hotpotqa

The official implementation of ICLR 2020, "Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answering".

Codes for NAACL 2021 Paper "Unsupervised Multi-hop Question Answering by Question Generation"

Implementation of Not All Contexts Are Equal: Teaching LLMs Credibility-aware Generation. Paper: https://arxiv.org/abs/2404.06809

Pytorch implementation of "A Simple Yet Strong Pipeline for HotpotQA" (Groeneveld, D., Khot, T., & Sabharwal, A.). Now developing!

RAG Project for COMP5423 of PolyU 25Fall

Ablation search over 55 agentic-RAG configs on HotpotQA, scored on EM/F1, retrieval recall and groundedness

This codebase implements a Retrieval-Augmented Generation (RAG) chatbot using the Gemini API and DSPy framework, designed to answer questions based on the HotPotQA dataset. It includes components for loading data, generating responses, and evaluating model performance through various QA strategies, including basic QA and multi-hop retrieval.

Hallucination-resistant multi-hop QA using hybrid BM25+FAISS retrieval, cross-encoder reranking, citation selection, and NLI-based verification. Evaluated on HotpotQA (7,405 examples), zero-shot.

Enhancing Retrieval-Augmented Generation with Document Link Structure for Multi-hop Web Question Answering

Component ablation study of an agentic RAG pipeline for multi-hop QA on 5K HotpotQA distractor questions. Evaluates query decomposition, iterative retrieval, Qdrant dense/sparse/RRF retrieval, cross-encoder reranking and adaptive routing. Finding: fixed hybrid retrieval outperformed adaptive routing.

Episodic Distillation for Verified RAG using Google Cloud + MongoDB Atlas

[EMNLP 2026] Official implementation of SAFE: An LLM-as-Verifier Framework for Evidence-Grounded Multi-Hop Reasoning.

프롬프트는 컴파일하고, 컨텍스트는 재사용하고, 틀린 부분만 고치는 RAG 파이프라인. 세 최적화의 상호작용을 8조합으로 측정

Agents Don't Always Do What They Think: Measuring Faithfulness in Multi-Step ReAct Agents

Implementation of concepts of NLP

LangGraph supervisor with parallel specialist agents. 1.99x speedup vs sequential multi-agent, reproducible in 6s with no API key. Planner, retrievers, analyzers, verifiers, synthesizer.

Multi-agent system for multi-hop information retrieval: query decomposition, tool-augmented retrieval, knowledge-graph traversal, and self-reflective validation. IR course project, Università di Napoli Federico II.

Corpus-level evaluation of diversity-aware retrieval for RAG on BEIR HotpotQA using BM25, DPR, Contriever, ColBERTv2, MMR, clustering, and DPP.

Exploration of retrieval methods on the HotpotQA corpus, combining dense retrieval and feature-based reranking. Achieved a mean nDCG@10 of 0.9416 using LambdaRank with features such as cross-encoder score, LLM score, BM25 score, and token-based statistics—surpassing dense retriever + cross-encoder baselines.

A two-stage query-aware context compression framework for RAG (QUITO-X + EP-EXIT) that cuts input tokens by 77% while achieving the best Token F1 among evaluated baselines on HotpotQA