New Research Explores LLM Agents for Engineering Optimization and Scientific Tasks
Several new research papers published on arXiv examine how large language model (LLM) agents can tackle complex technical workflows across multiple domains.
According to arxiv.org, researchers introduced ORFS-agent, an LLM-based system that automates parameter tuning in chip design flows. The system demonstrated improvements over standard Bayesian optimization approaches. Testing across six benchmarks showed that “thinking-model backends (Sonnet 4.6 and Kimi K2.5) improve the geometric-mean normalized wirelength, effective clock period, and co-optimization objectives by up to 1.0%, 1.3%, and 2.7% over OR-AutoTuner while using 40% fewer iterations,” according to the paper. The open-weight Kimi K2.5 model remained within 0.24% of Sonnet 4.6’s performance.
In a separate paper on arxiv.org, researchers presented a framework for evaluating tool-calling agents during inference rather than after execution. The approach achieved “+5.5% on irrelevance detection and +7.1% on multi-turn tasks” according to the authors, who introduced “Helpfulness-Harmfulness metrics” to measure whether reviewer feedback provides net positive value.
A third study on arxiv.org compared interaction paradigms for scientific visualization tasks, finding that “general-purpose coding agents achieve the highest task success rates but are computationally expensive, while domain-specific agents are more efficient and stable but less flexible.”
Finally, researchers on arxiv.org presented CARE (Collaborative Agent Reasoning Engineering), a methodology for engineering LLM agents in scientific domains through “reusable artifacts and systematic, stage-gated phases.”