Which Meaning-Inverting Edits Can a Retriever See, and Why? A Mechanism-Level Failure Map for Text Retrieval Architectures
That single-vector dense retrievers struggle with meaning-inverting edits such as negation is established. We ask two questions that result leaves open: which inverting edits each retrieval architecture can see, and why single-vector models fail: is the distinction encoded but unranked by cosine similarity, or absent from the representation? We build symmetric minimal-pair contrast sets over four axes (negation, role reversal, numeric, and a meaning-preserving paraphrase control; 388 validated items) and score 17 systems (BM25, eleven single-vector encoders including a within-family scale sweep, late interaction, cross-encoders and an LLM reranker) under one pairwise protocol. Dense paired accuracy is 0.18 on role reversal (below the 0.25 chance level) and 0.52 on negation, against 0.93 on the control; token-level architectures recover part of the gap, but not in the monotone order we pre-registered. A linear probe on frozen embeddings, scored in the same within-item ranking mode as cosine (a correction to our pre-registered, metric-mismatched probe), gives an axis-dependent answer: on negation it reaches 0.81, so the distinction is partly encoded but under-ranked; on role reversal it reaches only 0.53 against a measured 0.50 null, and a nonlinear read-out recovers no more. For argument order, the limit of single-vector retrieval looks representational, not metric. In an end-to-end RAG demonstration roughly 72% of retrieval errors become answer errors. The contrast sets, code and analysis are available from the authors on request. Preprint. Not peer reviewed.
Authors
- Khaled Shaheen
- Faisal Shaheen (ORCID: https://orcid.org/0009-0006-8905-6765)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24270177
- Primary Topic
- Information Retrieval and Search Behavior
- Type
- preprint