When LLM Agents Meet Reinforcement Learning
Section: Base Framework · Prime Intellect
FreshScore 89
Entry
prime-rl
Appears in 5 awesome lists
Prime Intellect's fully asynchronous RL framework — "agentic RL training at scale" (the INTELLECT-3 report) — built to train 1T+ parameter MoE models on 1,000+ GPUs through three abstractions: a lightweight CPU Orchestrator (dataflow/scheduling), an FSDP2 Trainer, and vLLM inference as…
Open github.comprimeintellect-ai/prime-rl
Found in these lists
Awesome Agent Evals
Section: 5e · RL-environment / verifiable-reward toolkits (eval ⇄ training) · Prime Intellect — https://github.com/PrimeIntellect-ai/prime-rl — 🆕 async RL trainer consuming verifiers envs (INTELLECT-3).
FreshScore 88
Awesome LLM Resources
Section: Agentic RL
FreshScore 90
Awesome Open Source AI
Section: 7. Training & Fine-tuning Ecosystem · Agentic RL Training at Scale from Prime Intellect. Framework for large-scale reinforcement learning capable of scaling to 1000+ GPUs with fully asynchronous RL, FSDP2 training, and vLLM inference. Apache 2.0 licensed.
FreshScore 88
Awesome Prompts
Section: Agent Training (RL Post-Training) · Prime Intellect's fully asynchronous RL framework — "agentic RL training at scale" (the INTELLECT-3 report) — built to train 1T+ parameter MoE models on 1,000+ GPUs through three abstractions: a lightweight CPU Orchestrator (dataflow/scheduling), an FSDP2 Trainer, and vLLM inference as…
FreshScore 96
Often listed next to it
TRL
Hugging Face's post-training library — the reference implementation of GRPO (the algorithm DeepSeek popularized) and the most-used single-GPU-to-cluster starting point for RL post-training; 75+ trainers across SFT / DPO / KTO / PPO / ORPO / CPO and more, with vLLM generation, PEFT/LoRA, and AMD…
In 10 listsDetails
slime
Zhipu AI / Tsinghua's (THUDM) RL scaling framework — the production trainer behind the GLM series from GLM-4.5 to GLM-5.3-Flash; connects Megatron training with SGLang rollout through a unified train–rollout–buffer architecture (parameters sync, router-based generation, prompt/verifier management)…
In 6 listsDetails
OpenRLHF
The Ray+vLLM pioneer of open RLHF/RL post-training — the first framework built around a fully decoupled Ray-based distributed architecture (vLLM generation + DeepSpeed ZeRO/FSDP training, placement-colocated hybrid engine with sleep-mode weight sharing between actor/critic/reward/reference); PPO /…
In 6 listsDetails
NeMo-RL
NVIDIA's scalable RL post-training library (successor to NeMo Aligner, re-architected with the NeMo Framework) — the megatron-core-backed option in the open stack: GRPO / GSPO / DAPO / PPO / DPO / SFT(+LoRA) / on-policy distillation / GDPO with training via PyTorch DTensor (FSDP2) or Megatron Core…
In 5 listsDetails
verl
The most widely adopted open-source RL post-training framework for LLMs (ByteDance Seed; HybridFlow architecture, EuroSys 2025) — a flexible single-controller/dataflow abstraction that hybrid-orchestrates rollout generation (vLLM/SGLang) and training (FSDP/Megatron) in one program; PPO / GRPO /…
In 5 listsDetails
agent-lightning
Microsoft Research's harnessed agentic RL trainer — v1.0 (~3,500 LOC, arXiv 2608.17528) trains agents inside their real production harnesses instead of reimplementing them within the trainer: an OpenAI-compatible API Gateway proxy records prompts/responses/logprobs of any agent (LangGraph, OpenAI…
In 4 listsDetails
RLinf
Scalable open-source RL infrastructure for post-training foundation models via reinforcement learning. Features M2Flow paradigm for embodied AI and agentic workflows with real-world robotics integrations. Apache 2.0 licensed.
In 3 lists
AReaL
Fully asynchronous RL system for LLM reasoning and agents (inclusionAI / Ant Group + Tsinghua IIIS; NeurIPS 2025) — decouples rollout generation from training with staleness-aware PPO, reporting 2.77× speedup over synchronous systems at equal GPU count and near-linear scaling to 512 GPUs; GRPO /…
In 3 lists