arXiv：cs.CL· Jing Chen, Giulia Loca, Simona Amenta, Marco Marelli·· 6 小时前AI 评分36
伪词探针研究：LLM 对支配人类伪词处理的子词汇线索敏感度不足
Pseudowords as probes: Large Language Models show little of the sublexical sensitivity that governs human pseudoword processing
AI 导读
研究用两个意大利语二选一伪词实验测试五个 LLM，并与人类行为基线对比。当选项含真实词提供词汇熟悉度线索时，LLM 与人类一致性更高；在纯伪词条件下则明显低于字符 n-gram 模型 fastText。驱动人类—fastText 一致的子词汇余弦相似度线索未能稳定迁移到人类—LLM 对齐，推理 token 消耗与人类处理难度也无一致关系。
正文
Abstract:Systematicity, the probabilistic mapping of form to meaning, permeates language at all levels, and sublexical cues have been shown to govern human pseudoword processing. Yet whether LLMs exhibit comparable sensitivity to these cues remains unclear. We tested five LLMs on two Italian two-alternative forced-choice pseudoword experiments and compared their responses with a human behavioural baseline. LLMs aligned more reliably with humans when real-word options provided a lexical familiarity cue than in the pseudoword-only condition, where they fell substantially below fastText, a character-n-gram model. In addition, the sublexical cosine-similarity cue that reliably drove human--fastText agreement did not consistently transfer to human--LLM alignment, and reasoning-token expenditure bore no consistent relation to human processing difficulty. These findings suggest that LLMs do not necessarily share the sublexical cues that govern human pseudoword processing; we discuss tokenization and training-data coverage as candidate explanations.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.07936 [cs.CL] |
| (or arXiv:2610.07936v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.07936 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jing Chen [view email]
[v1]
Tue, 6 Oct 2026 08:09:58 UTC (84 KB)
来源：arXiv：cs.CL · arxiv.org