v0.5.1

Release Highlights

What's Changed

• Record auto_metrics on trajectory (#440)
• Fix ruff formatting (#438)
• Document Qwen3 14B instead of Qwen2.5 14B in non-tutorial examples (#437)
• Small docs change (#433)
• Serverless documentation (#432)
• Retry GET and DELETE (#431)
• Release v0.5.0 (#428)
• Release v0.4.12 (#427)
• feat: Add VLM support (#412)

Full Changelog: v0.5.0...v0.5.1

v0.5.0

Release Highlights

What's Changed

• Release v0.4.12 (#427)

Full Changelog: v0.4.12...v0.5.0

v0.4.12

Release Highlights

What's Changed

• Add client-side error capturing (#426)
• Send wandb-artifact:/// prefix as part of inference name. (#425)
• Report serverless metrics (#424)
• feat: Add placeholder _log() implementation for ServerlessBackend (#422)
• Upgrade SkyPilot to 0.10.3.post1 (#421)
• fix: exit the loop in the monitor_openai_server (#418)
• feat: Upgrade Unsloth (#408)
• add playwright agent example to dev/ART (#402)
• feat: Update SkyPilot to 0.10.2 (#393)
• Update training details (#389)
• LocalBackend._monitor_openai_server improvement (#388)
• Add RULER scoring to LG integration example (#385)
• Show open deep research tutorial (#384)
• Minor open deep research doc edits (#383)
• Add deep research tutorial (#382)
• Show GSPO docs page (#381)
• Add MCP•RL doc, remove GSPO (#380)
• Add news item on LangGraph (#379)
• gather_trajectory_groups: fused after_each callback to group await (#378)
• Document RULER in README (#376)
• Release v0.4.11 (#374)
• Pin gql verion to fix (#372)
• Update nb links (#371)
• Allow run_checks.sh to succeed on mac (#370)
• Add art.mcp package, release 0.4.10 (#369)

Full Changelog: v0.4.11...v0.4.12

v0.4.11

Release Highlights

This patch release includes bug fixes, documentation improvements, and a new MCP package:

New Features

• MCP Package: Added art.mcp package for Model Context Protocol integration

Bug Fixes

• Fixed GraphQL dependency compatibility by pinning gql package to < 4
• Fixed run_checks.sh script to succeed on macOS

Documentation & Examples

• Updated notebook links and examples
• Improved LangGraph integration documentation
• Better documentation for wrap_rollout function

What's Changed

• Add art.mcp package (#369)
• Pin gql version to fix compatibility (#372)
• Allow run_checks.sh to succeed on mac (#370)
• Update nb links (#371)
• Properly document wrap_rollout (#368)
• Update LangGraph docs (#367)
• Link to LangGraph (#365)
• Update LangGraph integration doc (#363)

Full Changelog: v0.4.9...v0.4.11

v0.4.9 - LangGraph Integration

🕸️ LangGraph Integration - Build Smarter Multi-Step Agents

This release introduces seamless integration with LangGraph, enabling you to train sophisticated ReAct-style agents that improve through reinforcement learning without manual prompt engineering.

✨ Major Features

• 🆕 LangGraph Integration - Drop-in replacement for LangGraph's LLM initialization with automatic logging and trajectory capture
• 🔄 Multi-Step Agent Training - Train agents that reason, use tools, and adapt their behavior over time
• 📊 Auto Trajectory Generation - Automatic conversion of LangGraph agent executions into ART training data
• ⚡ RULER Compatibility - Use ART's general-purpose reward function without hand-crafted rewards

🔧 Improvements

• Type Safety - Enhanced type annotations and fixes for LangGraph integration
• Memory Management - Better CUDA cache management and garbage collection utilities
• Dependencies - Pinned litellm to version 1.74.1 for stability
• Code Quality - Refactored logger imports and async tokenizer methods

📚 Documentation & Examples

• New Documentation - Comprehensive LangGraph integration guide with examples
• Updated README - Featured LangGraph integration in main project description
• Example Notebook - ART•E LangGraph notebook for training email search agents
• License Updates - Updated third-party notices and licensing information

🔧 Code Example

import art
from art.langgraph import wrap_rollout, init_chat_model
from langgraph import create_react_agent

@wrap_rollout(model)
async def run_agent(scenario: str) -> art.Trajectory:
    agent = create_react_agent(init_chat_model(), tools)
    result = await agent.ainvoke({"messages": [("user", scenario)]})
    return art.Trajectory()  # Automatically captured

# Train with RULER - no reward engineering needed!
await art.train(model, reward_function="ruler")

🔗 Links

• 📖 LangGraph Integration Docs
• 🏋️ Try the Notebook

v0.4.8

What's Changed

Features

• Improve SkyPilotBackend by @bradhilton in 303279f
• Add integration tests in 696c230
• Add experimental standard deviation learning rate scheduling (#340) in 677d286
• Add preliminary unit testing support in 234cad9
• Add experimental support for truncated importance sampling in 3e71e4e

Bug Fixes

• Add OpenAI-compatible server monitoring with automatic process destruction after failed health check in d85f49b
• Fix tokenization issue in 94cb04c
• Disable v0 LoRA manager patching in 9c2468d
• Remove patch_patch_vllm in 44dbde9

Maintenance

• Update pyright dependency to include nodejs extras in 62f619a
• Update dependencies for Unsloth and Unsloth-Zoo to latest versions in fe526b3
• Various refactoring and code improvements

Full Changelog: v0.4.7...v0.4.8

v0.4.7

Release Highlights

What's Changed

• fix: Patch Unsloth issue (#333)
• Fix unsloth-zoo ascii encoding issue (#332)
• Release v0.4.6 (#331)
• Fix issues with dependencies not installing from git (#330)

Full Changelog: v0.4.6...v0.4.7

v0.4.6

Release Highlights

This patch release focuses on development quality improvements and bug fixes:

• Fixed dependency installation issues: Resolved problems with dependencies not installing properly from git repositories
• Added Pyright type checking: Enhanced code quality with static type analysis
• Improved documentation: Added guidance for starting RL training from existing SFT LoRA adapters

What's Changed

• Fix issues with dependencies not installing from git (#330)
• feat: Add Pyright type checking (#326)
• docs: Show how to start RL from an existing SFT LoRA adapter (#325)
• Add MCP•RL news link (#324)

Full Changelog: v0.4.5...v0.4.6

v0.4.5

Release Highlights

🎯 New Algorithm: GSPO (Group Sequence Policy Optimization) - Introducing the stable and efficient RL algorithm used to train state-of-the-art models like Qwen3-235B-A22B-Instruct-2507. GSPO improves training stability for Mixture-of-Experts models with sequence-level optimization and infrastructure-friendly design.

🚀 Major Performance Improvement: Upgraded to vLLM 0.10.0 with significant startup time improvements due to frozen garbage collection during CUDA graph capture (vLLM PR #21146)

🔧 Enhanced Developer Experience: Improved release workflow with human-curated release notes

What's Changed

• feat: Add GSPO support (#273)
• feat: Upgrade vLLM to v0.10.0 (#302)
• feat: Update Unsloth dependencies to 2025.8.1 and TRL to 0.20.0 (#293)
• Add configurable timeout for vLLM server initialization (#292)
• feat: Explore "Prime Intellect"-style clipping (#295)
• Launch H100-SXM by default (#289)
• Add AutoART notebook (#278)
• Improve release workflow with human-curated notes (#303)

Full Changelog: v0.4.4...v0.4.5

v0.4.4

ART 0.4.4 Release Notes

New Features

• SkyPilot Integration Enhancement: Added SkyPilot extras support for improved cloud deployment
  capabilities (#255)
• Reward System Improvements: Added experimental support to not scale rewards, providing more
  flexibility in reward configuration (#fd2a118)

Documentation & Examples

• New Tutorial: Added temporal-clue-7b.ipynb notebook demonstrating temporal reasoning capabilities
  (#a2802b3)
• Enhanced Documentation: Updated RULER documentation with comprehensive guidance on combining
  rewards (#250)
• ART•E Integration: Added ART•E notebook examples to documentation (#242, #240)

Bug Fixes & Improvements

• Dependency Management: Reverted to previous version of gql to resolve compatibility issues (#249)
• Unsloth Integration: Added experimental logprob pre-calculation support for Unsloth services
• Installation Fixes: Improved backend dependency installation when using local ART paths (#241)
• Documentation Updates: Various minor documentation improvements and clarifications

Technical Improvements

• Updated SkyPilot backend installation instructions
• Removed obsolete numpy installation cells from quickstart examples
• Enhanced dependency synchronization