Researchers have investigated the phenomenon of "message-free hallucination" in automatic speech recognition (ASR) and neural machine translation (NMT) systems, where models generate fluent text despite having no recoverable message in the input. The study focuses on the role of reserved null tokens within these encoder-decoder systems, examining whether the scores associated with ending generation can provide a usable abstention signal. Findings indicate that while models like Whisper often possess a useful abstention signal, standard decoding methods do not consistently leverage it, suggesting that improving null-token scores could reduce fabrication but requires careful balancing to avoid deleting valid output. AI
IMPACT Identifies a potential mechanism for reducing hallucinations in ASR and NMT, impacting model reliability and output quality.
RANK_REASON Academic paper detailing a new finding about model behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
