Towards grounded autonomous research: an end-to-end LLM mini research loop on published computational physics

Researchers demonstrate an autonomous LLM agent capable of executing a complete research loop—reading, reproducing, critiquing, and extending computational physics papers. Testing across 111 papers reveals the agent identifies substantive flaws in 42% of cases, with 97.7% of issues requiring actual computation to detect, and produces a publishable peer-review comment on a Nature Communications paper without human direction.

This research represents a significant milestone in autonomous AI-driven scientific discovery, moving beyond theoretical capability toward practical implementation in complex physical science domains. The agent demonstrates competence across three distinct tasks: computational reproduction of published results, identification of non-trivial methodological and analytical flaws, and generation of original scientific contributions at publication quality. The scale test (111 papers) validates the system's generalization ability, while the depth case study (MOSFET multiscale simulation) proves the agent can conduct independent analysis that revises established conclusions.

The finding that 97.7% of identified flaws require execution rather than static analysis is particularly significant—it demonstrates the agent moves beyond pattern-matching and actually engages in computational verification. This capability addresses a critical gap in scientific reproducibility, where many published results contain errors detectable only through replication. The autonomous generation of a publishable Comment on a Nature Communications paper suggests the system operates at research frontiers, not merely routine tasks.

For the AI research community, this validates scaling autonomous reasoning to domains requiring deep physical grounding and literature synthesis. It implies future scientific acceleration where LLM agents could systematically audit published work, identify gaps, and propose extensions. However, the work remains largely confined to computational physics—domains with well-defined mathematical frameworks, publishable code, and quantifiable outputs. Generalization to experimental sciences or fields with qualitative reasoning remains unclear. The implications extend beyond academia to scientific software verification and quality assurance in computational modeling used across industry.

• →LLM agents successfully autonomously reproduce, critique, and extend computational physics research at publication quality without human intervention.
• →The agent identified substantive flaws in 42% of 111 tested papers, with 97.7% of issues requiring actual computation to surface.
• →An unsupervised autonomous agent produced a publishable peer-review Comment revising a Nature Communications paper's headline conclusion.
• →This capability addresses scientific reproducibility gaps by systematically executing computational verification beyond static analysis.
• →Scaling potential exists for scientific audit and discovery, though generalization beyond computational domains remains uncertain.