RefineLoop RAG: Dynamic Query Refinement through Iterative
Evaluation by LLM
This report proposes a novel query optimization method for Retrieval-Augmented Generation (RAG) using iterative refinement and model-driven evaluation. The approach adjusts user queries based on generative and discriminative scoring of initial retrievals, leveraging few-shot and binary equivalence prompts with LLaMA and GPT models. Results show improved query precision and retrieval relevance, demonstrating the effectiveness of integrating retrieval and evaluation loops in knowledge-intensive tasks.
Babysitting a Small Language Model through One-Step Tree-of-Thoughts Knowledge Distillation
This paper introduces a One-Step Tree-of-Thought (ToT) prompting method with knowledge distillation to enhance Small Language Models (SLMs) using Large Language Models (LLMs). By prompting GPT-4o with a simplified ToT framework, we generate training data to fine-tune SmolLM-360M, enabling it to emulate ToT-style reasoning efficiently. Tested on the Game of 24 dataset, the approach achieves competitive reasoning performance to LLMs while retaining computational efficiency.
