Hello everyone!
Thank you for your interest in ROLL.

ROLL has recently introduced a host of new features. Below is a summary of the recent updates. We will continue to iterate and update ROLL, and we welcome you to join the ROLL community.

🚀 New Features

• Agentic RL

  • Refactored and optimized the Agentic RL design to provide a more powerful and flexible framework.
  • Introduced multi-turn interactive local development and debugging capabilities for Agentic RL, significantly boosting development and debugging efficiency. Example: tests/agentic/env_manager/test_traj_env_manager.py
  • Added Agentic RL async training to improve training efficiency. Example: examples/qwen2.5-0.5B-agentic/agent_val_frozen_lake_async.yaml

• New Training Capabilities

  • Supported Group Sequence Policy (GSPO) with importance_sampling: Literal["token", "seq"].
  • Introduced Distill Pipeline, providing knowledge distillation capabilities. Path: roll/pipeline/distill/distill_pipeline.py
  • Added VLM Multi-domain RLVR Pipeline, enabling multi-domain joint training for multi-modal models. Path: roll/pipeline/rlvr/rlvr_vlm_pipeline.py
  • New DPO Pipeline. Path: roll/pipeline/dpo/dpo_pipeline.py
  • Supported LoRA training. Example: examples/qwen2.5-7B-rlvr_megatron/rlvr_lora_zero3.yaml
  • Added the latest math test datasets, GPQA-Diamond, and a new MultipleChoiceBoxedRuleRewardWorker.

• Other Enhancements

  • Improved the functionality for restoring checkpoints from downloaded model paths.
  • Added CLAUD.md documentation.
  • Fixed issues caused by Automap concurrency.
  • Fixed directory issues when saving Critic checkpoints.
  • Resolved vllm_strategy Qwen3 Dense FP8 compatibility issues.