Interpreting the internal process of neural models has long been a challenge. This challenge remains relevant in the era of large language models (LLMs) and in-context learning (ICL); for example, ICL poses a new issue of interpreting which example in the few-shot examples contributed to identifying/solving the task. To this end, in this paper, we design synthetic diagnostic tasks of inductive reasoning, inspired by the generalization tests in linguistics; here, most in-context examples are ambiguous w.r.t. their underlying rule, and one critical example disambiguates the task demonstrated. The question is whether conventional input attribution (IA) methods can track such a reasoning process, i.e., identify the influential example, in ICL. Our experiments provide several practical findings; for example, a certain simple IA method works the best, and the larger the model, the generally harder it is to interpret the ICL with gradient-based IA methods.
View on arXiv@article{ye2025_2412.15628,
title={ Can Input Attributions Interpret the Inductive Reasoning Process Elicited in In-Context Learning? },
author={ Mengyu Ye and Tatsuki Kuribayashi and Goro Kobayashi and Jun Suzuki },
journal={arXiv preprint arXiv:2412.15628},
year={ 2025 }
}