Creative Ownership in the Age of AI
This paper discusses copyright law challenges posed by generative AI, proposing a new infringement criterion based on training corpus dependence.
Agentic Test-Time Scaling for WebAgents
Introducing CATTS, this research provides a technique for dynamically allocating compute in multi-step tasks, improving performance and efficiency.
AttentionRetriever: Attention Layers are Secretly Long Document Retrievers
Presentation of AttentionRetriever, a model for retrieving long-document context, outperforming existing models while maintaining efficiency.
CM2: Reinforcement Learning with Checklist Rewards for Multi-Turn and Multi-Step Agentic Tool Use
CM2 presents a novel reinforcement learning framework for optimizing agent tool usage, offering strong improvements over supervised approaches.
Think like a Scientist: Physics-guided LLM Agent for Equation Discovery
Introducing KeplerAgent, this study showcases a framework for scientific reasoning using LLMs to enhance accuracy in symbolic equation discovery.
Automated Test Suite Enhancement Using Large Language Models with Few-shot Prompting
Exploration of using few-shot prompting to enhance the quality and usability of LLM-generated unit tests, emphasizing the effective integration of human examples.
ExtractBench: A Benchmark and Evaluation Methodology for Complex Structured Extraction
ExtractBench offers a comprehensive framework for evaluating structured extraction from PDFs, emphasizing reliability and standardization.
When Scaffolding Breaks: Investigating Student Interaction with LLM-Based Writing Support in Real-Time K-12 EFL Classrooms
An exploration of LLMs in classroom settings reveals challenges of lower-proficiency students relying on AI scaffolding, leading to design guidelines for inclusivity.
Towards Autonomous Mathematics Research
Introducing Aletheia, a math research agent advancing AI capabilities in solving and generating research papers through autonomous reasoning.