AI Agent Collection

OG-RAG is a novel method that enhances LLM responses by grounding retrieval processes in domain-specific ontologies, using hypergraphs to represent domain documents and a greedy algorithm for efficient context retrieval, improving factual accuracy and context attribution in specialized domains. OG-RAG constructs a hypergraph representation of domain documents, retrieves the minimal set of hyperedges, and enables efficient retrieval while preserving the complex relationships between entities, increasing the recall of accurate facts and improving response correctness across different LLMs.

This paper introduces a generative AI system for the legal domain that integrates Retrieval-Augmented Generation (RAG), Vector Stores (VS), and Knowledge Graphs (KG) constructed via Non-Negative Matrix Factorization (NMF) to enhance legal information retrieval, AI reasoning, and minimize hallucinations, demonstrating its application in legal document clustering, summarization, and cross-referencing tasks. The system leverages web scraping to collect legal texts and bridges the gap between keyword-based searches and contextual understanding by using advanced semantic representations and hierarchical relationships.

NitiBench is introduced as a new benchmark for Thai legal QA systems, evaluating RAG and long-context LLM approaches. The benchmark includes datasets for Thai financial and tax law, with tailored metrics for multi-label retrieval and end-to-end evaluation, to address the limitations of current Thai legal NLP solutions and provide a foundation for future research in the field. Codes and datasets are open-sourced to support fair evaluation and future research.

The s3 framework improves Retrieval-Augmented Generation (RAG) by training a search agent using reinforcement learning and a novel Gain Beyond RAG reward, achieving strong performance with less data compared to existing methods. It optimizes retrieval quality independently of the generator LLM, enhancing modularity and efficiency in RAG systems for both general and medical question answering tasks

This paper introduces EVO-RAG, a curriculum-guided reinforcement learning framework for efficient multi-hop retrieval-augmented generation, which evolves a query-rewriting agent from broad exploration to concise refinement, improving accuracy and reducing retrieval depth in multi-hop question answering tasks. EVO-RAG uses a multi-objective reward mechanism and dynamic reward scheduler to optimize the retrieval and generation processes.

This study evaluates the reasoning capabilities of Large Language Models (LLMs) in dynamic environments, testing prompting techniques like self-reflection and heuristic mutation, and finds that strategic prompting can improve smaller models' performance, but advanced reasoning methods can also introduce instability and performance drops, revealing limitations in planning and spatial coordination.

This paper introduces the concept of 'sufficient context' in Retrieval Augmented Generation (RAG) systems, uses an LLM-based autorater to classify context sufficiency, and explores methods to reduce hallucinations by leveraging sufficient context information for selective generation, improving accuracy by 2-10% for Gemini, GPT, and Gemma.

NodeRAG is a graph-centric framework that introduces heterogeneous graph structures for seamless integration of graph-based methodologies into the RAG workflow, improving question-answering performance and efficiency. It constructs a heterograph with distinct node types representing entities, relationships, and semantic units, enabling fine-grained retrieval and unified information retrieval across different levels of information.

D-FINE is a real-time object detector that redefines bounding box regression in DETR models using Fine-grained Distribution Refinement (FDR) and Global Optimal Localization Self-Distillation (GO-LSD), achieving state-of-the-art accuracy and efficiency on the COCO dataset. It iteratively refines probability distributions and transfers localization knowledge from deeper to shallower layers, balancing speed and accuracy.