sjtu-sai-agents/ML-Master

The official implementation of "ML-Master: Towards AI-for-AI via Integration of Exploration and Reasoning"

What it solves

ML-Master tackles the challenge of automating complex machine learning engineering work, like solving Kaggle-style competition tasks on OpenAI's MLE-Bench benchmark. It helps an AI system act independently: it explores different approaches, reasons about them, and gradually accumulates knowledge, so it can handle long, multi-step problems without constant human guidance.

How it works

ML-Master combines exploration and reasoning in a single loop. It tries out different strategies, captures the insights and outcomes, and stores the most useful ones in an adaptive memory. Version 2.0 adds a Hierarchical Cognitive Caching (HCC) architecture: it distills short-term execution traces into stable long-term knowledge, keeping tactical execution separate from strategic planning while both evolve together. The system uses LLMs (e.g., DeepSeek for coding, GPT-4o for evaluation and feedback) to drive the process, and it runs on top of the MLE-Bench environment.

Who it's for

Researchers and practitioners working on AI agents, AutoML, or long-horizon task automation. It's especially valuable for those who want an open-source framework for autonomous machine learning engineering, where the agent must plan, execute, and learn over extended periods.

Highlights

  • Ranked #1 on OpenAI's MLE-Bench leaderboard with 56.44% overall performance, a 92.7% improvement over version 1.0.
  • Outperforms prior baselines across low (75.76%), medium (50.88%), and high (42.22%) complexity tasks, with medium complexity improving 152.2%.
  • Introduces a Hierarchical Cognitive Caching (HCC) architecture for ultra-long-horizon scientific exploration.
  • Open-source project with Docker image support and integration with MLE-Bench; later expanded into the self-evolving EvoMaster framework.

Related

  • Project
    securetron-gh/aifw

    AI-FW is a self-hosted gateway that sits between AI agents and the models they call, inspecting every prompt and response, enforcing governance rules, and producing regulator-ready audit logs. It supports fail-closed operation, agent identity, cost-saving caches, and integration with tools like Claude Code, Cursor, and M365 Copilot.

  • Project
    menkesu/awesome-pm-skills

    Awesome PM Skills 2.0 is a collection of 25 product-management workflows distilled from 340 episodes of Lenny's Podcast. Each skill takes a real work artifact like a pricing page or PRD and guides an AI assistant to produce a concrete deliverable such as a pricing plan or decision memo, with expert frameworks, grading, and verified quotes.

  • Project
    THUSI-Lab/GameScaling

    GameScaling is a set of vision-language models that were continued pre-training on 1,600 hours of keyboard-and-mouse gameplay so they can play computer games in real time. Each model takes the current screen (plus a short history) and outputs the next 200 ms of input as six 33 ms key steps plus one mouse move and scroll; the repo provides the inference client, system prompt template, examples, and tests for serving the released checkpoints.

  • Project
    adithyan-ak/AgentHound

    AgentHound is an open‑source offensive security framework that scans AI‑agent infrastructure from a foothold, harvests credentials, validates real access, and builds a graph‑based attack surface. It works via a static collector binary (active or stealth mode) and an optional Neo4j‑backed server for visual analysis.