# Agent-RL/ReCall

ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning & ReCall: Learning to Reason with Tool Call for LLMs via Reinforcement Learning

Repository: https://github.com/Agent-RL/ReCall
Canonical: https://ross.abutalabs.com/products/recall
Homepage: https://attractive-almandine-935.notion.site/ReCall-Learning-to-Reason-with-Tool-Call-for-LLMs-via-Reinforcement-Learning-1d7aec91e9bb8006ad40f9edbfe2191a
Language: Python
License: MIT
License Family: permissive
Topics: agent, function-calling, llm, reinforcement-learning, tool-use
Last push: 2025-05-16T10:31:58+00:00

## Health v2 (maintenance only)
Score: 30/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 21, release rhythm 35, longevity 39
- inputs: {"age_days": 548, "days_push": 474, "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 1431, forks 89 (observed 2026-08-28T04:04:42.838493+00:00)

## What it is
ReCall is a framework that trains LLMs to reason with arbitrary tool calls via reinforcement learning, without supervised tool-use trajectory data. It is the successor to ReSearch, built on verl and vLLM, and includes synthetic multi-step task generation for training agentic tool-based reasoning.

## Use cases
- train an llm to call tools via reinforcement learning
- teach a model to reason with search and function calling without sft data
- build agentic tool-use capabilities in open-source llms
- generate synthetic multi-step tool-use training environments
- reproduce the ReSearch reasoning-with-search training pipeline

## When to choose
- you want to RL-train an LLM for agentic tool calling without supervised trajectories
- you need a drop-in replacement for ReSearch with arbitrary user-defined tools
- you want to experiment with synthetic environments for multi-step tool-use training

## When to avoid
- you only need inference-time tool calling without training
- you lack multi-GPU infrastructure for RL training with vLLM rollouts
- you need a production-ready agent framework rather than a research training codebase

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, reinforcement-learning, agent-framework, rag
- domain: large-language-models, reinforcement-learning, machine-learning
- platform: python
- tags: tool-calling, function-calling, verl, vllm, synthetic-data, research, tool-use, ai-agents, gpu, linux

## Member repositories
- Agent-RL/ReCall (main) score 30

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:42.838493+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 2026-08-30T04:37:04.484522+00:00, confidence not recorded.
  - readme: https://github.com/Agent-RL/ReCall (fetched 2026-08-28T04:04:42.838493+00:00, sha d929109e8bde)
  - homepage: https://attractive-almandine-935.notion.site/ReCall-Learning-to-Reason-with-Tool-Call-for-LLMs-via-Reinforcement-Learning-1d7aec91e9bb8006ad40f9edbfe2191a (fetched 2026-08-29T11:48:23.755252+00:00, sha 73a6ba54b760)
- Data as of 2026-08-30T08:39:29.467469+00:00.