Researchers introduce RUBATO, a novel multistream positional encoding that decouples the decoding rate of internal thinking streams in Full-Duplex models, enabling them to process at hardware-adaptive rates and improve accuracy and response time.

AudioTETRIS is a training-free, two-stage framework for compressing audio tokens in Large Audio Language Models (LALMs) without losing performance. It uses the audio encoder's representation geometry as an acoustic prior to compute a per-token fused score, which drives progressive top-$k$ token selection. This results in near-uncompressed performance using only 25% of audio tokens.

Researchers propose a black-box causal test to address object hallucination in large audio-language models (LALMs). They introduce a method to localize and remove the evidence of hallucination, resulting in improved accuracy and a deeper understanding of the grounding regimes of different LALM models.

DuplexAgent-RSI is a full-duplex voice agent collaboration system that enables continuous listening and speaking while handling complex reasoning and tool use through a modular, verifiable loop. Its harness expresses the workflow as six editable modules, which are revised from interaction traces using a simulator that generates timed test conversations, identifies failures, and proposes targeted module changes.