Research & Projects

The objective of this research is to design and develop behavior-driven LLM agents that replicate human thought patterns and reasoning. By systematically extracting and analyzing behavioral traits, decision-making styles, and thought processes from public interviews and other data sources, the study aims to encode these insights into LLM agents. This will enable the agents to simulate the unique reasoning and thinking styles of specific individuals in a contextually accurate manner.

This research investigates the use of Retrieval-Augmented Generation (RAG) to improve legal Natural Language Processing (NLP) by fine-tuning Large Language Models (LLMs) in legal domain. The study shows that fine-tuned LLMs, when combined with advanced RAG techniques, lead to a 13% improvement in Hit Rate and a 15% increase in Mean Reciprocal Rank (MRR) compared to baseline LLMs.

Furthermore, the proposed Legal Query RAG (LQ-RAG) architecture, which incorporates a feedback loop, outperforms both naive RAG configurations and RAG with fine-tuned models. It achieves a 23% improvement in relevance score over the naive configuration and a 14% improvement over the RAG with Fine-Tuned LLMs. These results underscore the potential of RAG and fine-tuned LLMs to advance legal NLP applications.

• More detail will be published soon.

This research focuses on enhancing sentiment analysis for cryptocurrency markets, where social media sentiments significantly influence market dynamics. Fine-tuning large language models (LLMs) improved zero-shot performance by 40%, with larger models like FLAN-T5 achieving up to 75.16% accuracy. Optimal results were found with 6,000 data points, while MiniLM proved efficient for smaller datasets.

Clear, concise instructions outperformed complex ones by 12%, showcasing the importance of simplicity in model training. These findings pave the way for more effective sentiment analysis tools, unlocking smarter insights for navigating dynamic cryptocurrency trends.

• For more detail please click here.