Research / Open work

Research,
shared openly.

Selected contributions by our team, spanning agent collaboration, generative evaluation, and large-scale training trajectories. This work informs how we design benchmarks and data for the next generation of agents.

Ideas → experiments → shared progress

08 projects

Selected project / Agents

AFlow

Automating agent workflows

Zhang et al. · ICLR 2025 Oral

Represents workflows as code and automatically improves them through search and execution feedback.

Explore the project

Agents

MetaGPT

Multi-agent collaboration

Hong et al. · ICLR 2024 Oral

Encodes standard operating procedures into multi-agent workflows to support complex task decomposition and verification of intermediate results.

Read the paper

Evaluation

Video-Bench

Human-aligned video evaluation

Han et al. · CVPR 2025

Evaluates video generation quality with a multidimensional prompt suite and assessment strategies built around multimodal language models.

Read the paper

Evaluation

Terminal-Bench

Challenging terminal tasks

Merrill et al. · 2026

Measures long-horizon agent capabilities through real terminal tasks, executable environments, and independent verifiers.

Explore the benchmark

Training data

OpenThoughts

Data recipes for reasoning

Guha et al. · 2025

Opens up reasoning-data recipes and training and evaluation workflows to study how data selection shapes model capabilities.

Explore the project

Training data

OpenThoughts-Agent

Data recipes for agents

Raoof et al. · 2026

Uses open tasks, trajectories, and verification environments to investigate how to compose and scale training data for general-purpose agents.

Explore the project

Evaluation

Harbor-Index

A compact agentic benchmark

Shi et al. · 2026

Developed with contributions from Kevin Xiang Li, combining difficulty filtering, human audits, and an iterative process for improving task quality.

Explore the benchmark

Training data

SWE-ZERO-12M

Execution-free agentic trajectories

Li · 2026

Released by Kevin Xiang Li, a collection of 12.29 million agentic coding trajectories generated without executing repository-specific tests or commands.

Explore the dataset

Better questions. Shared progress.

Interested in a research collaboration?

Get in touch