Conceptual and Unbiased Reasoning in Language Models
This addresses the problem of improving generalization and reducing bias in AI systems for applications requiring high-level abstract reasoning, though it is incremental in enhancing existing models.
The paper tackles the problem of limited conceptual reasoning in large language models by proposing a novel framework that forces models to reason abstractly, showing a 9% to 28% drop in performance compared to direct inference. It then introduces techniques that improve conceptual reasoning by 8% to 11%, leading to more robust and less biased decision-making.
Conceptual reasoning, the ability to reason in abstract and high-level perspectives, is key to generalization in human cognition. However, limited study has been done on large language models' capability to perform conceptual reasoning. In this work, we bridge this gap and propose a novel conceptualization framework that forces models to perform conceptual reasoning on abstract questions and generate solutions in a verifiable symbolic space. Using this framework as an analytical tool, we show that existing large language models fall short on conceptual reasoning, dropping 9% to 28% on various benchmarks compared to direct inference methods. We then discuss how models can improve since high-level abstract reasoning is key to unbiased and generalizable decision-making. We propose two techniques to add trustworthy induction signals by generating familiar questions with similar underlying reasoning paths and asking models to perform self-refinement. Experiments show that our proposed techniques improve models' conceptual reasoning performance by 8% to 11%, achieving a more robust reasoning system that relies less on inductive biases.