Researchers propose a method called Triage that can predict and prune audio tokens in large audio language models (LALMs) before the language model runs, achieving significant compression and improving performance.

A game-theory-inspired framework, Stackelberg Alignment, is proposed for language models (LLMs) to collaborate and improve collectively. The framework uses an EXP3 bandit to select instructions for LLMs to respond to, and the LLMs learn from each other's responses through peer judgment and reputation-based matching.

PriceCheck is a new method for controlling the risk of serving answers from a large language model by selecting which checks to run and when to stop. It uses a compact family of decision rules built from label-free checks and prices them to guide the selection process. Experiments show that PriceCheck outperforms other methods in terms of serving answers and keeping selective risk low.

Researchers propose a novel framework, Less Uniform Diffusion (LUDI), to improve the scalability of Uniform Diffusion Language Models (UDLMs). LUDI addresses the issues of over-uniform training objectives and condition-target confusion in UDLMs by introducing a less uniform loss and token-level corruption hints. This allows for confidence-based few-step sampling, resulting in cleaner supervision and improved generation capabilities.

Researchers proposed InMyStyle, a system that adapts small language models to rewrite AI-edited text to an individual user's writing style without an instruction prompt. They found that small models (0.5B-7B parameters) are sufficient for the task, with model size mainly affecting efficiency rather than quality. The study evaluated the system using a single-user case study and a secondary evaluation with LLM judges.