidence recovers directional classes while preserving much of the baseline diagonal; Short side remains nearly absent.

RQ2: Interface Ablations and Corruption Controls

Pre-question injection. Putting evidence before the question lowers accuracy to 46.3%, but the aggregate understates the effect: 75 baseline errors are corrected while 90 correct predictions are broken. Numeric-only and neutral-language renderings behave similarly, ruling out wording as the main cause, and a single feature group (48.2%) outperforms the complete pre-question block. The evidence changes the answer distribution even when its net effect is negative.

Post-question placement and gating. Moving the same ungated block after the question raises accuracy to 53.4%, an increase of 7.1 points over pre-question injection. Confidence control then improves it to 54.5% with strict precision gates and to 56.2% with Youden gates.

Evidence interfaceControlAcc. (%)
Pre-question, fullnone46.3
Pre-question, numericnone45.1
Pre-question, one groupnone48.2
Post-question, fullnone53.4
Post-question, fullstrict gates54.5
Post-question, fullYouden gates56.2

Table 3. Interface ablation with the evidence source and Qwen2.5-VL-7B fixed. Wording alone does not repair pre-question injection; position supplies the largest step, and selective coverage adds the remaining gain.

A 3×33 \times3 sweep over the centre/thirds alignment threshold (0.05–0.12) and the strong-symmetry threshold (0.6–0.8) stays within 56.8–57.0%, so the conclusion does not depend on one threshold combination. Position protects the model’s task prior, and gating limits the influence of unreliable measurements.

Response to corrupted evidence. We replace the evidence with internally consistent but incorrect values by randomising the centroid, re-deriving the dependent quantities, flipping the mass direction, and randomising mirror similarity. Naive pre-question injection then falls to 33.6%, which is 15.9 points below the no-evidence baseline, whereas Channel A yields 48.4%. The calibrated interface therefore relies less on the block when its fields conflict with the image.

Component-wise corruption separates evidence quality from interface robustness. Position, mass, and mirror corruption leave Channel A at 54.1%, 51.1%, and 56.4%, so mass is both the most useful cue and the largest error source, consistent with its 67.8% precision and the remaining Balanced errors. The 7.8-point clean-to-corrupt drop shows that the tools matter, and the one-point distance to C0 shows that they no longer overwhelm the image.

Clean accuracy and corrupted-evidence fallback. Let A(∅)A(\varnothing), A(E)A(E), and A(E~)A(\widetilde{E}) denote image-only, measured, and corrupted-evidence accuracy. The difference A(E)−A(∅)A(E)-A(\varnothing) measures the useful signal a channel extracts, and A(E~)−A(∅)A(\widetilde{E})-A(\varnothing) measures how far it falls back to the visual prior. Both quantities matter, because an interface can repair a few errors while creating a much larger failure surface. Channel A moves the pair from (−3.1, −15.9) for naive injection to (+6.7, −1.0) points. The image, question, options, and decoder are fixed throughout, so the two differences isolate the model’s reliance on the injected fields.

Before/After Q&A demo showing the original shot, evidence overlay, and depth

Figure 3. A Before/After Q&A demo (shotbench-2638), drawn from the pool of items that C0 answers incorrectly and Channel A answers correctly. The supplementary material adds two further demos and a counterexample in which Channel A breaks a correct answer. Question: What’s the composition of this shot? A Right heavy B Symmetrical C Left heavy D Short side. Ground truth: A — Right heavy. Before (C0): C — Left heavy. The baseline treats the empty space on the right as visual weight. Measured evidence: Dominant-subject centroid x=0.64x=0.64; left/right visual mass 0.36/0.64; difference 0.27 (78th training percentile). After A / B: A — Right heavy / A — Right heavy. What changed: Calibrated geometry distinguishes occupied visual mass from negative space, reversing the directional error.

Evidence conditionAcc. (%)Δ\Delta vs. C0
C0: no evidence49.5—
Naive: clean46.3−3.1
Naive: all fields corrupted33.6−15.9
Channel A: clean56.2+6.7
position corrupted54.1+4.6
visual mass corrupted51.1+1.6
mirror similarity corrupted56.4+6.9
all fields corrupted48.4−1.0

Table 4. Clean utility and corrupted-evidence fallback are distinct. Naive injection fails both tests; calibrated rendering gains on clean evidence and returns close to the image-only prior when every field is wrong.

RQ3: Other Models and a Second Benchmark

Five models improve, two stay approximately flat, and two decline (Figure 4), and the gains are not monotone in baseline accuracy. The primary model with Channel A (56.2%) stays below all three commercial baselines, and Channel B (61.0%) exceeds gpt-5.4 and gemini-3.1-pro but remains 2.1 points below claude-opus-4-8 on unrounded accuracies. Channel B requires embedding access and was not run on the commercial APIs, so these are contextual comparisons and not paired ones. LLaVA-OneVision cannot reliably map the supplied directions to labels, and claude-opus-4-8, the strongest baseline in the group, loses 6.9 points under Channel A. Evidence helps only when a model can reconcile it with its own prior.

Cross-benchmark transfer. On the 120-item CineTechBench composition subset, zero-shot accuracy is 65.0% and Channel A reaches 64.2%, a difference of one item. Central rises from 78.6% to 100%, and Rule of Thirds drops by 3.5 points because the leakage guard suppresses answer-vocabulary wording. Transfer is neutral overall, and the outcome depends on how the geometry maps onto a benchmark’s terminology.

Effect on Non-Composition Questions

Appending composition evidence to the 3,093 questions in ShotBench’s seven non-composition dimensions raises accuracy from 52.9% to 56.4%. No dimension drops by more than one point (camera angle, −0.9-0.9), while lens size rises by 10.6 and lighting by 6.4. Roughly 2–3 of the aggregate points come from the answer-format instruction, which reduces parse failures from 3.8% to zero, and we do not attribute that part to geometry.

Discussion

The experiments show a trade-off between the model’s own prior and the injected evidence. Measurements placed before the task change how the model reads the image and question, while post-question placement with gating preserves more correct baseline predictions and still gains on the directional classes.

Evidence for the Integration Interpretation

Four controls point to the interface rather than to the geometry alone. The rule classifier and the evidence-free chain-of-thought control both stay below C0. Rewording the block at the pre-question position also leaves accuracy below C0, while moving the same facts after the question adds 7.1 points. Corruption affects the naive and the calibrated interface differently with the image and decoder fixed. Both channels consume the same vector and freeze the same VLM, yet Channel A gates statements and Channel B learns four input embeddings, which weakens a template-only explanation of the gain. Taken together, these controls locate the effect in the interface: what changes across conditions is how much of the injected block the model acts on.

We identify three error sources. Measurement error comes from a wrong box or pseudo-saliency map; mass corruption quantifies it. Integration error means a model over-follows a value or cannot map it to an option; prompt position and the nine-model study expose it. Terminology error leaves visible geometry disconnected from a cinematic label, as with Short side. These require, respectively, better perception, trust calibration, and targeted language grounding.

The persistent Short side failure separates visual evidence from language grounding. Every VLM, including the trained specialist, scores zero on this class under C0 and Channel A, while an F9 rule reaches 27.5%. ShotQA contains 39 Short side and 89 Symmetrical samples against roughly 1,700 per frequent class. The geometry is therefore recoverable from the features while the cinematic label itself remains underlearned.

Choosing Between the Two Interfaces

The two channels share the extraction stack and the 24-dimensional vector, and they occupy different deployment points. Channel A changes only the text sent to the model, which suits closed APIs and settings where a human has to inspect the facts that were exposed. Its cost is model dependence, since the gates control the reliability of the measurements and not the willingness of the receiving model to follow them. Channel B requires access to the input embeddings and improves the same C0 model by 11.5 points while training 0.1% of a 7B backbone. Added capacity does not help by itself, since an eight-token, one-epoch variant reaches 60.1%, below the 61.0% of the final setting.

A change of deployment distribution requires re-estimating the training-distribution percentiles and re-checking the receiving model on that distribution.

Threats to Validity and Scope

Construct validity. Composition labels compress a subjective visual judgement into one option, and exact-match accuracy cannot represent plausible secondary readings of an image. Our nine features cover subject placement, balance, symmetry, depth, and prominent lines, and they omit colour, motion, gaze, semantic emphasis, and temporal editing. The zero accuracy on Short side and the decline on Balanced mark that boundary.

Internal validity. All primary conditions use the same 479 items, official prompt, greedy decoder, and paired transition analysis; experts and VLM weights stay frozen, gates use only ShotQA training statistics, and test labels never enter feature construction. These controls do not eliminate selection effects from design iteration. The 3×33 \times3 threshold sweep and the fixed train/validation split show local stability; we did not run nested model selection across independent datasets. Detector failures are correlated, since a wrong dominant box changes centroid, area, depth, and visual mass together, and our all-field corruption reproduces one such failure family.

Statistical validity. The exact McNemar test applies because the systems answer the same questions, and the primary Channel A–C0 comparison remains significant (p=0.0018p=0.0018). With 479 questions the power is limited for rare labels and small model differences. Channel B exceeds Channel A by 4.8 points, but their paired comparison is borderline (p=0.058p=0.058). The per-class, threshold-sweep, and nine-model results are therefore diagnostic, and we report transition counts alongside them.

External and deployment validity. CineTechBench is neutral overall, which shows that a taxonomy change can erase the primary gain, and two of nine models regress. Neither scale nor a strong zero-shot score guarantees that a new VLM will consult the block appropriately. Channel A gates individual lines but never abstains, so an undetected and confidently wrong measurement can still change an answer. A new camera domain, language, or label set requires re-estimating evidence precision and validating the receiving model.

Conclusion

On ShotBench composition, pre-question evidence lowers clean accuracy and follows corrupted measurements. Post-question gating improves the frozen VLM by 6.7 points, and a soft-token projector improves it by 11.5 points. Across nine models the same text interface produces gains, negligible changes, and regressions. Explicit geometry therefore helps when its interface is calibrated both for the reliability of the measurements and for the receiving model.

Acknowledgments

This work was supported by the Innovation Fund Project for Graduate Students of Bohai University.

Competing Interests

The authors have no competing interests to declare that are relevant to the content of this article.

References

  1. [1]Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, et al. Flamingo: A visual language model for few-shot learning. In NeurIPS, volume 35, 2022.arxiv.org/abs/2204.14198
  2. [2]Shuai Bai et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025.arxiv.org/abs/2502.13923
  3. [3]Boyuan Chen, Zhuo Xu, Sean Kirmani, et al. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, pages 14455–14465, 2024.arxiv.org/abs/2401.12168
  4. [4]Michael Dorkenwald, Nimrod Barazani, Cees G. M. Snoek, and Yuki M. Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In CVPR, pages 13548–13558, 2024.arxiv.org/abs/2402.08657
  5. [5]Haoan Feng, Sri Harsha Musunuri, and Guan-Ming Su. Geometry-guided camera motion understanding in videollms. In CVPR Workshops, pages 8423–8433, 2026. arXiv:2603.13119.arxiv.org/abs/2603.13119
  6. [6]Yonatan Geifman and Ran El-Yaniv. Selectivenet: A deep neural network with an integrated reject option. In ICML, volume 97, pages 2151–2159, 2019.arxiv.org/abs/1901.09192
  7. [7]Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In ICML, volume 70, pages 1321–1330, 2017.arxiv.org/abs/1706.04599
  8. [8]Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, pages 14953–14962, 2023.arxiv.org/abs/2211.11559
  9. [9]Qingqiu Huang et al. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. arXiv:2007.10937.arxiv.org/abs/2007.10937
  10. [10]Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700–6709, 2019.arxiv.org/abs/1902.09506
  11. [11]Menglin Jia, Luming Tang, Bor-Chun Chen, et al. Visual prompt tuning. In ECCV, pages 709–727, 2022. doi: 10.1007/978-3-031-19827-4_41.DOI
  12. [12]Justin Johnson, Bharath Hariharan, Laurens van der Maaten, et al. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, pages 2901–2910, 2017.arxiv.org/abs/1612.06890
  13. [13]Amita Kamath, Jack Hessel, and Kai-Wei Chang. What’s “up” with vision-language models? investigating their struggle with spatial reasoning. In EMNLP, pages 9161–9175, 2023. doi: 10.18653/v1/2023.emnlp-main.568.
  14. [14]Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, pages 3045–3059, 2021. doi: 10.18653/v1/2021.emnlp-main.243.
  15. [15]Bo Li et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024.arxiv.org/abs/2408.03326
  16. [16]Debang Li et al. Composing good shots by exploiting mutual relations. In CVPR, 2020.DOI
  17. [17]Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, volume 202, pages 19730–19742, 2023.arxiv.org/abs/2301.12597
  18. [18]Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL-IJCNLP, pages 4582–4597, 2021. doi: 10.18653/v1/2021.acl-long.353.DOI
  19. [19]Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. TACL, 11:635–651, 2023. doi: 10.1162/tacl_a_00566.DOI
  20. [20]Hongbo Liu et al. Shotbench: Expert-level cinematic understanding in vision-language models. In NeurIPS, 2025. arXiv:2506.21356.arxiv.org/abs/2506.21356
  21. [21]Nelson F. Liu, Kevin Lin, John Hewitt, et al. Lost in the middle: How language models use long contexts. TACL, 12:157–173, 2024. doi: 10.1162/tacl_a_00638.DOI
  22. [22]Shilong Liu et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In ECCV, 2024. arXiv:2303.05499.arxiv.org/abs/2303.05499
  23. [23]Long Mai, Hailin Jin, and Feng Liu. Composition-preserving deep photo aesthetics assessment. In CVPR, pages 497–506, 2016.
  24. [24]Quinn McNemar. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2), 1947.pubmed.ncbi.nlm.nih.gov/20254758
  25. [25]Anyi Rao et al. A unified framework for shot type classification based on subject centric lens. In ECCV, 2020. arXiv:2008.03548.arxiv.org/abs/2008.03548
  26. [26]Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, et al. Toolformer: Language models can teach themselves to use tools. In NeurIPS, volume 36, 2023.arxiv.org/abs/2302.04761
  27. [27]Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In ICCV, pages 11888–11898, 2023.arxiv.org/abs/2303.08128
  28. [28]Rafael Grompone von Gioi et al. Lsd: A fast line segment detector with a false detection control. IEEE TPAMI, 32(4), 2010.DOI
  29. [29]Xinran Wang, Songyu Xu, Xiangxuan Shan, et al. Cinetechbench: A benchmark for cinematographic technique understanding and generation. arXiv preprint arXiv:2505.15145, 2025.arxiv.org/abs/2505.15145
  30. [30]Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE TIP, 13(4), 2004.pubmed.ncbi.nlm.nih.gov/15376593
  31. [31]Zijun Wei et al. Good view hunting: Learning photo composition from dense view pairs. In CVPR, 2018.DOI
  32. [32]Hang Wu et al. Refineshot: Rethinking cinematography understanding with foundational skill evaluation. arXiv preprint arXiv:2510.02423, 2025.arxiv.org/abs/2510.02423
  33. [33]Hang Wu et al. Camreasoner: Reinforcing camera movement understanding via structured spatial reasoning. arXiv preprint arXiv:2602.00181, 2026.arxiv.org/abs/2602.00181
  34. [34]Kaiyu Yang, Olga Russakovsky, and Jia Deng. Spatialsense: An adversarially crowdsourced benchmark for spatial relation recognition. In ICCV, pages 2051–2060, 2019.arxiv.org/abs/1908.02660
  35. [35]Lihe Yang et al. Depth anything v2. In NeurIPS, 2024. arXiv:2406.09414.arxiv.org/abs/2406.09414
  36. [36]Zhiyuan You, Ke Wang, He Zhang, et al. Photoframer: Multi-modal image composition instruction. In CVPR, pages 10197–10207, 2026. arXiv:2512.00993.arxiv.org/abs/2512.00993
  37. [37]Xuejun Zhang et al. Predicting camera pose from perspective descriptions for spatial reasoning. arXiv preprint arXiv:2602.06041, 2026.arxiv.org/abs/2602.06041
  38. [38]Jinguo Zhu et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025.arxiv.org/abs/2504.10479

Paper details

vlnc:W_f49b5s68
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents