Projects sharing features with ARPO

The GAIA game is over, and Alita is the final answer.

This repository contains the code and benchmark (ToolQA-D) for our research paper titled "Learning Evolving Tools for Large Language Models", which has been accepted at ICLR 2025.

Introduction

🌠 A PO st-training recipe for scaling R L on A dvanced R eason I ng model S 🚀

2025/10/23 🔥🔥Our paper is accepted by NeurIPS 2025.🔥🔥 - 2025/04/22 Released our Paper on arXiv. See here - 2025/03/24 We re-implement our algorithm based on verl. ✨✨ Key features: (1) add ~50 additional metrics to comprehensively monitor the training process and stability, (2) add a…

Zhi Gao, Yuntao Du, Xintong Zhang, Xiaojian Ma, Wenjuan Han, Song-Chun Zhu, Qing Li

This repository contains the code for our paper titled "Optimizing Test-Time Compute via Meta Reinforcement Finetuning." In this work, we introduce a novel approach to optimizing test-time compute through meta reinforcement learning, aiming to balance the efficiency and discovery capabilities of…

🔧✨Tool-Star: Empowering Multi-Tool Collaborative Web Agent via Reinforcement Learning

Princeton University \& PKU \& UIUC \& University of Chicago \& ByteDance Seed

December 5, 2025 🔍 We propose entropy ratio clipping​ (ERC) to impose a global constraint on the output distribution of the policy model. Experiments demonstrate that ERC can significantly improve the stability of off-policy training. 📄 The paper is available on arXiv.

25/08/20 : Aceept as EMNLP 2025 Main Conference paper

🚀 VLM-Powered Agent for Intelligent Tool Orchestration | 🔧 Open-source Framework for Multi-modal AI

Paper - Abstract - Updates - Quick Start - Installation - Download the datasets - Create Experiment Script - Single GPU Training (Only for Rho models) - Running the experiments - Code Structure - Initial SFT Checkpoints - Acknowledgement - Citation

JARVIS is a system for large language model task orchestration, deployment management, and automation benchmarking. It utilizes a task orchestrator to decompose complex requests into actionable steps and coordinates various expert models to synthesize final responses. The project includes an AI model deployment manager to handle the local deployment of expert models across different hardware scales. It further provides an AI workflow API consisting of web endpoints used to trigger automated task workflows and retrieve results from model selection stages. The framework incorporates an automat

``` STE/ ├─ toolmetadata/: tool related metadata ├─ prompts/: full prompts used ├─ savedresults/: prediction results in json ├─ {, FT, ICL}.json: results for baseline model, tool-enhanced w/ fine-tuning, tool-enhanced with ICL ├─ CLround.json: continual learning (each round) ├─ main.py: main…

💜 Confucius Demo | 🤗 Hugging Face | 🤖 ModelScope | ⌨️ GitHub | 📚 Paper | 💬 Wechat

This is the official implementation of paper Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning, where we present Nemotron-Research-Tool-N1, a family of tool-using reasoning language models. These models are trained with an R1-style reinforcement learning…

ICLR 2025 Tool-Planner: Task Planning with Clusters across Multiple Tools

This repository accompanies the paper Let's Verify Step by Step and presents the PRM800K dataset introduced there. PRM800K is a process supervision dataset containing 800,000 step-level correctness labels for model-generated solutions to problems from the MATH dataset. More information on…

ToolBench is an open platform for training, serving, and evaluating large language models that retrieve and call real-world APIs to complete user instructions. It provides an API-aware inference engine that selects relevant tools from a large corpus and generates sequences of tool calls to produce final answers, along with a custom API registration system that lets users add their own REST endpoints for the model to discover and invoke. The platform includes a complete instruction-tuning pipeline for training models on curated tool-use data, a multi-tool execution engine that coordinates sequ

Large Language Models (LLMs) have exhibited significant potential in performing diverse tasks, including the ability to call functions or use external tools to enhance their performance. While current research on function calling by LLMs primarily focuses on single-turn interactions, this paper…

Free Process Rewards without Process Labels

We release here the code for most of the main experiments in paper CREATOR: Tool Creation for Disentangling Abstract and Concrete Reasoning of Large Language Models, including the code of our framework and various baselines. We also release two new datasets: Creation Challenge and Tool Transfer…