Skip to content

kagnlp/xolver.github.io

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 

Repository files navigation

Xolver

This website showcases our work on Xolver: Multi-Agent Reasoning with Holistic Experience Learning Just Like an Olympiad Team.

Project Overview

Xolver is a training-free, multi-agent reasoning framework that equips black-box LLMs with persistent, evolving memory of holistic experience. Inspired by how expert problem solvers like Olympiad teams leverage diverse experiences, Xolver integrates external retrieval, tool use, agent collaboration, self-evaluation, and iterative refinement to achieve expert-level reasoning across mathematics and programming tasks.

Key Features

  • Multi-Agent Collaboration: Dynamic team of specialized agents (mathematicians, programmers, verifiers) that work together
  • Holistic Experience Learning: Dual-memory architecture combining episodic long-term memory with evolving shared memory
  • Tool-Augmented Reasoning: Seamless integration with external tools (Python execution, code debugging)
  • Iterative Refinement: Judge-mediated feedback and continuous improvement across iterations
  • Cross-Problem Learning: Accumulates knowledge from solved problems to enhance future performance
  • State-of-the-Art Results: 98.1% on GSM8K, 94.4% on AIME'24, 93.7% on AIME'25, 99.8% on Math-500, 91.6% on LiveCodeBench

Performance Highlights

Even with lightweight backbones (QWQ-32B), Xolver often surpasses the most advanced models including:

  • Qwen3-235B, Gemini 2.5 Pro, o1, o3, and o4-mini-high
  • Specialized reasoning agents like OctoTools, CheatSheet, Search-o1
  • With stronger backbones (o3-mini-high), achieves new best results across all benchmarks

Resources

Website License

Creative Commons License
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •