AI Research Env: An End-to-End Training Platform for Machine Learning Research Agents

AI Research Env is an OpenEnv-compatible simulation platform designed to train AI agents to complete a full scientific research workflow from literature reading and hypothesis formulation to result analysis, providing a standardized evaluation environment for autonomous scientific discovery agents. Through structured workflows, multi-difficulty tasks, and multi-dimensional evaluation mechanisms, the platform promotes the transformation of AI from simple question-answering to an autonomous scientific research paradigm.

Current large language models (LLMs) are mostly simple question-answering systems, while real scientific research requires completing complex processes such as literature reading, hypothesis formation, experiment design, and result analysis. The goal of AI Research Env is to bridge this gap and enable agents to become autonomous systems capable of handling the full research process.

The platform defines seven core actions to simulate the research process: read_paper (literature summary), propose_hypothesis (hypothesis formulation), design_experiment (experiment design), run_experiment (experiment execution), analyze_results (result analysis), refine_hypothesis (hypothesis iteration), and final_answer (conclusion and recommendation). It also provides three tasks with increasing difficulty: computer vision classification (easy), natural language processing sentiment analysis (medium), and healthcare tabular data (hard), covering real challenges in different machine learning domains.

The platform uses a phased scoring mechanism, including keyword coverage (50-65%), in-depth analysis (25-35%), and phase progress rewards (5%). The scoring range for each step is 0.0-1.0 (shaping reward), and the round reward is the sum of steps. Context prompts are unlocked after the second step to help agents adjust their direction, avoiding training difficulties due to sparse rewards.

The backend is built on FastAPI to provide RESTful APIs, including interfaces for health checks, round resets, and action submissions. The frontend is a React+Recharts dashboard that supports real-time progress visualization, action history tracking, and reward curve analysis. The core environment uses Pydantic typed models to ensure data consistency, with 27 test cases covering key functional paths.

Test results using Qwen/Qwen2.5-72B-Instruct show: computer vision classification task score of approximately 0.74 (6 steps), NLP sentiment analysis of approximately 0.68 (7 steps), healthcare tabular data of approximately 0.61 (8 steps), with an average score of about 0.68. These results indicate that advanced LLMs still have room for improvement in end-to-end research tasks, while validating the effectiveness of the platform's evaluation mechanism.

The innovative value of AI Research Env lies in providing a standardized evaluation benchmark to promote AI-assisted scientific discovery. Future outlooks include: adding more domain tasks, building stronger baseline models, exploring new training methods and agent architectures, and expanding applications in real scientific research scenarios. This is a solid step toward the vision of AI-assisted scientific discovery.