Abstract:Large Language Models (LLMs) have the potential for substantial common sense reasoning. However, these capabilities are often emergent in larger models. This means smaller models that can be run locally are less helpful and capable with respect to certain reasoning tasks. To meet our problem space requirements, we fine-tune smaller LLMs to disaster domains, as these domains involve complex and low-frequency physical common sense knowledge. We introduce a pipeline to create Field Ready Instruction Decoding Agent (FRIDA) models, where domain experts and linguists combine their knowledge to make high-quality seed data that is used to generate synthetic data for fine-tuning. We create a set of 130 seed instructions for synthetic generation, a synthetic dataset of 25000 instructions, and 119 evaluation instructions relating to both general and earthquake-specific object affordances. We fine-tune several LLaMa and Mistral instruction-tuned models and find that FRIDA models outperform their base models at a variety of sizes. We then run an ablation study to understand which kinds of synthetic data most affect performance and find that training physical state and object function common sense knowledge alone improves over FRIDA models trained on all data. We conclude that the FRIDA pipeline is capable of instilling general common sense, but needs to be augmented with information retrieval for specific domain knowledge.
Abstract:Large Language Models, despite their significant capabilities, are known to fail in surprising and unpredictable ways. Evaluating their true `understanding' of language is particularly challenging due to the extensive web-scale data they are trained on. Therefore, we construct an evaluation to systematically assess natural language understanding (NLU) in LLMs by leveraging Construction Grammar (CxG), which provides insights into the meaning captured by linguistic elements known as constructions (Cxns). CxG is well-suited for this purpose because provides a theoretical basis to construct targeted evaluation sets. These datasets are carefully constructed to include examples which are unlikely to appear in pre-training data, yet intuitive and easy for humans to understand, enabling a more targeted and reliable assessment. Our experiments focus on downstream natural language inference and reasoning tasks by comparing LLMs' understanding of the underlying meanings communicated through 8 unique Cxns with that of humans. The results show that while LLMs demonstrate some knowledge of constructional information, even the latest models including GPT-o1 struggle with abstract meanings conveyed by these Cxns, as demonstrated in cases where test sentences are dissimilar to their pre-training data. We argue that such cases provide a more accurate test of true language understanding, highlighting key limitations in LLMs' semantic capabilities. We make our novel dataset and associated experimental data including prompts and model responses publicly available.