New Studies Trace Long, Drifting Search Chains in Agentic RAG and Offer Architecture Fixes

A cluster of papers shows how repeated, interleaved retrieval calls create query-intent drift and proposes stateful, parallel and graph-based remedies that still need real-world validation.

Overview

Researchers published a cluster of papers on October 21–22 that together map how modern Retrieval-Augmented Generation (RAG) systems generate progressive sub-queries during reasoning and why that matters for QA performance.
The ASQ dataset collected about 1 million agentic search traces and found that single user queries can trigger extremely long multi-step retrieval chains, sometimes up to 186 calls, with large variation across agent and retrieval settings.
LASAGNA ran controlled sweeps of 450 pipeline configurations and showed that individually strong components often fail to add up in combination and that the best pipeline choices change as the document corpus grows.
Two proposed fixes report benchmark gains: MAPS reduces 'search momentum' by retaining retrieved evidence, refining a reasoning state, and parallelizing complementary sub-queries; ThreadRAG preserves context with a semantic-topology graph and energy-driven pathfinding and reports sizable EM and token-cost improvements.
All teams call for full-stack, large-scale validation because the reported gains come from controlled benchmarks and the agentic retrieval behaviors raise practical concerns about latency, token cost, and answer faithfulness in production systems.