Awesome Auto Research Hub Papers · Datasets · Projects
← Back to papers

MLR-Copilot: Autonomous Machine Learning Research based on Large Language Models Agents

arXiv 2024 59.6 method

TLDR

MLR-Copilot is an autonomous ML research framework using LLM agents for idea generation, experiment implementation, and code execution.

Reasoning

The paper presents a clear three-stage pipeline with RL-tuned LLM and HuggingFace integration, showing practical design. However, the abstract lacks detailed evaluation results and does not specify the nature of the five tasks, limiting assessment of empirical rigor.

Read-first score

Read-first score 59.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 83.

Recency 8%
75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Topical relevance 42%
69.2

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Methodology quality 25%
60

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=experiment,result

Reproducibility 25%
38

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Field roles

Bridge

Rank sensitivity

Stability: volatile; rank range: 117.

Keyword Scores

autonomous research agent
9
automated research
9
automated experimentation
9
research automation
9
automated scientific discovery
8
experiment design agent
8
AI for scientific research
8
scientific discovery agent
8
AI scientist
7
literature review agent
5
survey generation
2
paper writing agent
1

Deep Analysis

Innovations

  • Autonomous ML research framework with three stages: idea generation, experiment implementation, and code execution
  • IdeaAgent powered by an RL-tuned LLM to generate feasible ideas and experiment plans from existing research papers
  • ExperimentAgent that retrieves prototype code and optionally retrieves candidate models and data from HuggingFace to convert plans into executable code
  • Iterative debugging and human feedback loop to improve success of executable outcomes
  • Enhancement of ideation through training after initial generation

Methodology

The framework uses three stages: IdeaAgent (RL-tuned LLM) generates ideas and experiment plans from research papers; ExperimentAgent converts plans to code using retrieved prototype code and optionally HuggingFace models/data; then runs experiments with iterative debugging and human feedback. Evaluated on five ML research tasks.

Key Results

Experiment results demonstrate the potential of the framework to facilitate ML research progress and innovation.

Tags

AICLLG