Srijan Shakya has successfully completed his master’s thesis, “Enhancing LLMs Reasoning Capabilities by Including Retrieval.” The work was supervised by Univ.-Prof. Dr. Sepp Hochreiter and Dr. Korbinian Pöppel, MSc.
In this thesis, the reliability of large language models (LLMs) in complex reasoning tasks was investigated by treating retrieval as a form of dynamic in-context learning. The research evaluates an adaptive retrieval-augmented architecture in which the LLM agent actively decides when to query an external knowledge base depending on the perceived difficulty of a task.
Experiments conducted on the GSM8K and MATH-500 benchmarks show that this adaptive strategy significantly outperforms standard static retrieval approaches. In particular, the model achieved a +6.4 percentage point improvement in specialised mathematical reasoning tasks.
These findings suggest that the ability of an AI system to self-assess its knowledge and selectively incorporate external information – acting as a form of metacognitive signal – represents an important principle for building more robust and reliable generative models.
