AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery

Abstract: Scientific research is shifting from isolated AI assistance to integrated, longer-horizon automated workflows (spanning literature, planning, execution, validation, and reporting). This repository synthesizes the transition toward AutoResearch, detailing a 5-level autonomy spectrum from L0 (Human Only) to L4 (Fully Autonomous). We focus on the currently active Vibe Research zone (L1-L3), outlining its technical foundations and evaluation paradigms.

Figure 1: Level-wise Progression of AutoResearch Across Five Workflow Stages.

The paradigm of AI for Science is moving beyond narrow, task-specific models (e.g., AlphaFold) toward workflow-level research automation. AI agents are now connecting idea generation, coding, experimentation, and manuscript drafting into integrated pipelines.

To systematically understand this landscape, we propose a Five-Level Autonomy Spectrum defining the allocation of control, execution, and validation.

Figure 2: How workflow control, task execution, and validation authority shift across L0-L4.

Rather than treating AutoResearch as a single benchmark, we view it as a layered evolution of capabilities, strongly conditioned by scientific domains.

Figure 3: Overview of the Survey Framework: Concept, Technical Foundations, Evaluation, and Domains.

Figure 4: Historical trajectory of AutoResearch mapping representative systems onto the L0-L4 spectrum.

Scientific autonomy relies on how robustly five core workflow stages are coupled. Strong automation requires that evidence, plans, execution, validation, and reporting maintain strict scientific alignment throughout the loop.

Figure 5: The five recurring technical stages of AutoResearch workflows.

Transforms raw literature access into durable, reusable evidential states to anchor downstream reasoning.

• Key Regimes: Search-centered, Evidence-centered, Structure-centered, Literature-memory grounding.

Converts grounded context into operationalizable, testable, and comparable candidate directions.

• Key Regimes: Proposal-centered, Deliberative multi-agent, Structure-guided, Search-based evolutionary planning.

Binds candidate plans to actionable substrates (code, APIs, lab instruments) for concrete execution.

• Key Regimes: Code-native, Tool-orchestrated, Laboratory-robotic, Human-gated execution.

Introduces rejection pressure to filter weak results and errors before they harden into scientific claims.

• Key Regimes: Execution-coupled reruns, Critique-mediated review, Expert-grounded evaluation.

Translates workflow state into inspectable artifacts while preserving exact claim-evidence alignment.

• Key Regimes: Draft-centered, Review-centered, Artifact-linked reporting.

Workflow automation is advancing faster than workflow verifiability. Evaluation must shift from measuring isolated task accuracy (e.g., benchmark scores) to assessing the holistic Scientific Quality of the entire pipeline.

Figure 11: Five dimensions of scientific quality essential for evaluating AutoResearch workflows.

1. Novelty: Originality of ideas, hypotheses, and expansion of the search space.
2. Validity: Correctness of methods, execution accuracy, and evidence-claim alignment.
3. Impact: Plausible scientific usefulness and downstream adoption potential.
4. Reliability: Rerun stability, robustness against perturbations, and failure recovery.
5. Provenance: Traceability of claims back to exact data, tools, and literature sources.

• In L1/L2 (Human-Verified): Evaluation asks whether AI support improves throughput without degrading validity. Outcome quality alone is insufficient; we must also measure oversight load and human intervention points.
• In L3/L4 (AI-Led): The evaluative burden increases dramatically. High-autonomy systems must self-supply the rigorous rejection pressure, provenance, and accountability that human reviewers traditionally provide.

Scientific autonomy does not progress evenly across all fields. It is heavily constrained by the underlying domain's structure—specifically, the manipulability of the research object, feedback speed, observability, and the burden of accountability.

Figure 12: Domain-conditioned autonomy ceilings, showing the gradient from computationally closed domains to highly constrained empirical and social domains.

• 💻 Computational & Formal Sciences (High Autonomy Potential): Code-native, replayable environments with explicit correctness signals. These domains currently support the strongest workflow closure (approaching L3).
• 🔬 Physical, Chemical & Biological Sciences (Intermediate Autonomy): Rely on simulators, robotic platforms, and closed-loop optimization. Automation is powerful but bounded by physical instrumentation and hardware interfaces.
• 🏥 Clinical, Social & Earth Sciences (Highly Constrained): Characterized by delayed validation, human embodiment, and non-manipulable systems. Autonomy is strictly limited by regulatory, ethical, and causal constraints (remaining firmly in L2).

This list complements the survey narrative above. It is organized around the paper's workflow-centered framing rather than a single model family or benchmark. Level labels are intentionally conservative: a broad pipeline is listed as human-verified workflow automation when routine human verification is still structurally necessary.

• Automated scientific discovery: from equation discovery to autonomous discovery systems. Paper
• Hypothesis generation with large language models. Paper
• From automation to autonomy: A survey on large language models in scientific discovery. Paper
• Agent4S: The Transformation of Research Paradigms from the Perspective of Large Language Models. Paper
• Agentic AI for scientific discovery: A survey of progress, challenges, and future directions. Paper
• From AI for Science to Agentic Science: A survey on autonomous scientific discovery. Paper
• A Survey of AI Scientists. Paper
• AI4Research: A survey of artificial intelligence for scientific research. Paper
• A vision for auto research with LLM agents. Paper
• The evolving role of large language models in scientific innovation: Evaluator, collaborator, and scientist. Paper

• Autonomous LLM-driven research: from data to human-verifiable research papers. Paper
• The AI Scientist: Towards fully automated open-ended scientific discovery. Paper
• The AI Scientist-v2: Workshop-level automated scientific discovery via agentic tree search. Paper
• Curie: Toward rigorous and automated scientific experimentation with AI agents. Paper
• PiFlow: Principle-aware scientific discovery with multi-agent collaboration. Paper
• AI-Researcher: Autonomous scientific innovation. Paper
• CodeScientist: End-to-end semi-automated scientific discovery with code-based experimentation. Paper
• DeepScientist: Advancing frontier-pushing scientific findings progressively. Paper
• AlphaEvolve: A coding agent for scientific and algorithmic discovery. Paper
• EvoScientist: Towards multi-agent evolving AI scientists for end-to-end scientific discovery. Paper

• Many heads are better than one: Improved scientific idea generation by a LLM-based multi-agent system. Paper
• CycleResearcher: Improving automated research via automated review. Paper
• Robin: A multi-agent system for automating scientific discovery. Paper
• Towards an AI co-scientist. Paper
• SciSciGPT: advancing human-AI collaboration in the science of science. Paper
• Build your personalized research group: A multiagent framework for continual and interactive science automation. Paper
• OmniScientist: Toward a Co-evolving Ecosystem of Human and AI Scientists. Paper
• AgentRxiv: Towards collaborative autonomous research. Paper
• Rethinking the AI Scientist: Interactive Multi-Agent Workflows for Scientific Discovery. Paper

• STORM. GitHub
• GPT Researcher. GitHub
• OpenScholar. GitHub
• PaperQA2. GitHub
• Open Deep Research. GitHub
• DeerFlow. GitHub
• OpenHands. GitHub
• Aider. GitHub
• SWE-agent. GitHub
• autoresearch. GitHub
• Agent Laboratory. GitHub
• Auto-Claude-Code-Research-in-Sleep (ARIS). GitHub

• LitLLM: A toolkit for scientific literature review. Paper
• DeepResearcher: Scaling deep research via reinforcement learning in real-world environments. Paper
• SciSage: A multi-agent framework for high-quality scientific survey generation. Paper
• ResearchAgent: Iterative research idea generation over scientific literature with large language models.
• DeepResearchAgent. GitHub
• HypER: Literature-grounded hypothesis generation and distillation with provenance.
• CiteME: Can Language Models Accurately Cite Scientific Claims?
• PaperGuide: Making Small Language-Model Paper-Reading Agents More Efficient. Paper
• Sparks of science: Hypothesis generation using structured paper data. Paper
• Tongyi DeepResearch. GitHub

• AIGS: Generating science from AI-powered automated falsification. Paper
• Scideator: Human-LLM scientific idea generation grounded in research-paper facet recombination. Paper
• Chain of ideas: Revolutionizing research via novel idea development with LLM agents. Paper
• Jr. AI Scientist and Its Risk Report: Autonomous Scientific Exploration from a Baseline Paper. Paper
• MirrorMind: Empowering OmniScientist with the expert perspectives and collective knowledge of human scientists. Paper
• Idea2Paper. GitHub
• Nova: An iterative planning and search approach to enhance novelty and diversity of LLM generated ideas. Paper
• NovelSeek: When Agent Becomes the Scientist.

• ChemCrow: Augmenting large-language models with chemistry tools. Paper
• MM-Agent: LLM as agents for real-world mathematical modeling problem. Paper
• BioPlanner: automatic evaluation of LLMs on protocol planning in biology.
• Biomni: A Multi-Modal Agent for Comprehensive Biomedical Research. Paper
• OR-Agent: Bridging Evolutionary Search and Structured Research for Automated Algorithm Discovery. Paper
• SciDER: Scientific Data-centric End-to-end Researcher. Paper
• SciDataCopilot: An Agentic Data Preparation Framework for AGI-driven Scientific Discovery. Paper
• AutoLabs: Cognitive multi-agent systems with self-correction for autonomous chemical experimentation. Paper
• ORGANA: A robotic assistant for automated chemistry experimentation and characterization.
• A multiagent-driven robotic AI chemist enabling autonomous chemical research on demand.
• Agents for self-driving laboratories applied to quantum computing. Paper

• ControlA: Agentic Workflow Control Mechanisms for Reliable Science.
• BLADE: Benchmarking language model agents for data-driven science.
• EXP-Bench: Can AI conduct AI research experiments? Paper
• PaperBench: Evaluating AI's Ability to Replicate AI Research. Paper
• SPOT: When AI co-scientists fail. Paper
• ResearchBench: Benchmarking LLMs in scientific discovery via inspiration-based task decomposition. Paper
• ResearcherBench: Evaluating deep AI research systems on the frontiers of scientific inquiry. Paper
• LLM-REVal: Can We Trust LLM Reviewers Yet? Paper
• AIRS-Bench: Benchmarking AI-driven research systems for scientific discovery.
• BioDisco: Multi-agent hypothesis generation with dual-mode evidence, iterative feedback and temporal evaluation. Paper
• BioDSA-1K: Benchmarking data science agents for biomedical research. Paper
• FIRE-Bench: Evaluating Agents on the Rediscovery of Scientific Insights. Paper
• The Story is Not the Science: Execution-Grounded Evaluation of Mechanistic Interpretability Research. Paper
• Towards Execution-Grounded Automated AI Research. Paper

• WritingBench: A comprehensive benchmark for generative writing. Paper
• AgentRxiv: Towards collaborative autonomous research. Paper
• AIssistant: An Agentic Approach for Human-AI Collaborative Scientific Work on Reviews and Perspectives in Machine Learning. Paper
• AutoPR: Let's Automate Your Academic Promotion! Paper
• DRPG: An Agentic Framework for Academic Rebuttal. Paper
• PaperGuide: Making Small Language-Model Paper-Reading Agents More Efficient. Paper

• Auto-Bench: An automated benchmark for scientific discovery in LLMs. Paper
• ResearchBench: Benchmarking LLMs in scientific discovery via inspiration-based task decomposition. Paper
• ResearcherBench: Evaluating deep AI research systems on the frontiers of scientific inquiry. Paper
• AIRS-Bench: Benchmarking AI-driven research systems for scientific discovery.

• EXP-Bench: Can AI conduct AI research experiments? Paper
• MLAgentBench: Evaluating language agents on machine learning experimentation. Paper
• PaperBench: Evaluating AI's Ability to Replicate AI Research. Paper
• BLADE: Benchmarking language model agents for data-driven science.
• SPOT: When AI co-scientists fail. Paper

• CiteME: Can Language Models Accurately Cite Scientific Claims?
• LLM-REVal: Can We Trust LLM Reviewers Yet? Paper
• The 2025 AI Agent Index: Documenting technical and safety features of deployed agentic AI systems. Paper

• The AI Scientist-v2: Workshop-level automated scientific discovery via agentic tree search. Paper
• DeepScientist: Advancing frontier-pushing scientific findings progressively. Paper
• Agent Laboratory: Using LLM agents as research assistants. Paper
• Curie: Toward rigorous and automated scientific experimentation with AI agents. Paper
• MLR-Copilot: Autonomous machine learning research based on large language models agents. Paper
• Olympiad-level formal mathematical reasoning with reinforcement learning.
• AlphaEvolve: A coding agent for scientific and algorithmic discovery. Paper

• AI Feynman: A physics-inspired method for symbolic regression. Paper
• Probabilistic weather forecasting with machine learning. Paper
• Scaling deep learning for materials discovery. Paper
• PhysMaster: Building an Autonomous AI Physicist for Theoretical and Computational Physics Research. Paper
• SciAgents: automating scientific discovery through bioinspired multi-agent intelligent graph reasoning. Paper
• Agents for self-driving laboratories applied to quantum computing. Paper

• EmbodiedClaw: embodied-AI development workflow automation.
• RoboClaw: long-horizon robot-learning lifecycle automation.
• AutoRT: real-world robot fleet orchestration.
• AutoEval: automated policy evaluation and scene reset.
• RoboTwin and RoboTwin 2.0: generative digital twins and scalable bimanual data generation.
• GenSim and GenSim2: language-model-based task and demonstration generation.
• MimicGen and SkillMimicGen: demonstration amplification for embodied skills.

• Autonomous chemical research with large language models. Paper
• An autonomous laboratory for the accelerated synthesis of inorganic materials. Paper
• A mobile robotic chemist. Paper
• ChemCrow: Augmenting large-language models with chemistry tools. Paper
• ORGANA: A robotic assistant for automated chemistry experimentation and characterization.
• AutoLabs: Cognitive multi-agent systems with self-correction for autonomous chemical experimentation. Paper
• ChemOS: An orchestration software to democratize autonomous discovery. Paper
• GNoME: graph networks for materials exploration. Article

• Towards an AI co-scientist. Paper
• CellVoyager: AI compbio agent generates new insights by autonomously analyzing biological data.
• BioPlanner: automatic evaluation of LLMs on protocol planning in biology.
• Genesis: towards the automation of systems biology research. Paper
• SpatialAgent: An autonomous AI agent for spatial biology.
• BioDSA-1K: Benchmarking data science agents for biomedical research. Paper
• Biomni: A Multi-Modal Agent for Comprehensive Biomedical Research. Paper

• LLM-based agentic systems in medicine and healthcare.
• Exploring the dilemma of AI use in medical research and knowledge synthesis: A perspective on deep research tools.
• AgentClinic: a multimodal agent benchmark to evaluate AI in simulated clinical environments. Paper
• Autonomous Agentic AI for Clinical Workflow Orchestration: Self-Managing Healthcare Operations.
• Agentic AI framework for end-to-end medical data inference.
• Towards an AI biomedical scientist: Accelerating discoveries in neurodegenerative disease.
• Towards a Medical AI Scientist. Paper

• HybridQuestion: Human-AI Collaboration for Identifying High-Impact Research Questions. Paper
• HLER: Human-in-the-Loop Economic Research via Multi-Agent Pipelines for Empirical Discovery. Paper
• AI for social science and social science of AI: A survey.
• Can we automatize scientific discovery in the cognitive sciences? Paper
• The adoption of Large Language Models in economics research.

• A foundation model for the Earth system.
• Probabilistic weather forecasting with machine learning. Paper
• Accurate medium-range global weather forecasting with 3D neural networks. Paper
• Learning skillful medium-range global weather forecasting. Paper
• ClimaX: A foundation model for weather and climate. Paper