← News

ResearchSources: arXiv · 20 Nov 2026

DEO offers training-free fix for negation-aware retrieval

A new arXiv paper proposes Direct Embedding Optimization, a method that handles negation and exclusion queries without fine-tuning or extra training data.

In plain words

Search systems often struggle when a query asks to exclude something, such as finding images without people. The proposed method, DEO, splits a query into what should match and what should not, then adjusts the query embedding directly. It needs no retraining or extra data, according to the paper's abstract.

What was announced

The authors, writing on arXiv, note that existing retrieval methods often fail on negation and exclusion queries, and that prior fixes rely on embedding adaptation or fine-tuning, which add computational cost and deployment complexity. DEO instead decomposes a query into positive and negative components and optimizes the query embedding with a contrastive objective, requiring no additional training data or model updates.

The reported results come from the paper's abstract alone. The authors state that DEO outperforms baselines on the NegConstraint benchmark, with gains of +0.0738 nDCG@10 and +0.1028 MAP@100, and improves Recall@5 by +6% over OpenAI CLIP in multimodal retrieval. They present this as evidence that DEO is practical for negation- and exclusion-aware retrieval in real-world settings. Because only the abstract is available, the benchmark setup, baselines and evaluation details cannot be verified from this source.

What it does not show

  • Only the paper's abstract was reviewed; the full methodology, benchmark construction and experimental details were not available for verification.
  • The reported gains are the authors' own claims and have not been independently confirmed.

For security and IT teams

Scope
DEO targets negation- and exclusion-aware text and multimodal retrieval, described as training-free, so it may be relevant where fine-tuning-based retrieval adaptation is too costly.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.