Open research systems

Research that becomes infrastructure.

DIG develops reusable systems together with the Gen-Verse open research ecosystem. We view code, environments, training stacks, models, and evaluation tools as research artifacts — not merely supplements to papers.

From ideas to reusable research systems.

The portfolio spans reasoning, reinforcement learning, interactive agents, environments, multimodal diffusion language models, and post-training infrastructure.

Reasoning · Process Reward

ReasonFlux

A family of systems for hierarchical reasoning, thought templates, trajectory-aware process reward models, and co-evolving coders.

GitHub ↗
General Agentic RL

RLAnything

Dynamic environment, policy, and reward learning for general reinforcement learning and agent training.

GitHub ↗
Interactive Agent RL

OpenClaw-RL

Fully asynchronous reinforcement learning for interactive agents operating across conversation, terminal, GUI, SWE, and tool-use settings.

GitHub ↗
Environment Learning

GenEnv

Environment construction and evolution for agent learning, exploration, and environment–policy co-evolution.

GitHub ↗
Multimodal dLLM

MMaDA

Multimodal large diffusion language models and systems for unified multimodal generative modeling.

GitHub ↗
Diffusion LLM RL

dLLM-RL

Trajectory-aware post-training and reinforcement learning infrastructure for diffusion language models.

GitHub ↗
Multi-Agent Reasoning

LatentMAS

Efficient multi-agent collaboration through reasoning and communication in latent space.

GitHub ↗
Open Research

Gen-Verse

The broader open research community connecting models, environments, datasets, evaluation frameworks, and algorithms.

Organization ↗
Discovery Intelligence

DIG

The academic research group connecting learning, evolution, and discovery into a unified long-term research program.

Research agenda →
Open by design

A useful research artifact should make the next experiment easier to run, the next hypothesis easier to test, and the next researcher faster to start.