Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

What can LLMs tell us about the mechanisms behind polarity illusions in humans? Experiments across model scales and training steps

Dario Paape

Abstract

I use the Pythia scaling suite (Biderman et al. 2023) to investigate if and how two well-known polarity illusions, the NPI illusion and the depth charge illusion, arise in LLMs. The NPI illusion becomes weaker and ultimately disappears as model size increases, while the depth charge illusion becomes stronger in larger models. The results have implications for human sentence processing: it may not be necessary to assume "rational inference" mechanisms that convert ill-formed sentences into well-formed ones to explain polarity illusions, given that LLMs cannot plausibly engage in this kind of reasoning, especially at the implicit level of next-token prediction. On the other hand, shallow, "good enough" processing and/or partial grammaticalization of prescriptively ungrammatical structures may both occur in LLMs. I propose a synthesis of different theoretical accounts that is rooted in the basic tenets of construction grammar.

What can LLMs tell us about the mechanisms behind polarity illusions in humans? Experiments across model scales and training steps

Abstract

I use the Pythia scaling suite (Biderman et al. 2023) to investigate if and how two well-known polarity illusions, the NPI illusion and the depth charge illusion, arise in LLMs. The NPI illusion becomes weaker and ultimately disappears as model size increases, while the depth charge illusion becomes stronger in larger models. The results have implications for human sentence processing: it may not be necessary to assume "rational inference" mechanisms that convert ill-formed sentences into well-formed ones to explain polarity illusions, given that LLMs cannot plausibly engage in this kind of reasoning, especially at the implicit level of next-token prediction. On the other hand, shallow, "good enough" processing and/or partial grammaticalization of prescriptively ungrammatical structures may both occur in LLMs. I propose a synthesis of different theoretical accounts that is rooted in the basic tenets of construction grammar.

Paper Structure

This paper contains 7 sections, 4 figures.

Table of Contents

  1. Introduction
  2. LLM sentence completion experiments
  3. Materials
  4. Models
  5. Results for NPI sentences
  6. Results for depth charge sentences
  7. Discussion

Figures (4)

  • Figure 1: Difference in summed log probabilities between negative and positive continuations for NPI sentences across prompts and model sizes. More positive y-values mean more negative continuations. Error bars show 95% confidence intervals.
  • Figure 2: Difference in summed log probabilities between negative and positive continuations for NPI sentences in the 1.4b model across prompts and training steps. More positive y-values mean more negative continuations. Error bars show 95% confidence intervals.
  • Figure 3: Difference in summed log probabilities between negative and positive continuations for depth charge sentences across prompts and model sizes. More positive y-values mean more negative continuations. Error bars show 95% confidence intervals.
  • Figure 4: Difference in summed log probabilities between negative and positive continuations for depth charge sentences in the 1.4b model across prompts and training steps. More positive y-values mean more negative continuations. Error bars show 95% confidence intervals.