Three recent papers published on arXiv address different efficiency and capability challenges facing large language models.
According to a paper published March 20, 2026, researchers propose a utility-guided orchestration policy for tool-using LLM agents that aims to balance answer quality with execution costs. The paper addresses what it describes as “a fundamental tension between answer quality and execution cost” in LLM agents, where methods like ReAct may improve performance but result in “excessive tool calls, longer trajectories, higher token consumption, and increased latency.” The proposed approach selects among actions including respond, retrieve, tool call, verify, and stop by “balancing estimated gain, step cost, uncertainty, and redundancy.”
Separately, researchers introduced Generative Active Testing (GAT) in a paper submitted February 26, 2026, addressing the high cost of creating test sets for LLM benchmarking. According to the paper, GAT uses “LLMs as surrogates” and a “Statement Adaptation Module” to modify generative tasks into a pseudo-classification format. The authors report their zero-shot acquisition functions “reduce estimation error by ~40% compared to traditional sampling baselines.”
A third paper, accepted at DASFAA 2026, presents LLM-MRD (LLM-Guided Multi-View Reasoning Distillation) for fake news detection. According to the authors, the approach addresses “prohibitive reasoning inefficiency due to the high computational costs of LLMs” and demonstrates “a comprehensive average improvement of 5.19% in ACC and 6.33% in F1-Fake when evaluated across all competing methods and datasets.”