EvoMap/AutoResearch

AI/ML research agents from idea to paper-ready evidence. An EvoMap open-source project.

What it solves

AutoResearch is designed to eliminate hallucinations and unsupported claims in AI/ML research by grounding the entire process—from idea generation to evidence collection—in real-world signals and rigorous cross-verification. It prevents the common issue of AI agents inventing details or overinterpreting negative results by enforcing a stateful, traceable workflow that requires independent reviews and empirical evidence.

How it works

The system operates as a stateful agent workflow divided into two primary phases:

  1. Idea Generation: It collects research signals from public channels (papers, community trends), intersects them with a local knowledge base of domain constraints, and uses at least three distinct LLMs to independently develop and cross-review candidate ideas.
  2. Idea Execution: Using the Claude Code CLI and a coordinator, it moves an idea through a recoverable pipeline: planning, pilot testing for feasibility, scaling to full experiments, result analysis, and finally, an independent blind review by a "critic" model to challenge conclusions.

All progress, including code, logs, and critic reports, is persisted to disk, allowing researchers to inspect or take over the process at any time.

Who it’s for

It is built for AI and machine learning researchers who want to automate the discovery of novel research directions and the execution of experiments while maintaining the scientific rigor required for paper-ready evidence.

Highlights

  • Cross-Domain Discovery: Combines external research signals with local domain knowledge to find novel problems.
  • Independent Multi-Model Review: Requires multiple distinct model identities to prevent self-approval of ideas.
  • Stateful & Recoverable: Persists all plans, code, and logs to disk so long-running experiments can resume after interruption.
  • Pilot-First Approach: Validates feasibility at a low cost before scaling to full compute resources.
  • Support for Negative Results: Preserves evidence of failed hypotheses rather than forcing a success story.

Related

  • Project
    weave-os/router

    A drop-in proxy that automatically routes AI requests to the best available model from multiple providers using a local embedder rather than manual selection.

  • Project
    rust-ndarray/ndarray

    A Rust crate providing n-dimensional containers for general and numeric elements, enabling efficient multi-dimensional array manipulation and linear algebra.

  • Project
    mljar/mercury

    A framework that turns Python notebooks into interactive web applications, allowing users to deploy chats, AI agents, and dashboards directly from .ipynb files.

  • Project
    crmne/ruby_llm

    A Ruby-native AI framework that provides a consistent API for building chats, agents, and multimodal features across multiple AI providers in Ruby and Rails applications.