CONTEXT-AWARE ADAPTATION OF GENERATIVE LLM RESPONSES
DOI:
https://doi.org/10.34185/1991-7848.itmm.2025.01.089Keywords:
Context-aware generation, Large Language Models, Adaptive alignment, Neuron modulation, Crisis informatics, HumAID dataset, Layerwise intervention, Transformer architectures, Model behavior adaptationAbstract
In crisis situations, the speed of response becomes critically important. Large Language Models (LLMs) are capable of generating useful recommendations; however, their default behavior often fails to account for the specific context of emergency events. This paper proposes the Adaptive Injectable Realignment Model (Adaptive IRM) as a method for achieving context-aware response generation. Adaptive IRM is a lightweight neural module that integrates into the LLM’s forward pass and injects contextual signals to adjust the model’s internal representations without modifying its original weights. Our approach is focused on disaster scenarios, using N neurons (in our case, N = 4) corresponding to natural hazards such as earthquakes, floods, fires, and hurricanes. The output signals from the Adaptive IRM, injected at various transformer layers, modulate attention mechanisms to emphasize information relevant to the given context. The paper outlines the IRM architecture, describes the proposed extensions, and presents a plan for using the HumAID dataset to train the Adaptive IRM. Experimental results are not yet available; instead, the concept, motivation, and implementation strategy of the proposed system are discussed.
References
Smith, B., Baker, D., Chase, C., et al. The Mysterious Case of Neuron 1512: Injectable Realignment Architectures Reveal Internal Characteristics of Meta’s Llama 2 Model // arXiv [preprint]. – 2024. – ID: arXiv:2407.03621. – Available at: https://arxiv.org/abs/2407.03621 (Accessed: [date not specified]).
Dathathri, S., et al. Plug and Play Language Models: A Simple Approach to Controlled Text Generation // Proceedings of the 8th International Conference on Learning Representations (ICLR). – 2020.
Alam, F., Qazi, U., Imran, M., Ofli, F. HumAID: Human-Annotated Disaster Incidents Data from Twitter with Deep Learning Benchmarks // Proceedings of the ICWSM 2021 Conference. – 2021.
Otal, H. T., Canbaz, M. A. LLM-Assisted Crisis Management: Building Advanced LLM Platforms for Effective Emergency Response and Public Collaboration // arXiv [preprint]. – 2024. – ID: arXiv:2402.10908. – Available at: https://arxiv.org/abs/2402.10908 (Accessed: [date not specified]).
DRAGN Labs. Injectable Alignment Model – README [GitHub Repository, electronic resource]. – 2024. – Available at: https://github.com/DRAGNLabs/InjectableAlignmentModel (Accessed: [date not specified]).
