An Existence Proof for Language Models That Can Explain Garden-Path Effects via Surprisal

ACL ARR 2026 January Submission4209 Authors

Keywords: Surprisal Theory, Garden-Path Effect, Fine-Tuning

Abstract: Surprisal, defined as the negative log-probability of a word given its context, has been advocated as a metric for modeling human sentence-processing difficulty. While surprisal from recent neural language models (LMs) generally captures human processing difficulty on naturalistic corpora, it severely underestimates processing difficulty on syntactically ambiguous sentences (garden-path effects), leading to the claim that the processing difficulty of such sentences cannot be reduced to surprisal. In this paper, we investigate whether it is truly impossible to construct an LM that can explain garden-path effects via surprisal. Specifically, while previous work has evaluated off-the-shelf pretrained neural LMs, we fine-tune these LMs on garden-path sentences to align surprisal-based reading-time estimates with actual human reading times, and evaluate both the success of this approach and its impact on predictive power for reading times on naturalistic corpora. Our results show that fine-tuning succeeds without degrading (and in fact improves) predictive power for human reading times on naturalistic corpora, providing an existence proof for an LM that can explain both garden-path effects and naturalistic reading times via surprisal.

Paper Type: Long

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Research Area Keywords: cognitive modeling, computational psycholinguistics

Contribution Types: Theory

Languages Studied: English

Submission Number: 4209