MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering

MLExpert is a 35B Mixture-of-Experts agentic model for autonomous machine learning engineering. It organizes experiments, writes and revises code, and improves solutions through sustained interaction with execution environments. Hierarchical agentic supervised fine-tuning (SFT) develops general tool-use and domain-specific optimization skills, while Windowed Interaction Reinforcement Learning (WIRL) trains multi-turn decisions using feedback from completed experiments.

• Long-horizon optimization: Plans experiments, interprets execution feedback, revisits previous solutions, and selects submission candidates within a persistent workspace.
• MLE performance: Achieves a 72.7% medal rate on MLE-Bench Lite, matching GLM-5.3 and Kimi-K3 under the MLExpert Harness.
• General-purpose harness compatibility: Achieves a 69.7% medal rate with Claude Code, compared with 42.4% for its base model.
• Scientific task optimization: Matches published scientific results on seven of ten NatureBench Lite tasks and surpasses them on five.

The model selects subsequent actions based on experimental states and execution feedback. The interaction harness maintains candidate solutions and experiment history throughout long-horizon optimization.

The Kaggle-MLE Factory constructs executable machine learning tasks. The interaction harness supports model-directed experimentation and teacher trajectory collection. Hierarchical agentic SFT transfers general coding and tool-use capabilities to MLE through outcome-guided supervision. WIRL groups decisions between successive experiment results into execution windows and learns from intermediate execution feedback through asynchronous collection and policy correction.

MLE-Bench Lite comprises 22 historical Kaggle competitions for end-to-end machine learning engineering. The main metric is the percentage of tasks receiving any medal. Under the MLExpert Harness, MLExpert raises its base model's medal rate from 37.9% to 72.7%. With Claude Code, the rate improves from 42.4% to 69.7%.

NatureBench Lite comprises ten scientific experimentation tasks. M denotes Match-SOTA, and S denotes Surpass-SOTA. With Claude Code, MLExpert raises the Surpass-SOTA rate from 0.0% to 50.0% and the Match-SOTA rate from 40.0% to 70.0% relative to its base model.

The Kaggle-MLE Factory builds executable MLE tasks from user-supplied Kaggle competition slugs, including downloading, data analysis, LLM-generated preparation and scoring scripts, validation, and repair. Kaggle credentials and the model service are user-configurable. Generated data splits and scoring code require human review, especially for complex competition metrics. Start with its offline example, which runs without Kaggle credentials or a model service.

MLExpert model weights are available on Hugging Face. The commands below serve an OpenAI-compatible API at http://localhost:8000/v1 with tensor parallelism across two GPUs.

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate
uv pip install vllm --torch-backend=auto

Text generation

vllm serve InternScience/MLExpert \
    --host 127.0.0.1 \
    --port 8000 \
    --tensor-parallel-size 2 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --language-model-only

Tool calling

vllm serve InternScience/MLExpert \
    --host 127.0.0.1 \
    --port 8000 \
    --tensor-parallel-size 2 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --language-model-only

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate
uv pip install sglang

Text generation

python -m sglang.launch_server \
    --model-path InternScience/MLExpert \
    --host 127.0.0.1 \
    --port 8000 \
    --tp-size 2 \
    --context-length 262144 \
    --reasoning-parser qwen3 \
    --language-model-only

Tool use

python -m sglang.launch_server \
    --model-path InternScience/MLExpert \
    --host 127.0.0.1 \
    --port 8000 \
    --tp-size 2 \
    --context-length 262144 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder \
    --language-model-only

These parameters apply to both thinking and non-thinking modes. Set chat_template_kwargs.enable_thinking to true to enable thinking or false to disable it. Machine-readable profiles are provided in the model repository's inference_profiles.json.