Eliciting Attributions from LLMs with Minimal Supervision

Ramakanth Pasunuru, Koustuv Sinha, Armen Aghajanyan, LILI YU, Tianlu Wang, Daniel M. Bikel, Luke Zettlemoyer, Maryam Fazel-Zarandi, Asli Celikyilmaz

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Attribution, Large Langauge Models, Instruction Tuning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We show that its remarkably easy to align a LLaMA model to elicit source attributions for a given question by leveraging on its memorization abilities. LLaMA models also improve their downstream performance when provided their own attributions!

Abstract:

Large Language Models (LLMs) have quickly become popular tools for recalling information; however, the specific mechanisms they utilize to encode and store vast information within their parameters is not well understood. As a step towards improving interpretability and reliability of LLMs, we develop AttributeLLaMA, which involves fine-tuning LLaMA models to enable them to attribute their responses. By utilizing a mere 100 expert-annotated attribution examples, we are able to achieve this capability. Our experimental studies demonstrate that these attributions significantly improve the performance of a strong LLM on downstream tasks such as MMLU and StrategyQA by more than 3.5% and 4.0%, respectively. Furthermore, our analyses on the attributions obtained from AttributeLLaMA reveal the remarkable memorization capabilities of LLMs

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6014