CyberArena is a live benchmark for symmetric attack-and-defense CTF for frontier coding agents.

DR.DOCBENCH is a difficulty-aware benchmark for expert-level document parsing.

SuperGPQA is a large-scale and highly challenging benchmark created to evaluate the advanced reasoning capabilities of Large Language Models (LLMs). Its purpose is to test model performance on expert-level, graduate qualification questions across an unprecedented 285 academic and professional disciplines.

KINA is a high-density knowledge benchmark encompassing 261 fine-grained disciplines, the first to incorporate disciplinary representativeness as a core metric. It features a reusable, game-theoretic data collection pipeline that mitigates annotation vulnerabilities.

Discover VeriWeb, a pioneering benchmark for long-horizon web agents. It offers a reproducible environment with subtask verification to rigorously evaluate long-chain web agent performance.

OmniDocBench is a comprehensive benchmark for evaluating AI in document parsing and content extraction across diverse real-world PDF documents.

Discover PIN, a new data format and two large-scale datasets (PIN-200M & PIN-14M) designed to help Large Multimodal Models learn richer, more grounded representations.

EDITREWARD is trained with our new large-scale human preference dataset, meticulously annotated by trained experts following a rigorous protocol containing over 200K preference pairs.

VideoScore2 is a multi-dimensional, interpretable, and human-aligned framework that explicitly evaluates visual quality, text-to-video alignment, and physical/common-sense consistency while producing detailed chain-of-thought rationales.

Knowledge-Orthogonal Reasoning Benchmark (KOR-Bench) is designed to evaluate models' intrinsic reasoning and planning abilities by minimizing interference from pretrained knowledge.

MMAR (Massive Multi-disciplinary Audio Reasoning) is a new and challenging benchmark designed to evaluate the deep audio understanding and reasoning capabilities of AI models.

OmniHD-Scenes is a massive multimodal autonomous driving dataset featuring 450K+ synchronized frames with comprehensive sensor data for next-generation perception research.

Chain-of-Agents (CoA) is a novel framework for training end-to-end agent foundation models (AFM) using multi-agent distillation and agentic reinforcement learning.

YuE is a family of open foundation models based on the LLaMA2 architecture that generates up to five minutes of music while maintaining lyrical alignment, coherent musical structure, and engaging vocal melodies.

AFM-Datasets is the official training dataset released alongside Chain-of-Agents, an end-to-end chain-of-agents framework for training agent foundation models.

TaskCraft is a multi-modal benchmark dataset featuring tasks ranging from simple (1-step) to expert-level (multi-step) agentic task processing.

COIG-Writer is a novel Chinese creative writing dataset that captures both diverse outputs and their underlying thought processes through systematic reverse-engineering of high-quality texts.

AutoKaggle is a powerful and user-centric framework that assists data scientists in completing daily data pipelines through a collaborative multi-agent system.

Matrix is a massive, open-source pretraining dataset containing approximately 4.7 trillion tokens, designed as a comprehensive bilingual resource for large language model pretraining.

CriticLeanBench is a specialized benchmark designed to evaluate the critical reasoning of AI models in the domain of mathematical formalization using the Lean4 proof assistant.

FormalMATH is a large-scale benchmark designed to evaluate and advance the capabilities of Large Language Models in formal mathematical reasoning using the Lean4 proof assistant.

COIG-P (Chinese Open Instruction Generalist - Preference) is a high-quality, large-scale Chinese preference dataset for aligning large language models to human values.