Agent Era Information Retrieval: A Denoising-First Perspective (SIGIR 2026 Review)

01 Core Narrative: The Bottleneck in IR Keeps Shifting

The review frames the history of information retrieval (IR) as a series of migrating bottlenecks, culminating in the fourth era where the "consumer" of retrieval results changes from humans to large language models (LLMs). This creates tension: traditional IR optimizes for high recall, but LLMs are sensitive to the noise that high recall introduces.

When the "consumer" of retrieved results shifts from humans to LLM agents, the core optimization goal of IR migrates from recall to denoising.

The authors identify three intrinsic fragilities of the fourth era: fragmentation, context dilution, and cascading failures.

02 Controlled Experiments: Noise Impact on Generation Quality

A reproducible experiment uses LLaMA‑2‑7B‑Chat on 500 Natural Questions, each paired with 100 DPR passages labeled as gold or noise. By varying the signal‑to‑noise ratio, the study shows that increasing noise dramatically degrades exact match (EM) scores, far more than merely moving gold passages.

Core observation: under this setup, noise‑induced degradation is about five times larger than positional perturbation.

03 Attribution Framework: Three Entry Points for Noise

Corpus‑level noise : polluted indexes, duplicated or outdated content, and the influx of AI‑generated text that can cause model collapse (Shumailov et 2024).

Retriever‑level noise : hard distractors—highly relevant‑looking passages that do not support the answer, especially problematic for dense retrievers.

Context‑construction noise : the final prompt assembly amplifies noise via ranking‑attention misalignment, contradictory evidence, and indirect prompt injection (Greshake et al., 2023).

These layers can cascade, prompting the authors to redefine evidence assembly as an active denoising stage.

04 Methodology: Five‑Stage Denoising Pipeline

The core contribution organizes existing denoising techniques into five stages aligned with the information flow lifecycle:

§3.1 Controlled Indexing

Source credibility layering (e.g., timestamps, C2PA signatures, AI‑watermarks).

Quality filtering with MinHash, SemDeDup, and cleaning pipelines from RefinedWeb/FineWeb/Dolma.

Temporal management via VersionRAG or EraRAG.

Structural defenses using graph‑based RAG (GraphRAG, HippoRAG, RAPTOR).

§3.2 Robust Retrieval

Emphasizes that matching relevant documents is itself a denoising step; in the LLM era, precision and robustness must outrank raw recall.

§3.3 Context Assembly

Models the interaction between retriever and generator, aiming to maximize information density within context windows.

§3.4 Retrieval Verification

Introduces measurable verification metrics and feedback loops to audit and refine upstream components.

§3.5 Closed‑Loop Training

Looped orchestration (IRCoT, ChainRAG) to mitigate "lost‑in‑retrieval".

End‑to‑end policy learning (Search‑R1, Toolformer) that embeds denoising decisions in model weights.

Self‑evolution (Reflexion, MemGPT, AutoRAG, DSPy) for continual improvement.

05 Application Scenarios: Domain‑Specific Noise Manifestations

The review summarizes four high‑retrieval‑dependence domains, highlighting characteristic failure signatures and corresponding denoising recipes (see original Table 1).

Conclusion

In the LLM era, a retrieval system’s key duty is to act as a noise gate—expanding recall is easy, but controlling the quality of evidence fed to the model is far more challenging.

Paper Title: LLM‑Oriented Information Retrieval: A Denoising‑First Perspective
Authors: Lu Dai, Liang Sun, Fanpu Cao, Ziyang Rao, Cehao Yang, Hao Liu, Hui Xiong
Institution: Hong Kong University of Science and Technology / HKUST (Guangzhou)
Conference: SIGIR 2026
Link: https://www.alphaxiv.org/abs/2605.00505