On the Tip of the Tongue: Analyzing Conceptual Representation in Large Language Models with Reverse-Dictionary Probe

TL;DR

In-context learning is used to guide the models to generate the term for an object concept implied in a linguistic description to probe LLMs' capacity for conceptual inference in the reverse dictionary task.

Abstract

Probing and enhancing large language models' reasoning capacity remains a crucial open question. Here we re-purpose the reverse dictionary task as a case study to probe LLMs' capacity for conceptual inference. We use in-context learning to guide the models to generate the term for an object concept implied in a linguistic description. Models robustly achieve high accuracy in this task, and their representation space encodes information about object categories and fine-grained features. Further experiments suggest that the conceptual inference ability as probed by the reverse-dictionary task predicts model's general reasoning performance across multiple benchmarks, despite similar syntactic generalization behaviors across models. Explorative analyses suggest that prompting LLMs with description$\Rightarrow$word examples may induce generalization beyond surface-level differences in task construals and facilitate models on broader commonsense reasoning problems.

Figures (13)

• Figure 1: Illustration of the reverse-dictionary probe. A list of $N$ description--word demonstrations is used to prompt an LLM to favorably evoke its conceptual inference capacity. The model generates a word/phrase for the object concept that is described in the query.
• Figure 2: Performance of LLMs in the prompted reverse dictionary task when provided with $N$ description--word pairs. Model performance is measured by exact match between the word/phrase decoded from the model and the name of the specific object for that description. Colored bands denote 95% confidence intervals.
• Figure 3: A t-SNE visualization of representations derived from LLaMA2-13B under different task conditions. Representations are extracted at the "$\Rightarrow$" symbol. Category assignments are based on the THINGS data.
• Figure 4: Correlation between LLMs' overall performance averaged across different reasoning tasks and their average conceptual inference performance in the reverse dictionary task with 24 demonstrations provided.
• Figure 5: Correlation between the LLMs' syntactic generalization ability, as measured by BLiMP (Left) and SyntaxGym (Right), and their average performance in the conceptual inference task with 24 demonstrations.