Reproduce “Think2Listen: Evidence-Centric Learning For Agentic Audio Understanding With LALMS” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jin Wang, Liang Lin, Kaiwen Luo, Weiliu Wang, Yuxuan Li, Zengxin Han, Gan Zhang, Jingtong Wu and 2 more

Think2Listen addresses audio question answering by training an agent to ask cue-specific questions about recordings and revise candidate answers using returned observations. Its Transition-as-Credit training signal rewards useful changes in plausible answer sets. On MMAR, it reports 69.25% ± 0.28, versus AudioToolAgent’s 65.90%, a 3.35-point difference. This benchmark-level result is not a fully controlled comparison across external baselines.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.