Open-source alternatives to Stable Alignment

(Aug 29, 2025) It is based on the TRL (Transformer Reinforcement Learning) package. We are also planning to use Verl to make this easy to use, especially for the inference time scaling, so stay tuned!

[Website](http://craftjarvis-jarvis1.github.io/) [Paper](https://arxiv.org/abs/2311.05997) [Twitter](https://twitter.com/jeasinema/status/1723900032653643796)

This repo contains the source code for making plans based on problems decribed by natural language.

This repository contains code released by Curai Health. For a list of all publications at Curai Health, see our blog.

MetaGPT is an agentic workflow orchestrator and multi-agent framework designed to transform natural language requirements into complete software deliverables. It functions as an AI software engineering suite that automates the creation of technical documentation, data structures, and source code by treating natural language as a programming environment. The system distinguishes itself by assigning professional roles to large language models, creating specialized agent teams that collaborate through a shared communication structure. It utilizes standard operating procedures to convert organiza

Official implementation of our ScoreFlow. Affiliation: University of Chicago, Princeton University, University of Oxford.

Original Flan (2021) | The Flan Collection (2022) | Flan 2021 Citation | License

Teaching Language Models to Critique via Reinforcement Learning

CORAL is a robust, lightweight infrastructure for multi-agent autonomous self-evolution, built for autoresearch. Works with Claude Code, Codex, Cursor, OpenCode, Kiro, and more.

This project introduces a novel agent architecture that simulates the attitudes and behaviors of real individuals by applying large language models (LLMs) to qualitative interviews about their lives. These agents replicate participants' responses on various social science measures, providing a…

🐝 The First Self-Improving agents with RL / Prompting Optimization

This framework provides a development environment for building collaborative systems where autonomous agents interact to solve complex tasks through conversational workflows. It functions as a conversational workflow engine and event-driven runtime, coordinating multi-step processes by translating high-level goals into structured dialogue sequences between specialized agents. The system distinguishes itself through its message-passing orchestration, which manages state transitions and task delegation between independent participants. It supports dynamic conversation state management to provid

RD-Agent is an autonomous framework designed to orchestrate multi-step software engineering and data science workflows. By leveraging large language models, the system decomposes complex technical requirements into actionable research, planning, and execution phases, ultimately generating and running code to solve specific development tasks. The platform distinguishes itself through a containerized execution sandbox that ensures secure dependency management and system stability for all autonomously generated code. It employs multi-agent orchestration to manage iterative feedback loops, allowi

​TextWorld is a sandbox learning environment for the training and evaluation of reinforcement learning (RL) agents on text-based games.

TinyTroupe is a multi-agent simulation framework designed to create populations of persona-based agents that interact to generate synthetic behavioral data and business insights. It serves as a persona-based agent orchestrator and synthetic data generator, allowing for the definition of agents with specific personality traits and goals to coordinate their interactions through structured workflows. The project features an extensible plugin system for connecting simulated agents to external tools and servers to execute code and access remote data. It includes an agentic simulation dashboard tha

Voyager is an autonomous embodied agent and lifelong learning framework that uses a large language model to explore virtual environments. It functions as a code-based action controller, translating natural language instructions into executable scripts to interact with its surroundings. The system features an automatic curriculum generator that creates sequences of exploration goals to discover new items and behaviors without human intervention. It maintains a skill library manager that stores learned behaviors as reusable code fragments, which can be composed to execute complex tasks. The fr

Swarm is a framework for building conversational systems that coordinate multi-agent workflows. It functions as an orchestration engine that manages persistent, multi-turn dialogues by routing tasks between specialized agents and executing local functions. The system is designed to handle complex, multi-step processes by maintaining shared state and context across agent interactions. The framework distinguishes itself through its approach to dynamic task delegation and execution control. It enables agents to hand off tasks to one another by returning agent objects, allowing for modular, domai

Agentverse is a multi-agent framework and orchestrator designed for deploying and managing multiple large language model agents. It provides a simulation environment where agents interact based on custom personas and defined interaction rules to solve tasks or simulate social dynamics. The system features a tool integration layer that connects agents to external functional plugins and specialized tools, extending their capabilities beyond text generation. It uses a combination of persona-based prompt injection and state-managed memory to maintain agent consistency and specialized skill sets d