Large Language Model Agents for Autonomous Scientific Research: A Lifecycle Survey of Methods, Systems, Benchmarks, and Risks

Large language models have moved from passive text generators to autonomous agents that plan, use tools, and act across the scientific research process. A fast-growing body of work now targets every stage of that process, from literature search and hypothesis generation through experiment code, wet-lab automation, manuscript writing, and peer review, together with discipline-specific agents for chemistry, biology, materials, medicine, and mathematics and a large family of benchmarks built to measure them. This review combines a systematic search with a narrative synthesis. A protocol-driven search of arXiv, Semantic Scholar, bioRxiv, and fourteen conference proceedings, run to 30 May 2026, retrieved 2928 records, of which 883 met the eligibility criteria and were coded, and 253 were analysed in depth. We organize that literature along the scientific research lifecycle and cross-cut it by four design axes, namely agent architecture, learning mechanism, autonomy level, and validation regime, and we report the publication status and evidence level of every claim we rely on. The field has matured unevenly. Retrieval-grounded deep research and machine learning engineering show repeatable gains on peer-reviewed benchmarks, whereas open-ended ideation, end-to-end discovery, and trustworthy reporting remain fragile. Realistic benchmarks report success rates between one sixth and one third, with a Kaggle bronze medal on 16.9 percent of engineering competitions, a best replication score of 21.0 percent across 20 published papers, 32.4 percent of scientific coding tasks solved, and 31.5 percent accuracy on memorization-resistant equation discovery. On a three-paper subset of the replication benchmark, an o1 agent scores 26.6 percent against 41.4 percent for machine learning PhDs. Closed-loop discovery systems report results confirmed in the wet lab, and all three such systems in our master comparison remain preprints, which places those claims on weaker evidential footing than the benchmark results. A reliability gap runs through the whole lifecycle, as agents that retrieve and write fluently still fabricate results, miscite sources, and overstate confidence. We contribute a documented search and coding protocol with the full coded corpus made available, a unified taxonomy, a master comparison of systems, a structured account of benchmarks and metrics, and a risk-to-mitigation map, and we close with a maturity assessment and a research roadmap for safe and verifiable autonomous science.