6arXiv cs.AI (Artificial Intelligence)·
SearchGen-20K benchmark and teach-then-search co-training framework for world-knowledge-grounded visual generation
Researchers introduce SearchGen-20K and SearchGen-Bench, a dataset of 20,839 prompts across 12 failure categories targeting the world-knowledge bottleneck in visual generation, paired with a 1M-item multimodal search corpus. Frontier open visual generators score only 21–28/100 on the new benchmark, a gap invisible to existing evaluations. The paper proposes a teach-then-search co-training framework that discovers a model's evolving knowledge boundary and uses search tools selectively, achieving monotonic improvement and laying groundwork for recursive self-improvement in agentic visual generation. All datasets and corpora are released publicly.
Related guides (3)
Related events (8)
MAGMaR 2026 Shared Task findings on multimodal retrieval and grounded generation
The overview paper for the MAGMaR 2026 shared task reports results from two tracks: video retrieval and grounded article generation from retrieved videos. Two teams submitted 17 retrieval systems, all surpassing last year's winning baseline, while four teams submitted 16 generation systems with each team producing at least one human-preferred report. The shared task advances evaluation of multimodal retrieval-augmented generation pipelines combining video and text.
Meta Research Improves Image Generation via Staged Planning and Self-Revision Fine-Tuning
Researchers from Meta and collaborating universities propose a fine-tuning method that teaches image generators to compose images through discrete plan-sketch-inspect-refine cycles rather than generating all at once. Starting from BAGEL-7B, they construct ~62,000 training examples using GPT-4o and FLUX.1 Kontext to supervise each stage, achieving 83% on GenEval versus 77% for the base model and a competing method (PARM) that required 11x more training data and ~8x more inference steps. The approach improves spatial relationship accuracy, object attribute fidelity, and real-world knowledge grounding in generated images.
Semantic Generative Tuning (SGT) for Unified Multimodal Models
This paper introduces Semantic Generative Tuning (SGT), a post-training paradigm for unified multimodal models (UMMs) that bridges the gap between visual understanding and visual generation. The authors find that image segmentation tasks serve as optimal generative proxies, providing structural semantics that improve both perception and generative layout fidelity. SGT aligns representation spaces across understanding and generation objectives, improving feature linear separability and visual-textual attention allocation. Evaluations show consistent gains on multimodal comprehension and generative fidelity benchmarks.
SkillGenBench: Benchmarking Skill Generation Pipelines for LLM Agents
SkillGenBench is a new benchmark designed to evaluate the ability of LLM agents to generate correct, reusable, and executable skills from raw repositories and documents, rather than merely using pre-provided skills. It covers two generation regimes (task-conditioned and task-agnostic) and two procedural sources (repository-grounded and document-grounded), with standardized execution-based evaluation protocols. Experiments across multiple skill-generation methods reveal substantial performance variation and distinct failure modes depending on source type. The benchmark aims to establish skill generation as an independent research problem within agent systems.
WikiVQABench: A Knowledge-Grounded Visual Question Answering Benchmark from Wikipedia and Wikidata
WikiVQABench is a new human-curated VQA benchmark that requires external knowledge beyond visual perception, constructed by combining Wikipedia images, captions, and Wikidata structured knowledge with LLM-generated question candidates reviewed by human annotators. The benchmark evaluates knowledge-intensive reasoning in vision-language models, covering 15 VLMs ranging from 256M to 90B parameters. Accuracy spans 24.7% to 75.6%, indicating meaningful discrimination across model scales. The dataset and code are publicly released.
GPIC: Stanford Releases 28-Trillion-Pixel Permissively Licensed Image Corpus for Visual Generation Research
Stanford Vision Lab introduces GPIC, a Giant Permissive Image Corpus of approximately 28 trillion pixels comprising 100M training, 200K validation, and 1M test images, all permissively licensed for research and commercial use. Images are captioned by a state-of-the-art vision-language model, safety-filtered, deduplicated, and hosted on Hugging Face. The release includes a benchmarking protocol for generative modeling and a reference baseline using pixel-space flow matching. The dataset addresses a key gap in scalable visual generative modeling research by providing a large, stable, and openly licensed resource.
Peak-Then-Collapse: RLVR Tool-Use Failures on Knowledge-Graph APIs
This paper investigates RLVR-based tool-use training (GRPO on Qwen2.5-7B-Instruct) on a minimal knowledge-graph API (Freebase over Complex WebQuestions) and documents a 'peak-then-collapse' pattern where tool-grounded answer rates rise then fall to zero within 50 steps, replicated across four seeds and seven reward designs. The authors identify a key structural difference between knowledge-graph APIs and other tool types (Python, web search, JSON): sparse, non-natural-language feedback signals (e.g., empty brackets '[]') prevent the model from recovering via pretraining-familiar error signals. A direct oracle ablation shows relation selection is not the bottleneck—95.4% of errors are retrieval-composition failures—and self-distillation reaches 40% EM at 7B, with capacity scaling to 14B yielding only marginal gains, suggesting an interface-bound ceiling.
TailorMind: Preference-aligned multimodal content generation via hypergraph collaborative filtering
TailorMind is a new system for personalized multimodal content generation that translates user behavioral traces into generation-ready preferences without relying on existing item pools or user-generated content. The approach combines hypergraph collaborative filtering for sparse user histories, ranking-error feedback with textual gradient descent for profile optimization, and retrieval-augmented style control. The authors introduce TailorBench, a benchmark across three platforms evaluated on coherence, novelty, aesthetic quality, hallucination, and profiling, reporting up to 29% Recall gains in reranking over baselines.