ETD

• AI
• artificial intelligence
• AwareBench
• awareness
• AwareXtend
• collaboration
• large language models
• LLMs

Large Language Models (LLMs) have achieved strong results in reasoning and text generation, but they continue to struggle with contextual sensitivity, cultural adaptability, and alignment with human intentions. This thesis investigates awareness as a complementary skill to factual accuracy, formalized along five dimensions: capability, mission, emotion, culture, and perspective. To enhance awareness in LLMs, two collaborative frameworks are proposed. The first, Hierarchical Answer
Aggregation (HA), integrates the outputs of multiple models by weighting their answers according to each agent dimension-specific competence. The second, Peer Debate (PD), allows models to iteratively exchange, critique, and refine their answers. Experiments are conducted on the AwareBench dataset and on the newly developed AwareXtend benchmark, designed to provide a more demanding awareness evaluation to challenge state-of-the-art models. Results show that collaboration improves awareness beyond what individual models can achieve: HA yields consistent gains, especially in capability and culture awareness, while PD provides broader and more generalisable improvements across dimensions. These findings indicate that collaborative strategies can mitigate the limitations of single LLMs and advance the development of systems that are more socially and contextually aware.