Resources

This page collects selected public resources from my recent work, including systems, models, benchmarks, datasets, and research platforms. For a complete list of papers, please see Publications.

Agents and research tools

• A5 — a proactive personal assistant that builds a long-term understanding of the user and their workspaces, notices changes in project materials, and suggests next steps.
• Atria Dawn — an open-source agentic model for research and engineering workflows, built on a 744B-parameter MoE foundation model with a 256K context window.
• Agentic Harness Engineering (AHE) — a framework for letting a coding agent automatically evolve its own harness and transfer the evolved harness to new tasks and model families.

Benchmarks, datasets, and platforms

• WorkForge — a framework for constructing verifiable long-horizon work-agent environments from real-world resources, with 16.7K environments across 40 professional domains.
• CL-bench — a benchmark for context learning, with 500 expert-crafted contexts, 1,899 tasks, and detailed verification rubrics.
• SocioVerse2 — an open-source research cockpit for social science, supporting interventions, counterfactual branches, and long-horizon social simulation.