arXiv Trend Report: AI Research Highlights for February 23, 2026

This daily arXiv trend report covers the most-noticed machine learning and AI papers on February 23, 2026, helping researchers and engineers quickly grasp frontier developments.

**Today's high-attention papers**: ① New research on 'inference-time compute scaling' exploring how to improve performance by increasing inference computation without adding model parameters; ② New multimodal RAG architecture capable of simultaneously retrieving text, images, and tabular data; ③ Lightweight RLHF (Reinforcement Learning from Human Feedback) method reducing alignment training costs by 70%.

Trend-wise, papers on 'Test-Time Compute' have tripled in the past two weeks, becoming one of the hottest AI research sub-directions — reflecting strong research community influence from the success of o1/o3 series models.