Agentic-RL

Agentic-RL is a research integration for multimodal agentic search with reinforcement learning. The original project work is focused on the rllm/vision_deepresearch_async_workflow/ pipeline:

• remote-sensing multimodal DeepResearch workflow and tool execution
• Qwen3-VL training launch recipes over rLLM / veRL / Megatron
• failure-aware trajectory masking for long tool-call chains
• reward composition for answer correctness, tool/query utility and format quality

Large frameworks (rllm, verl, Megatron-LM, mbridge) are vendored only to make the training integration reproducible. See THIRD_PARTY_NOTICE.md for attribution and license boundaries.

Start

Install the main RL package:

cd rllm
pip install -e .

Run a single-node Qwen3-VL training job:

bash vision_deepresearch_async_workflow/run/qwen3-vl-8b-single-node.sh

Other launch scripts are available in rllm/vision_deepresearch_async_workflow/run/.

Reproducibility Notes

The repository excludes generated package metadata, local parquet datasets, compiled objects, checkpoints and trajectory dumps. Prepare datasets with the scripts in rllm/vision_deepresearch_async_workflow/data_prepare/, then run the launch script that matches your GPU topology.

The 8B recipes currently use RLOO-style grouped rollouts, while the 2B recipe keeps a GRPO configuration. Resume descriptions should use the more precise phrase "failure-aware trajectory masking / advantage handling" unless referring specifically to the 2B GRPO script.