InternAgent-1.5:一个用于长期自主科学发现的统一代理框架
Autonomous Discovery Across All Sciences
2026.5.07: We have open-sourced InternAgent-1.5 with advanced features, including enhanced algorithm discovery tasks, autonomous scientific paper reproduction, memory module, and deep research for tackling complex research challenges.
2026.3.17: We provide public access to InternAgent's Deep Research capabilities, enabling developers and researchers to seamlessly integrate its advanced deep research functionality into their own workflows.
2026.2.14: ❤️❤️ We open-source MLEvolve, the core implementation of InternAgent's solution optimization subsystem for algorithm design tasks. As the open-source method to achieve #1 on MLEBench, MLEvolve demonstrates powerful capabilities in solution optimization within bounded hypothesis spaces.
2026.2.10: Official release of the InternAgent 1.5 Technical Report. InternAgent 1.5 achieves leading performance on scientific reasoning benchmarks including GAIA, HLE, GPQA, and FrontierScience, and supports end-to-end autonomous scientific discovery tasks across Physical, Biology, Earth, and Life Science domains, enabling both algorithm discovery and empirical discovery (dry/wet-lab experiments).
2025.10.13: InternAgent-1.0 code has been fully open-sourced, supporting end-to-end automation and autonomous evolution across 12 scientific research tasks.
2025.07.17: The source code of InternAgent has been partially open-sourced. The complete version of InternAgent (covering 12 types of tasks for autonomous scientific research) will be open-sourced soon. This code repository can be used for full-cycle autonomous scientific research, ranging from hypothesis generation to automated experimental execution.
2025.07.10: NovelSeek has been renamed to InternAgent. This change embodies our hopeful vision for autonomous scientific research framework, and we hope it will empower all researchers to achieve great scientific discoveries.
conda create -n InternAgent python=3.11
conda activate InternAgent
pip install -r requirements.txt
Rename .env.example to .env and fill in your API keys:
mv .env.example .env
Key fields in .env:
OPENAI_API_KEY= # OpenAI or compatible API key (used for embeddings and memory)
OPENAI_API_BASE_URL= # Base URL for OpenAI-compatible endpoints
OPENROUTER_API_KEY= # OpenRouter API key (when using the openrouter provider)
ANTHROPIC_API_KEY= # Anthropic API key (for Claude-based experiment backends)
To use OpenRouter as the model gateway, set OPENROUTER_API_KEY and run with
config/openrouter_config.yaml. See docs/openrouter.md
for setup details.
AutoDebug is a self-contained toy task (no dataset or model downloads required) and is the recommended first run to verify your setup.
python launch_discovery.py \
--config ./config/default_config.yaml \
--task AutoDebug \
--exp_backend claudecode
QA mode uses InternAgent's deep research pipeline to answer a research question directly — no experiment loop, just a synthesized answer grounded in literature.
python launch_qa.py --question "What are recent advances in memory-augmented LLMs?"
# Optionally save the answer to a file
python launch_qa.py -q "What are recent advances in memory-augmented LLMs?" -o answer.md
launch.py is a unified entry point for both modes:
python launch.py --mode discovery --task AutoDebug --exp_backend claudecode
python launch.py --mode qa --question "What are recent advances in memory-augmented LLMs?"
Configuration tips:
config/ — default_config.yaml is the main starting point.results/, logs under logs/.--skip_idea_generation --idea_path <path/to/ideas.json>.scripts/ for ready-to-use example scripts.Algorithm discovery tasks live under tasks/. Each contains a prompt.json (task description), baseline code/, and a launcher.sh. Datasets and environment-specific setup vary per task — refer to the code in each task folder.
Scientific paper reproduction tasks (sci_tasks) are a distinct mode where InternAgent is given a published paper and its data, and asked to reproduce the key findings autonomously. These tasks live under sci_tasks/tasks/ (from the ResearchClawBench benchmark). See docs/sci_tasks.md for a full guide.
Ready-to-use example scripts are provided under scripts/:
run_discovery.sh
Run a full discovery experiment (idea generation → experiments)
run_skip-idea.sh
Run experiments from an existing idea file, skipping generation
run_sci.sh
Run a scientific paper reproduction task (defaults to Astronomy_000)
run_qa.sh
Answer a research question via deep research
InternAgent 1.5 includes a persistent memory module that records experiment outcomes across sessions, helping the agent avoid previously failed directions and build on successful ones. See docs/memory_module.md for configuration and setup.
The Deep Research (DR) module decomposes a research question into subtasks, gathers information from academic databases and the web in parallel, and synthesizes findings into a direct answer or structured report. See docs/deep_research.md for configuration and setup.
…
@article{team2025internagent,
title={InternAgent: When Agent Becomes the Scientist--Building Closed-Loop System from Hypothesis to Verification},
author={Team, InternAgent and Zhang, Bo and Feng, Shiyang and Yan, Xiangchao and Yuan, Jiakang and Ma, Runmin and Hu, Yusong and Yu, Zhiyin and He, Xiaohan and Huang, Songtao and others},
journal={arXiv e-prints},
pages={arXiv--2505},
year={2025}
}
@article{hu2025flowsearch,
title={FlowSearch: Advancing deep research with dynamic structured knowledge flow},
author={Yusong Hu and Runmin Ma and Yue Fan and Jinxin Shi and Zongsheng Cao and Yuhao Zhou and Jiakang Yuan and Xiangchao Yan and Wenlong Zhang and Lei Bai and Bo Zhang},
journal={arXiv preprint arXiv:2510.08521},
year={2025}
}
@article{du2025automlgen,
title={AutoMLGen: Navigating Fine-Grained Optimization for Coding Agents},
author={Shangheng Du and Xiangchao Yan and Dengyang Jiang and Jiakang Yuan and Yusong Hu and Xin Li and Liang He and Bo Zhang and Lei Bai},
journal={arXiv preprint arXiv:2510.08521},
year={2025}
}
暂无开放 Issues，或尚未同步最近议题。