Research map / Multi-agent systems and games / LLM agents

Agentic Design Review System

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 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 Agentic-DRS, a training-free multi-agent MLLM system (meta, static and dynamic agents) for graphic design review, using graph-matching exemplar retrieval (GRAD) and structured design descriptions (SDD), evaluated on the new DRS-Bench.

Problem

Holistic evaluation of graphic designs across many attributes (alignment, composition, color, typography) and generation of actionable feedback, which heuristic and learned scoring methods handle poorly.

Approach

Combines graph-based in-context exemplar retrieval (Wasserstein plus Gromov-Wasserstein distances over CLIP element embeddings) and structured textual design descriptions with a meta-agent that plans static and dynamic reviewer agents and summarizes their ratings and feedback, all without weight updates.

Key results

• Agentic-DRS with GPT-4o reaches 75.29% accuracy on Afixa and 69.53% on Infographic, versus 62.91% and 58.26% for vanilla GPT-4o.
• On GDE, Agentic-DRS with GPT-4o improves Pearson correlation for alignment from 0.597 to 0.722 over the single-agent baseline.
• Ablation shows adding GRAD and SDD progressively raises Infographic accuracy from 63.11% to 69.53% for the GPT-4o variant.

Limitations

• Fails to detect alignment problems under small perturbations.
• Relies on proprietary MLLMs (GPT-4o, Gemini-1.5-Pro) and API calls.
• Small benchmark sizes (55 to 137 designs for several datasets).

Experimental setup

Datasets: GDE Afixa Infographic Internal Design Dataset (IDD)

Benchmarks: DRS-Bench (GDE, Afixa, Infographic, IDD)

Metrics: Accuracy Sensitivity Specificity Pearson correlation AIM

Baselines: Heuristic-based Evaluation Gemini-1.5-Pro GPT-4o Gemini-1.5-Pro + GRAD GPT-4o + GRAD + SDD

Builds on: GPT-4o Gemini-1.5-Pro CLIP Gromov-Wasserstein distance

Closest papers by meaning

• Agent Reviewers: Domain-specific Multimodal Agents with Shared Memory for Paper Review (ICML 2025 · similarity 0.82)
• Agent-as-a-Judge: Evaluate Agents with Agents (ICML 2025 · similarity 0.81)
• How can we assess human-agent interactions? Case studies in software agent design (ICML 2026 · similarity 0.81)
• Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.81)
• Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations (ICLR 2026 · similarity 0.81)
• SketchMind: A Multi-Agent Cognitive Framework for Assessing Student-Drawn Scientific Sketches (NeurIPS 2025 · similarity 0.81)