# RUC-NLPIR/ARPO

[ICLR 2026] Agentic Reinforced Policy Optimization (ARPO)

Repository: https://github.com/RUC-NLPIR/ARPO
Canonical: https://ross.abutalabs.com/products/arpo
Language: Python
License Family: other
Last push: 2026-08-20T07:12:37+00:00

## Health v2 (maintenance only)
Score: 54/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 98, release rhythm 13, longevity 29
- inputs: {"age_days": 407, "days_push": 13, "days_rel": 367, "gap_med": null, "n_releases_24m": 1}
- flags: no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1109, forks 60 (observed 2026-08-28T04:03:37.290946+00:00)

## What it is
ARPO (Agentic Reinforced Policy Optimization) is a reinforcement learning algorithm and training framework for LLM agents, published at ICLR 2026. It provides training code, models, and datasets for improving multi-turn agentic reasoning and tool use via RL.

## Use cases
- train llm agents with reinforcement learning
- improve multi-turn tool use of a language model
- reproduce ARPO paper results
- fine-tune Qwen or Llama models for agentic search
- apply RL post-training to deep search agents

## When to choose
- you need a research-grade agentic RL training algorithm for LLMs
- you want to RL-train models for multi-turn tool calling or deep search
- you want to build on ICLR 2026 ARPO/AEPO methods

## When to avoid
- you need a production-supported training framework with commercial support
- you need simple SFT-only fine-tuning without RL
- you lack GPU infrastructure for large-model RL training

## Facets
- artifact type: library
- maturity: active
- function: llm-training, reinforcement-learning, agent-framework
- domain: large-language-models, reinforcement-learning, machine-learning
- platform: python
- tags: rlhf, policy-optimization, agentic-rl, iclr-2026, research-code, llm-post-training, ai-agents, gpu, linux

## Member repositories
- RUC-NLPIR/ARPO (main) score 54

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:37.290946+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 2026-08-30T06:43:33.058581+00:00, confidence not recorded.
  - readme: https://github.com/RUC-NLPIR/ARPO (fetched 2026-08-28T04:03:37.290946+00:00, sha 08ad02062300)
- Data as of 2026-08-30T08:39:29.467469+00:00.