Skip to content

ARK: Answer-Centric Retriever Tuning via KG-augmented Curriculum Learning

Conference: ACL 2026 arXiv: 2511.16326 Code: GitHub Area: Graph Learning Keywords: Answer-Centric Retrieval, Knowledge Graph Augmentation, Curriculum Learning, Contrastive Learning, Long-Context RAG

TL;DR

ARK filters positive samples through three-dimensional answer sufficiency scoring (Forward + Backward + Retriever alignment) and generates progressively difficult hard negatives via LLM-constructed knowledge graphs for curriculum contrastive learning, averaging +14.5% F1 across 10 datasets.

Background & Motivation

Key Challenge: The gap between retriever training objective (query-document similarity) and RAG's ultimate goal (generating correct answers).

Core Idea: Use KG subgraph-generated augmented queries to mine progressively difficult hard negatives through curriculum contrastive learning, teaching the retriever to distinguish "sufficient" from "seemingly relevant but insufficient" evidence.

Method

Key Designs

  1. Three-Dimensional Answer Sufficiency Scoring: Forward alignment \(S_f\) = whether a chunk suffices to generate the answer; Backward alignment \(S_b\) = whether the question can be reconstructed from answer + chunk; Parameter alignment \(S_v\) = original retriever cosine similarity.

  2. KG-Driven Hard Negative Mining: Large subgraph (\(Q_L^{aug}\)) generates easier negatives; small subgraph (\(Q_S^{aug}\)) generates harder negatives — more focused subgraphs produce queries closer to the correct answer's "semantic neighborhood."

  3. Curriculum Contrastive Learning: Three-stage curriculum progressing from in-batch random negatives to hard negatives from \(Q_L^{aug}\) then \(Q_S^{aug}\).

Key Experimental Results

  • Average +14.5% F1 across 10 datasets
  • SOTA on 8/10 datasets (Ultradomain + LongBench)

Highlights & Insights

  • Redefines KG's role in RAG — from "retrieval index" to "training signal generator" — drastically reducing KG usage cost
  • Plug-and-play without changing retriever architecture

Rating

  • Novelty: ⭐⭐⭐⭐
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐
  • Writing Quality: ⭐⭐⭐⭐
  • Value: ⭐⭐⭐⭐⭐