Several new research papers on arXiv explore different approaches to improving large language model efficiency and performance.
According to a paper submitted to IEEE BigData 2026, researchers propose a behavior-aware framework for constructing diverse LLM crowds for future prediction tasks. The study evaluated 25 LLMs and found that “a three-model medoid crowd based on K-means++ behavioral clustering outperforms conventional voting over all 25 models on both prediction benchmarks, while reducing model calls by 88% and inference cost by approximately 80%.” The research suggests that representative behavioral diversity is more important than simply maximizing diversity.
Another paper introduces Parason, a system that reveals both subtask and trial parallelism in LLM reasoning. According to the research, trial parallelism accounts for “65.5% of parallelizable reasoning computation in DeepSeek-V4’s reasoning steps,” and becomes increasingly dominant on harder problems. The system achieves “an average acceleration about 1.7× while maintaining competitive accuracy” on mathematical reasoning benchmarks including AIME24 and AIME25.
Additionally, research accepted at TMLR 2026 proposes COPE, a test-time collaboration framework where small and large models exchange plans in a multi-stage cascade. According to the paper, COPE “achieves performance comparable to large proprietary models, while drastically reducing the inference API cost.”
A separate paper examines quality degradation in LLM inference under compute constraints, analyzing how service degradation affects retry rates and customer churn.