Anagrafe della ricerca

Large Language Models (LLMs) are one of the most disruptive technologies of the last years as they are revolutionizing information access, creativity, and text processing tasks. Beyond understanding and generating text in natural language, recently proposed LLMs also support visual content as part of the instruction prompts or responses. How ever, they still show limited capabilities to process time series data, understand temporal relations, and effectively handle time-evolving scenarios. Making LLMs fully aware of the concept of time is of primary importance because the process of knowledge discovery from data is inherently time-dependent and the amount of real-world numeric time series, acquired for instance from sensors or measurement instruments, is ever-increasing. The TIME-AWARE-LLM project (TA-LLM in short) aims at addressing the current limitations of LLMs while coping with time series data and temporal aspects. The project goal is threefold: 1. Allow LLMs to effectively and efficiently cope with time series data. We investigate how LLMs can leverage the knowledge hidden in timestamped series of numeric values. To this end, TA-LLM envisages the design and testing of new LLM-based approaches based on in-context learning and parameter-efficient or lightweight fine-tuning. 2. Mitigate the effects of multimodal content misalignment in LLM data sources. We study to what extent LLMs’ performance is influenced by the temporal misalignment of the input data available in different modalities. Specifically, in several multimodal data such as educational videos, news streams, and audio podcasts, the visual, textual, and acoustic sources are likely to be not fully aligned or partially misaligned. TA-LLM aims at first detecting LLMs’ challenges related to temporal content misalignment and then proposing ad hoc mitigation strategies. 3. Propose Time-Aware LLM-based Approaches. We design and test new LLM-based time-evolving approaches to address common LLM tasks such as data summarization, question answering, and intent recognition. To make LLM-based systems robust to time-evolving scenarios TA-LLM studies, develops, and tests reactive systems to detect temporal drifts or inconsistencies in LLM outputs and and promptly implement corrective actions.