Rethinking Scientific Discovery in the Agentic Era
TLDR
SCION is an agentic scientific operating system that coordinates tasks, tools, and agents via a Research Execution Plan for long-horizon discovery.
Reasoning
The paper presents a novel framework (SCION) that integrates multiple scientific tasks and agents, with clear methodology (REP, hierarchical execution) and real-world applications in materials, molecules, and antibodies. However, the abstract is cut off, limiting visibility of quantitative results and comparisons, and some claimed capabilities (e.g., literature review) are not explicitly detailed.
Read-first score
Read-first score 70.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 80.
Field roles
Rank sensitivity
Stability: volatile; rank range: 12.
Keyword Scores
Deep Analysis
Innovations
- SCION as an agentic scientific operating system that acts as an organizational nexus, with a Science Agent as Meta-Harness connecting tasks, tools, agents, artifacts, and memory
- Research Execution Plan (REP) that compiles high-level scientific intent into staged objectives, dependencies, verification checkpoints, tool requirements, expected artifacts, and fallback conditions
- Hierarchical multi-agent execution with profile-driven specialization, selective context construction, governed delegation, and layered epistemic memory for long-horizon scientific work
- Formulation of discovery as Target-conditioned Inverse Search, extended to hidden-target settings via batch active search under finite experimental budgets
- Demonstrated strong performance in decomposition, verification, refinement, and memory reuse over existing autonomous research-agent baselines
Methodology
SCION is an agentic system where a Science Agent as a Meta-Harness orchestrates scientific tasks, tools, agents, artifacts, and memory. The core is a Research Execution Plan (REP) that breaks down high-level goals into staged objectives with verification and fallback. It employs hierarchical multi-agent execution, profile specialization, selective context, governed delegation, and layered epistemic memory. Discovery is formalized as target-conditioned inverse search, extended with batch active search for hidden targets. Evaluation includes materials analysis, molecule design, protein/antibody screening, and specific tasks like scientific reading, idea generation, molecule generation, and antibody screening, compared against existing autonomous research-agent baselines.
Key Results
SCION outperforms existing autonomous research-agent baselines, particularly in decomposition, verification, refinement, and memory reuse, across tasks in scientific reading, idea generation, molecule generation, and antibody screening.