A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Unknown
An in-depth exploration of factors affecting multi-modal in-context learning, using semantic diversity analysis to understand performance drivers.
Unknown
The paper investigates whether quality-diversity archive coverage can certify the robustness of process reward models under stress tests, finding that coverage alone is insufficient.
Unknown
The paper compares input-level test-time augmentation against output-level sampling for LLMs at matched compute, finding input diversity yields better accuracy.
Zhouyuan Ma, Yutao Wu, Hanxun Huang, et al.
HarmProfile is a content-centric benchmark dataset of over 80,000 validated harmful artifacts from 23 frontier LLMs, defining model-level risk profiles and revealing that harmfulness and diversity grow with model capability.
Ananya Sahu, Mohit Bansal, Elias Stengel-Eskin
CreativeInstruct teaches LLMs to balance creativity and quality via special [StartCreativity] spans, matching or exceeding diversity baselines without sacrificing quality.
Unknown
Cycle3D proposes a unified 3D generation framework that enhances diversity and texture consistency through a generation-reconstruction cycle during the denoising process.
Unknown
This paper investigates in-context learning in world models, showing that context and diversity are key to enabling self-adaptation to novel or rare configurations.
Bohan Li, Yutai Hou, Wanxiang Che
This survey categorizes data augmentation methods in NLP into paraphrasing, noising, and sampling based on the diversity of augmented data, and reviews their applications and challenges.
Ilias Driouich, Hongliu Cao, Eoin Thomas
Introduces a multi-agent framework for generating synthetic QA datasets for RAG evaluation, prioritizing semantic diversity and privacy preservation.
Kaiyu Li, Zepeng Xin, Zixuan Jiang, et al.
Introduces OVEarth-Bench, a benchmark for open-vocabulary Earth observation with broad category coverage and diverse query types, revealing current method limitations and MLLM superiority.
Siyuan Huang, Pengyu Cheng, Haotian Liu, et al.
Skill Self-Play (Skill-SP) co-evolves a proposer, solver, and skill controller via RL to reconcile structured verification with open-ended task diversity for LLM self-evolution.
Nuo Chen, Yicheng Tong, Yuzhe Yang, et al.
This paper systematically studies diversity collapse in multi-agent LLMs, showing that interaction structures, not model insufficiency, primarily cause reduced exploration diversity.