Where Does the Relative Clause Attach? Probing Prosodic and Semantic Sensitivity in Large Audio Language Models

Despite extensive research on human sentence processing, little is known about whether large audio language models (LALMs) exhibit similar sensitivity to prosodic and semantic cues during syntactic ambiguity resolution. We evaluated the ability of five state-of-the-art LALMs to resolve relative clause (RC) attachment ambiguities across three typologically diverse languages: English, Chinese, and Arabic. Following the Implicit Prosody Hypothesis (Fodor, 1998, 2002) and based on the MultiWho dataset (Lee et al., 2025), we constructed stimuli including two conditions: (1) prosodically biased sentences with manipulated intonational phrasing and boundary placement, and (2) semantically biased sentences where world knowledge favors one attachment site. We generated audio stimuli via text-to-speech synthesis with controlled prosodic contours signaling either high or low RC attachment. Each stimulus was probed using yes/no and forced-choice prompt formats. The findings revealed that LALMs show limited and inconsistent sensitivity to prosodic manipulations, with low attachment remaining dominant regardless of prosodic condition. Forced-choice prompts yielded more stable attachment preferences and fewer uncertain responses than yes/no prompts. Models demonstrated greater sensitivity to semantic and world knowledge cues than to prosodic cues. Considerable cross-linguistic and cross-model variation was observed. These results raise questions about the cognitive capacities of current LALMs, suggesting that their syntactic processing is driven primarily by semantic information, with prosodic cues playing a limited and inconsistent role in ambiguity resolution.