The Null Token Knows: Reducing Message-Free Hallucination in ASR and NMT

2026-08-16

AI Summary

This study addresses the issue of fluent hallucinations generated by ASR and NMT models in the absence of valid input, proposing the native null token as a diagnostic lens. Through auditing null token scores, scalar logit offsets, and decoder state probing, we demonstrate that the null token encodes valid abstention signals. We advocate evaluating abstention strategies by jointly considering suppression efficacy and deletion costs. Experiments indicate that while boosting null token scores significantly mitigates hallucinations, it necessitates balancing the risk of erroneously deleting legitimate content. Consequently, this work establishes a novel paradigm for alleviating no-signal hallucinations that effectively reconciles safety with practical utility.

Abstract

Modern encoder-decoder systems can produce fluent text even when their input contains no recoverable message. We study this failure in ASR and NMT through the models' reserved null tokens, asking whether the score for ending generation already carries a usable abstention signal. Across speech recognizers and translation models, we audit native null-token scores and scalar logit shifts. In Whisper, we additionally probe decoder states and compare supervised row edits with conventional external gates. The evaluated models often expose a useful abstention signal, but stock decoding does not reliably act on it. Raising the null-token score can sharply suppress fabrication, but aggressive intervention also deletes valid speech or shortens legitimate translations. These findings turn the null token into a diagnostic lens on hallucination and motivate evaluating abstention methods by both suppression and deletion costs, rather than by hallucination reduction alone.

Problem

• Hallucination
• ASR
• NMT
• Null Token
• Abstention

Innovation

• Null Token
• Hallucination
• Abstention Signal
• ASR
• NMT