Research map / Detection and segmentation / Open-set and few-shot detection

Detect Anything in Graphic Design: Element-Level Rewards for Autoregressive Detection

NeurIPS 2026 · Detection and segmentation › Open-set and few-shot detection · approach: Industrial anomaly detection · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DAD, an autoregressive VLM detector for graphic designs that predicts elements in compositional order with amodal boxes, trained with EleRPO, an element-level GRPO variant using martingale-difference rewards.

Problem

Graphic design detection treats layered elements as an unordered set and only detects visible regions, so compositional order and occluded parts are lost, hindering layer-level decomposition and editing.

Approach

Decode elements in compositional (back-to-front) order with amodal boxes, and convert the sequence-level F1 reward into unbiased element-level rewards via martingale difference decomposition, estimated from rollout clusters grouped by prefix F1 without extra reward models or re-rollouts.

Key results

• On the DAD test set, DAD achieves F1@[.5:.95] 72.57 and occluded recall 65.35 versus 40.89 and 34.26 for GPT-5.5.
• EleRPO reaches 79.7 IoU-weighted F1 on the DAD test set versus 74.2 for GRPO and 73.3 for PRIME.
• Human evaluation: DAD's amodal box IoU falls within the inter-annotator IoU range across occlusion levels, while GPT-5.5 and Gemini-3.1-Pro fall below it.

Surprising findings

• Policy-signal-based fine-grained RL methods (GRPO-lambda, GTPO, PRIME) do not improve over GRPO, and re-rollout methods (VinePPO, SPO) perform worse than GRPO on detection.

Limitations

• Evaluated on graphic design data where elements have layer annotations; ground truth for the main dataset is partly synthetic or annotated by Gemini-3-Pro.
• Element-level reward relies on approximating expected improvements via prefix-F1 clustering (Pearson 0.81 with a re-rollout estimator).

Experimental setup

Datasets: DAD dataset (10.4M designs, internal and augmented) Design-Multi-Layer-Bench Crello PrismLayersPro LICA ChartGalaxy COCO LVIS BDD100K Objects365 Rico PosterLayout

Benchmarks: DAD test set Design-Multi-Layer-Bench Crello PrismLayersPro LICA ChartGalaxy Rico PosterLayout DocLayNet COCO LVIS BDD100K

Metrics: F1@[.5:.95] IoU-weighted F1 Occluded R@[.5:.95] POA PSNR SSIM RGB L1 Alpha SIoU inference time

Baselines: GPT-4o GPT-5 GPT-5.4 GPT-5.5 Gemini-3.1-Pro GLIP-L Detic-B Grounding DINO-B Florence2-L Qwen3-VL-2B Qwen3-VL-8B InternVL3.5-2B

Builds on: GRPO Qwen3-VL PRIME VinePPO Rex-Omni

Closest papers by meaning

• Decomposition of Graphic Design with Unified Multimodal Model (ICML 2025 · similarity 0.86)
• Spot the Error: Non-autoregressive Graphic Layout Generation with Wireframe Locator (AAAI 2024 · similarity 0.85)
• Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · similarity 0.83)
• Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.83)
• Graphic Design with Large Multimodal Model (AAAI 2025 · similarity 0.83)
• Few Shot Part Segmentation Reveals Compositional Logic for Industrial Anomaly Detection (AAAI 2024 · similarity 0.83)