The paper presents a novel fine-tuning approach for LLMs named EXAR (EXclusive AutoRegressive fine-tuning) aimed at injecting metaknowledge about a particular task (in this case Question-Answering) in order to improve or self-correct previous errors.

Title: Eliciting metaknowledge in Large Language Models

Abstract:

The introduction of Large Language Models (LLMs) able to exhibit a number of linguistic and extra-linguistic capabilities has represented, in the last years, one of the main frontiers in Artificial Intelligence (AI) research. Researcher from various disciplines debate about whether or not, among the capabilities of LLMs, there is the one of using knowledge about knowledge – usually considered one of the antechambers of meta-cognition in cognitive agents – about a particular task in order to improve or self-correct previous errors. In this work we propose a novel fine-tuning approach for LLMs, named EXAR, based on a multi-stage process leveraging past predictions from an early version of the same, and aimed at injecting metacognitive features for the task of Question-Answering. The conducted experiments on Llama-2-7B-chat showed promising improvements on the quality of the outcomes, due to the fact that the LLM acquired the ability to detect its own wrong predictions forcing itself to repeat submissions, thorough a prompt designed to fix inadmissible predictions, whenever detected. Such detection is achieved by enquiring the same LLM acting as meta-validator, through another prompt specifically designed for such purpose.

In June 17-19, 2024, the Conference Advance in Cognitive Systems held in Palermo, where, together with Misael Mongiovì, Luana Bulla and Antonio Lieto, I submitted a paper titled: "Eliciting Metaknowledge in Large Language Models". The contribution investigated on improving predictions on Large Language Models, through a self-corrective inference taking in account of past inferences, which can be considered also kind of Metacognition. The work was accepted as poster and presented by Antonio Lieto. 

Paper: https://www.dropbox.com/scl/fi/n6supie11h6qkbsj7eb48/COGSYS_24.pdf?rlkey=crrq7802f1rsfqjb7tgx8hydz&dl=0

Poster: https://www.dropbox.com/scl/fi/zs7ejw6o9cb3z3m9owxu3/POSTER_ACS.pdf?rlkey=4t9qsdsj7tnfurbykg8qq3s2b&dl=0

In the 13th International Conference on Data Science, Techonology and Applications, the organization impressed me for the capillary communication for every aspect of the conference, the keynotes were interesting and I met several scholars with whom there could be future collaborations. For the occasion I presented an article on hierarchical classification of texts, in the absence of training datasets, using the Large Language Model Llama2 for the generation of synthetic texts.

https://www.scitepress.org/Papers/2024/127907/127907.pdf

14th International Conference on Formal Ontology in Information Systems