Closing the Retriever Gap in Agentic Search Systems, Offline Negative Item Filtering at Scale, and More!

2025-01-31 · Source: Top Information Retrieval Papers of the Week · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Expert, long

Summary

This week's research highlights significant advancements in information retrieval, focusing on Retrieval-Augmented Generation (RAG) systems and recommender systems. Ant Group's Deep GraphRAG introduces a hierarchical retrieval and adaptive integration method, achieving 94% of a 72B model's performance with a 1.5B parameter model on Natural Questions. Alibaba's CoNRec improves negative feedback modeling in recommendation systems, showing a 24.9% gain on long-tail items on Taobao data. Shi et al. present RT-RAG for multi-hop question answering, using tree-structured reasoning to prevent error propagation. Spišák et al. enhance collaborative filtering with sparse autoencoders for interpretability and steerability. ByteDance's HyFormer unifies sequence modeling and feature interaction for CTR prediction, outperforming baselines on billion-scale datasets. Jiao et al.'s PruneRAG boosts RAG efficiency for multi-hop QA, achieving higher F1 scores and running 4.9x faster. Liu et al.'s Agentic-R optimizes retrievers for agentic search, outperforming general-purpose retrievers across seven QA benchmarks. The University of Glasgow formalizes RPP and GPP tasks for RAG, combining QPP and perplexity-based predictors. TU Delft's PopSteer uses sparse autoencoders to interpret and mitigate popularity bias in recommenders, improving fairness while maintaining accuracy. Fan et al.'s Rank4Gen introduces a generator-aware document ranking model for RAG, showing consistent improvements across generators. Additionally, three new tools are introduced: RAGExplorer for visual analytics of RAG systems, Docs2Synth for synthetic data training of visual retrievers, and SearchGym for simulating real-world search environments for agent training.

Key takeaway

For AI Engineers building advanced RAG or recommender systems, these papers offer concrete strategies to enhance performance and interpretability. Consider integrating hierarchical retrieval and adaptive reward mechanisms from Deep GraphRAG to improve efficiency. If you are addressing bias in recommendation, explore PopSteer's neuron steering for interpretable debiasing. For multi-hop QA, RT-RAG and PruneRAG provide robust frameworks to prevent error propagation and boost efficiency.

Key insights

Recent advances enhance RAG and recommender systems through hierarchical retrieval, negative feedback modeling, and interpretable bias control.

Principles

• Hierarchical retrieval improves efficiency and context.
• Adaptive reward weighting prevents metric over-optimization.
• Generator-aware ranking optimizes RAG performance.

Method

Deep GraphRAG uses a three-stage hierarchical retrieval with beam search and DW-GRPO for adaptive reward rebalancing. CoNRec employs RQ-VAE for semantic IDs and Progressive GRPO for negative feedback modeling. PopSteer uses sparse autoencoders to identify and steer popularity-aligned neurons.

In practice

• Use Deep GraphRAG for efficient graph-based RAG.
• Implement CoNRec for better negative item filtering.
• Apply PopSteer to mitigate popularity bias in recommenders.

Topics

• Retrieval-Augmented Generation
• Recommender Systems
• Sparse Autoencoders
• Agentic Search
• CTR Prediction

Code references

• sakura20221/RT-RAG
• Fdioa/PruneRAG
• 8421BCD/Agentic-R
• parepic/PopSteer2.0
• JOHNNY-fans/Rank4Gen

Best for: AI Engineer, NLP Engineer, Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Deep Research for Recommender Systems, Improving Search Agent with One Line of Code, and More! — Recent research focuses on enhancing information retrieval through agent reasoning, semantic search, and RAG systems.
• Agent-Orchestrated Adaptive RAG: A Comparative Study on Structured and Multi-Hop Retrieval — Agentic RAG enhancements are domain-dependent and cost-intensive, necessitating selective, adaptive orchestration.
• A Late Chunking Approach for Visual Documents, Does Agentic Search Make GraphRAG Obsolete? and More! — New research enhances LLM retrieval, recommendation systems, and RAG through novel architectures and training methods.
• Retrieval-Augmented Generation Must Move Beyond Factual Grounding to Represent Diverse Opinions — Current RAG systems exhibit factual bias, necessitating Opinion-Aware RAG to represent diverse perspectives by preserving aleatoric uncertainty.
• Rethinking the Necessity of Adaptive Retrieval-Augmented Generation through the Lens of Adaptive Listwise Ranking — Adaptive RAG's role shifts from noise mitigation for weaker LLMs to efficiency optimization for stronger ones.
• The Real Leap Is Not Better Retrieval. It’s Better Control Over Retrieval. — Agentic RAG shifts focus from static retrieval to dynamic control over the search process.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Top Information Retrieval Papers of the Week.