Apples, Oranges, and Tennis Balls: A Neuro-Symbolic Approach to Facilitate Flexible Retrieval Strategies

ACL ARR 2025 July Submission1103 Authors

Abstract:

Humans exhibit flexible retrieval strategies that allow them to adaptively access different types of semantic knowledge depending on context and goals. In contrast, LLMs struggle with tasks requiring this kind of controlled, adaptive memory access.
In this work, we propose a neuro-symbolic approach to implement flexible retrieval strategies. We demonstrate our
ideas on the NYT Connections puzzle.
The Connections puzzle embodies many cognitive science themes, from how
we store concepts to how we flexibly retrieve them, and its study can offer a new lens to explore this topic.

Our approach significantly outperforms LLM-only baselines, improving average scores by 2-7x across models, enabling models that previously solved 0-5 puzzles to perfectly solve up to 32. We also show that combining smaller, open-source LLMs with symbolic reasoning can outperform larger proprietary models. We make our code and data publicly available.

Paper Type: Long

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Research Area Keywords: cognitive modeling

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 1103