# Alibaba-NLP/ZeroSearch

ZeroSearch: Incentivize the Search Capability of LLMs without Searching

Repository: https://github.com/Alibaba-NLP/ZeroSearch
Canonical: https://ross.abutalabs.com/products/zerosearch
Homepage: https://arxiv.org/pdf/2505.04588
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2025-08-16T17:34:58+00:00

## Health v2 (maintenance only)
Score: 36/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 37, release rhythm 35, longevity 34
- inputs: {"age_days": 483, "days_push": 382, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1307, forks 119 (observed 2026-08-28T04:04:19.091912+00:00)

## What it is
ZeroSearch is a reinforcement learning framework from Alibaba's Tongyi Lab that trains LLMs to use search by simulating search engine results with a fine-tuned LLM instead of calling real search APIs. It includes simulation LLMs, policy models, tuning datasets, and support for REINFORCE, GRPO, and PPO algorithms.

## Use cases
- train an LLM to use search without paying for search API calls
- reinforcement learning fine-tuning of LLMs for search agents
- simulate search engine results for LLM training
- train a search-capable agent with PPO or GRPO
- research on incentivizing search capabilities in language models

## When to choose
- you want to RL-train LLMs for search/retrieval without real search API costs
- you need reproducible simulated search environments for agent training
- you want pretrained simulation and policy models for Google or Wikipedia search

## When to avoid
- you need a production search engine or RAG pipeline rather than training
- you lack GPU resources for LLM training
- you need a plug-and-play search tool for inference only

## Facets
- artifact type: library
- maturity: active
- function: llm-training, rag, machine-learning, agent-framework
- domain: large-language-models, machine-learning
- platform: python
- tags: reinforcement-learning, search-simulation, llm-finetuning, research, verl, retrieval-augmented-generation, ai-agents, gpu, linux

## Member repositories
- Alibaba-NLP/ZeroSearch (main) score 36

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:19.091912+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 2026-08-30T04:50:07.074684+00:00, confidence not recorded.
  - readme: https://github.com/Alibaba-NLP/ZeroSearch (fetched 2026-08-28T04:04:19.091912+00:00, sha 81eb52eacb49)
- Data as of 2026-08-30T08:39:29.467469+00:00.