A new arXiv paper tackles the problem of choosing which demonstration examples to include in a prompt when using in-context learning with large language models. Because the space of possible example subsets is combinatorially large, the authors propose a data-efficient approach to selecting good combinations without exhaustive search. The work aims to improve how LLMs adapt to new tasks without fine-tuning.
A newly posted arXiv paper introduces a method for adapting graph-based prediction models to new data with limited supervision or feedback. The approach, called reliability-aware prototype learning, aims to make agentic systems more data-efficient when reusing prior knowledge after deployment. It falls under the machine learning category and is a preprint, not yet peer-reviewed.
An arXiv paper describes a long-horizon autonomous research program by Qiushi Engine targeting the BabyLM 2026 Strict-Small benchmark, which caps training data at 10 million words. The work examines how models can learn efficiently from limited text while retaining capabilities and generalizing to new inputs.