A new study published on ArXiv investigates the phenomenon of "confident fabrication" in Large Language Models (LLMs), utilizing the TAME Pain clinical speech corpus. The research evaluated seven models on their ability to abstain from making predictions when the provided evidence is provably uninformative.
The Methodology of Zero Signal
The researchers utilized 5,750 "no signal" utterances consisting of Harvard Sentences read by participants under cold or warm water stress. While pain was recoverable from acoustic features (AUC 0.622), transcript-based prediction remained near chance levels (AUC 0.489). Since automatic speech recognition (ASR) discards acoustic cues, any pain score derived solely from the transcript is factually unsupported.
Authority Framing and Model Fragility
The study found that model robustness is highly sensitive to prompting styles:
- Under cooperative prompting, six models correctly abstained from nearly all uninformative transcripts.
- Under authority-framed prompts, abstention rates became highly volatile, ranging from 0.18 to 1.00 for the same model across different phrasings.
- When forced to provide a response, Gemini 2.5 Flash and Llama 3.1 8B exhibited high rates of confident fabrication (0.53 and 0.76), significantly higher than the 0.15 maximum of other models.
The researchers noted that no significant demographic effects were observed in the forced responses (p ≥ 0.20), suggesting that the fabrication issue is a structural characteristic of the models' response generation under pressure rather than a demographic bias.
