# langfengQ/verl-agent

verl-agent is an extension of veRL, designed for training LLM/VLM agents via RL. verl-agent is also the official code for paper "Group-in-Group Policy Optimization for LLM Agent Training"

Repository: https://github.com/langfengQ/verl-agent
Canonical: https://ross.abutalabs.com/products/verl-agent
Homepage: https://huggingface.co/papers/2505.10978
Language: Python
License: Apache-2.0
License Family: permissive
Topics: llm-agents, llm-training, reinforcement-learning, large-language-models, deepseek-r1, grpo, agent-framework, gigpo
Last push: 2026-06-09T13:35:49+00:00

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 86, release rhythm 28, longevity 37
- inputs: {"age_days": 528, "days_push": 85, "days_rel": 265, "gap_med": null, "n_releases_24m": 1}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2256, forks 216 (observed 2026-08-28T04:06:31.402415+00:00)

## What it is
verl-agent is an extension of the veRL framework for training LLM and VLM agents via reinforcement learning, featuring step-independent multi-turn rollouts and customizable history/memory management. It is also the official implementation of the GiGPO (Group-in-Group Policy Optimization) algorithm for fine-grained credit assignment in long-horizon agent training.

## Use cases
- train llm agents with reinforcement learning
- implement gigpo for long-horizon agent training
- run multi-turn RL rollouts on ALFWorld or WebShop
- fine-tune Qwen models as interactive agents
- compare group-based RL algorithms like GRPO and GiGPO
- train vision-language agents in visual environments

## When to choose
- you need RL post-training of LLM agents over multi-turn, long-horizon tasks
- you want critic-free group-based RL with fine-grained step-level credit assignment
- you need customizable per-step input structures and memory modules during rollouts
- you want to reproduce or extend GiGPO, HGPO, or GraphGPO research

## When to avoid
- you only need single-turn RLHF or SFT without agent-environment interaction
- you need a production inference or serving framework rather than a training framework
- you lack GPU resources for large-scale RL rollouts

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, reinforcement-learning, agent-framework, machine-learning
- domain: reinforcement-learning, large-language-models, machine-learning, deep-learning
- platform: python
- tags: rl-training, grpo, gigpo, multi-turn-agents, verl, llm-agents, research-code, ai-agents, gpu, linux

## Member repositories
- langfengQ/verl-agent (main) score 56

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:31.402415+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:43:42.885487+00:00, confidence not recorded.
  - readme: https://github.com/langfengQ/verl-agent (fetched 2026-08-28T04:06:31.402415+00:00, sha 37bad6a6e72d)
  - homepage: https://huggingface.co/papers/2505.10978 (fetched 2026-08-29T10:23:42.713766+00:00, sha fb194e7508b9)
  - site_page: https://huggingface.co/docs (fetched 2026-08-29T10:23:42.722887+00:00, sha bdec26667b98)
  - site_page: https://huggingface.co/pricing (fetched 2026-08-29T10:23:42.724645+00:00, sha de6b7a178be5)
  - site_page: https://huggingface.co/huggingface (fetched 2026-08-29T10:23:42.726958+00:00, sha 0a25e6fef275)
- Data as of 2026-08-30T08:39:29.467469+00:00.