LN — From AI for Science to Agentic Science (Zhang et al., 2025)

Full Reference: Zhang, X., et al. (2025). From AI for science to Agentic Science: A survey on autonomous scientific discovery. arXiv preprint arXiv:2508.14111. https://doi.org/10.48550/arXiv.2508.14111


Pass 1 — Bird’s Eye

Type: Survey paper Main Claim: “Agentic Science” represents a new research paradigm where AI systems manage complete investigation cycles — from hypothesis to conclusion — with limited human oversight. Relevance to PUMA: ⭐⭐⭐⭐ High — provides the theoretical framework for classifying PUMA’s Smart PMO as applied Agentic Science.

Pass 2 — Key Content

Three-Phase Framework (aligned with LeCun’s levels)

  1. AI for Science (tool phase): Domain-specific models, literature synthesis, data analysis
  2. AI in Science (collaborative phase): Co-pilot systems, closed-loop experiments in specific domains
  3. Agentic Science (autonomous phase): AI agents manage full research cycles across domains

Key Application Domains Surveyed

  • Life Sciences: drug discovery, protein design, genomics
  • Chemistry: synthesis planning, catalyst design, reaction prediction
  • Materials: crystal structure prediction, property optimization
  • Physics: model discovery, simulation acceleration

Critical Requirements for Agentic Science

  1. Hypothesis formulation: Ability to generate specific, testable predictions
  2. Experiment planning: Selection of experimental conditions to maximally inform hypotheses
  3. Data interpretation: Automated extraction of conclusions from experimental data
  4. Knowledge integration: Grounding new results in the existing knowledge base

Identified Challenges

  • Scalability of data infrastructure for robotic experiments
  • Interpretability of AI-generated discoveries
  • Attribution and credit for AI contributions
  • Reproducibility of stochastic AI experiments

Pass 3 — PUMA Re-implementation

The survey’s framework maps directly to PUMA’s research design:

  • PUMA Phase 1 (Research) = “AI for Science” phase: using AI tools for literature synthesis
  • PUMA Phase 2 (Development) = “AI in Science” phase: building the benchmark with AI assistance
  • PUMA Stage 5 (Smart PMO) = “Agentic Science” phase: autonomous PM workflow cycles

MIT Critical Questions

  1. How can I use this? → Cite in PUMA Section 1.1 to position PUMA within the Agentic Science trajectory.
  2. Does it really support its claims? → Yes — well-documented with concrete examples per domain.
  3. What if PM doesn’t fit this framework? → PM lacks the “ground truth” verification mechanism of physical experiments; PUMA addresses this with labeled datasets (Jira SR, TAWOS).

APA7 Citation

Zhang, X., et al. (2025). From AI for science to Agentic Science: A survey on autonomous scientific discovery. arXiv preprint arXiv:2508.14111. https://doi.org/10.48550/arXiv.2508.14111


MOCs