Primary metric
Overall9
Prompt-macro average over the canonical ten applicable FF-judge+PP components, reported on a 0–100 scale. Inapplicable components are skipped at the prompt level before aggregation.
Results
Model leaderboard
Loading benchmark data…
Click any column header to sort. Test-mini, Mini-easy, and Mini-hard are overlapping tagged subsets, not a partition. Domains and failure modes are multi-label, so group counts overlap. Finalized 3d5 scores use present-only aggregation; coverage shows the number of valid evaluator sidecars.
Reference
Domain abbreviations
Twenty-two content domains represented in AgentGen Bench.
ABSAbstract — Abstract concepts, emotions, and symbolism.ARCArchitecture — Buildings, structures, interiors, and urban spaces.ARTArt & Design — Artistic styles, illustration, visual design, and aesthetics.COMCommercial — Advertising, branding, marketing, and product imagery.CULCulture — Cultural artifacts, traditions, heritage, and customs.EDUEducation — Educational diagrams and instructional content.FASFashion — Apparel, accessories, style, and textiles.FOOFood — Cuisine, dishes, ingredients, and culinary arts.GAMGaming — Video games, virtual worlds, and game characters.HISHistory — Historical events, figures, periods, and artifacts.INFInformation Design — Data visualization, infographics, and diagrams.MEDMedical — Anatomy, medical illustration, and health.MILMilitary — Weapons, military equipment, and warfare.MUSMusic — Instruments, performances, and music-related imagery.OBJObjects — Products, everyday objects, and artifacts.OTHOther — Miscellaneous or uncategorized content.PEOPeople — Portraits, human figures, and fashion on people.PSYPsychology — Mental states and psychological concepts.SCIScience & Nature — Biology, physics, natural phenomena, and scientific illustration.SPOSports — Athletics, sports, fitness, and recreation.TECTechnology — Devices, interfaces, software, and technical diagrams.TXTText & Typography — Text-heavy content, calligraphy, typography, and signage.Reference
Failure-mode abbreviations
Twelve recurring categories distilled from real production requests.
TK-RTemporal — Recent — Newly emerged or recently changed entities, events, products, and trends.TK-CTemporal — Current — Up-to-date states such as rankings, live events, locations, or social conditions.EIKEntity & IP Knowledge — Accurate identity and appearance of named people, characters, brands, products, or landmarks.CSVConcept & Symbol Visualization — Precise concepts, flags, symbols, and diagrams requiring exact representation.FHAFactual & Historical Accuracy — Correct facts, periods, artifacts, scientific details, and historical depictions.CSCultural Specificity — Authentic culture-bound practices, aesthetics, artifacts, and conventions.VURVisual / UI / UX — Specific interfaces, layouts, rendering conventions, spatial logic, and visual relationships.DVZData Visualization — Accurate charts, graphs, timelines, diagrams, and infographics.THGText / Typography — Correct, readable, and stylistically appropriate text rendered inside images.CCSComplex Composite — Coherent multi-element scenes, spatial arrangements, and cross-concept compositions.VUAVague / Abstract — Visualizing underspecified ideas, moods, metaphors, and abstract concepts.IRRImplicit Reasoning — Requirements that must be inferred rather than stated directly, including contextual and relational constraints.Inside the generator's knowledge boundary
Production prompts selected where strong generators already perform well without retrieval.
Outside the knowledge boundary
387 visually search-intensive and 264 textually search-intensive prompts expose gaps that retrieval can address.