New DEO Method Fixes AI Retrieval Failures in Negation Queries
Researchers have introduced Direct Embedding Optimization (DEO) to solve how AI fails at exclusion queries. Learn how this training-free approach works.
Source: Grafy Research LabRetrieval-Augmented Generation (RAG) systems frequently struggle to respect user instructions that involve negation or exclusion. Researchers have developed a new method called Direct Embedding Optimization (DEO) that allows AI to accurately retrieve content by explicitly separating what a user wants to include from what they want to exclude.
Why do AI retrievers struggle with negation?
Standard dense retrieval methods encode queries and documents into low-dimensional vectors to find semantic similarity. However, according to the research paper by Taegyeong Lee, Jiwon Park, Seunghyun Hwang, and JooYoung Jang, these systems often fail when a query includes phrases like "not," "exclude," or "without." This happens because standard models focus on the presence of semantic concepts rather than the logical relationship between them. For instance, a query for "earnings forecast without 2024 results" might still retrieve 2024 data because the model identifies the strong semantic link between the query and the 2024 documents, failing to prioritize the "without" constraint.
How does Direct Embedding Optimization work?
Existing solutions to this problem typically involve fine-tuning embedding models or using complex adaptation techniques, which require massive datasets and significant GPU resources. DEO offers a training-free alternative by following a three-step process.
First, the method uses a Large Language Model (LLM) to decompose a single user query into two distinct parts: positive sub-queries (what to look for) and negative sub-queries (what to avoid). For example, a request to "show latest earnings but exclude 2024" is broken down into positive components like "earnings forecast for 2025" and negative components like "2024 financial report."
Second, the system embeds both the original query and these decomposed sub-queries using a pre-trained model. It then performs direct optimization of the original query embedding using a contrastive loss function. This mathematical process pulls the query embedding closer to the positive sub-query embeddings while pushing it further away from the negative ones.
Finally, the optimized embedding is used to perform the actual retrieval, resulting in results that more closely align with the user's specific constraints.
Does DEO work for images and video?
Yes, the researchers demonstrated that DEO is modality-agnostic, meaning it can be applied to both text and multimodal retrieval tasks. In experiments involving multimodal retrieval on the COCO-Neg benchmark, the authors found that DEO improved Recall@5 when using OpenAI CLIP. Specifically, the performance increased from 0.4792 to 0.5392. This suggests that the method can help vision-language models better handle attribute negation, such as "not red," or relational negation, such as "A is not to the left of B."
What are the performance gains of this method?
In text-based benchmarks, the improvements were significant. Using the BGE-large-en-v1.5 embedding model, the researchers reported that DEO improved the Mean Average Precision (MAP) on the NegConstraint benchmark from 0.6299 to 0.7327. Additionally, the Normalized Discounted Cumulative Gain at 10 (nDCG@10) rose from 0.7139 to 0.7877. These gains indicate that the method provides a more stable and precise way to handle complex user intent without the overhead of retraining entire models.
How can researchers apply these findings?
Because DEO does not require additional training data or model updates, it is particularly useful for developers working in resource-constrained environments or those using closed-source embedding models where fine-tuning is impossible. The method's ability to work with existing pre-trained models makes it a highly practical tool for improving RAG pipelines in real-world applications.
For researchers looking to dive deeper into complex academic papers or technical documentation regarding new AI architectures, Grafy Reader can be used to search through and question dense PDFs, helping to parse new methodologies like DEO as they emerge in the field.
Frequently asked questions
Does DEO require training a new model?
No, DEO is a training-free method. Instead of updating the weights of an embedding model through expensive fine-tuning, it optimizes the query embedding itself using a contrastive objective based on decomposed positive and negative sub-queries.
Can DEO be used for image retrieval?
Yes, the method is multimodal. The researchers demonstrated its effectiveness in multimodal settings, showing that it can improve the performance of vision-language models like OpenAI CLIP when dealing with negated descriptions in image and video retrieval tasks.
What is the main advantage of DEO over fine-tuning?
DEO avoids the high computational costs and the need for large-scale datasets required by traditional fine-tuning. It provides a lightweight, practical way to improve negation-aware retrieval without the complexity of model adaptation or deployment hurdles associated with new model versions.
Source: Grafy Research Lab — DEO: Training-Free Direct Embedding Optimization for Negation-Aware Retrieval


