AI Science Tracker
EN

IBM

1 discovery from 2026

Discoveries

IBM

Retrieval improvements shrank when combined in a RAG pipeline

Source check: Source checked. Not independent scientific confirmation.

Finding: Five changes to a retrieval-augmented generation pipeline each improved document ranking when tested alone. Adding their individual gains gave an expected score of 0.667, but the combined system scored 0.483 against a baseline of 0.412. The result concerns the tested retrieval components and queries.

Technology: HyDE query augmentation, cross-encoder reranking

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

AI contribution
Learned query augmentation and cross-encoder reranking change query–document rankings; matched component perturbations test whether those changes add together.
Test and result
A matched component-composition experiment compared individual changes and the combined pipeline using NDCG@10, with a paired bootstrap over 93 queries.
Limitation / still unknown
The result does not establish that every retrieval pipeline shows the same interaction. The paper’s answer-faithfulness comparison used different component configurations and did not directly establish additivity.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.29722 · https://arxiv.org/html/2610.29722v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: arXiv v1 front matter, §§4–5, Table 6, §D.14/Figure 9 and limitations on faithfulness additivity.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open the interactive profile for the timeline, or browse all organizations and model families.