multi-hop-qa

Asking for What Was Never Requested: horizontal and vertical proactivity in LLM agents. Need-graph metrics (no LLM judge) and Q&D, which trains a questioner from the consequences of its questions.

EvLink: Source-grounded evidence linking for Graph RAG

7-layer RAG framework that eliminates semantic drift + context poisoning. Faithfulness 0.94 vs 0.71 Naive RAG. +13.6 EM on multi-hop QA.

Evaluation code for studying how upstream ASR errors affect multi-hop retrieval-augmented generation. EMNLP 2026 Main Conference.

Index-time reasoning for multi-hop RAG. Bridging facts precomputed during indexing enable single-pass retrieval and a single LLM call at inference. AACL-IJCNLP 2026 Findings.

When is enough evidence enough? Sufficiency judgments and stopping rules in multi-step retrieval agents. Technical report, code and full logs.

Binary evidence-sufficiency dissociation in reasoning-model hidden states for fixed-question, changed-context multi-hop QA.

Multi-turn retrieval agent post-training with Qwen3-4B, BM25 SearchTool, QLoRA SFT, and paired behavior evaluation on HotpotQA.

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

On-policy state training for chain dense retrieval (arXiv:2610.40409): code and the HotpotQA-Long / 2Wiki-Long benchmarks

Knowledge-graph RAG library for multi-hop QA — iterative retrieve-and-reason (IRCoT), graph retrieval, and cross-document enrichment at ingest. Production retrieval pipeline in Python.

What does agentic RAG actually cost? A measured cost-per-correct-answer benchmark of agentic RAG, iterative RAG, cross-encoder reranking and naive RAG on multi-hop QA (MuSiQue). Meters every LLM call.

HippoRAG 2 (graph) vs. ChromaDB (vector) retrieval, benchmarked head-to-head on 3 multi-hop QA datasets — real recall, significance tests, and cost

SAFE-KV (Findings of AACL-IJCNLP 2026): causal suffix repair makes non-prefix KV-cache reuse in RAG reproduce full-prefill answers. One-file implementation, plus the code, data, and per-example outputs behind every number in the paper.

A per-instance cost model for AI: compute and verification scale with the difficulty of one question.

A fine-tuned Qwen2.5-1.5B model that breaks complex multi-hop questions into atomic sub-queries with dependency graphs, doubling RAG retrieval accuracy in ~50ms

Evidence-based arbitration for multi-hop RAG/QA using FLARE, IRCoT, and FARR candidate trajectories.

Multi-hop graph + vector RAG over SEC 10-K/20-F filings. Neo4j knowledge graph + pgvector joined on chunk ids; the LLM never writes Cypher (enum router, template-only queries) and every claim carries a validated citation. Multi-hop accuracy 38% -> 63% vs vector-only RAG on a 64-question benchmark.

Graduated Action Routing for ReAct Agents: reproducible multi-hop QA routing experiments, measured inference costs, statistical analysis, and PAgE 2027 paper draft.

Fully local multi-hop RAG that cross-checks answers across news sources: Ollama, LanceDB, hybrid search, cross-encoder reranking, Llama 3.1 8B