ant-research / Drift

What is Drift?

Drift is a reinforcement learning training framework built specifically for diffusion language models. While most RL training tools target autoregressive transformers, Drift focuses on models that generate text by progressively unmasking tokens—like LLaDA and Dream. It handles the full RLVR (reinforcement learning from verifiable rewards) pipeline: sampling outputs, executing code or verifying math answers as reward signals, and updating the model with policy gradient methods.

The tool runs entirely on PyTorch with DeepSpeed ZeRO-3 for memory-efficient training across multiple GPUs or nodes. Users configure experiments through YAML files, point at pretrained models, and launch training with a single Accelerate command.

Why is not gaining traction?

Diffusion language models are a hot research area, but most teams are stuck using RL tools built for autoregressive models. Drift solves this mismatch directly. It comes with built-in reward functions for code execution and math verification out of the box—no need to wire up your own test runners. The block-wise parallel decoding is particularly useful for speeding up generation during rollouts, which matters when you're sampling thousands of responses per training step. With support for multiple masking strategies and temperature-based sampling, it's flexible enough to experiment with different denoising schedules.

Who should use this?

NLP researchers working on training or fine-tuning diffusion language models with RL. Specifically: anyone trying to improve code generation on HumanEval/MBPP, improve math reasoning on MATH500, or experiment with RLVR on custom tasks. If you're evaluating diffusion LLMs for production use and need better training tooling, this is worth a look. Academic labs exploring multi-node RL training for large diffusion models will find the distributed setup straightforward.

Verdict

Drift solves a real gap in the ML tooling landscape and comes from Ant Group's research division, which adds credibility. However, with only 19 stars and minimal documentation, expect to do some spelunking to get things running. The framework is functional but clearly early-stage—no polished tutorials, limited examples beyond the config files. If you're an ML practitioner comfortable reading YAML configs and debugging training scripts, it's a solid starting point. For teams needing production-ready stability, wait for more community iteration. The 1.0% credibility score reflects this: credible source, nascent project.