Researchers have introduced several new approaches for improving AI model performance during inference through test-time adaptation techniques.
According to arxiv.org, a framework called Query-Conditioned Test-Time Self-Training (QueST) enables large language models to adapt their parameters during inference using supervision derived directly from input queries. The system generates query-conditioned problem-solution pairs for parameter-efficient fine-tuning at test time. Across seven mathematical reasoning benchmarks and the GPQA-Diamond scientific reasoning benchmark, QueST consistently outperformed existing test-time optimization baselines, according to the paper.
In a separate development, arxiv.org reports that AgenticRecTune, a multi-agent framework leveraging Gemini, addresses configuration optimization in recommendation systems. The system comprises five specialized agents (Actor, Critic, Insight, Skill, and Online) and features a “self-evolving Skillhub” that summarizes historical results and updates skills.
According to arxiv.org, researchers also introduced Seg-Agent, a training-free framework for language-guided image segmentation. The system uses “Explicit Multimodal Chain-of-Reasoning” with Set-of-Mark visual prompting, allowing models to iteratively reason about spatial relationships in the visual domain. Seg-Agent achieved “performance comparable to state-of-the-art training-based methods without any parameter updates,” according to the paper.
Additionally, arxiv.org reports ChatSR, a multimodal language model designed for scientific data understanding that can automatically generate mathematical formulas based on user-specified constraints.