TL;DR
Dense retrieval models suffer from a "search-by-answer" bias, where they favor passages containing answers they already "know" from pretraining over passages with valid alternative answers. This bias is so strong that models will often prefer a passage containing a known answer even if that passage expresses the completely wrong semantic relation.
What Problem Does This Paper Address?
Dense retrievers are the workhorses of modern search and Retrieval-Augmented Generation (RAG) systems. Ideally, they should perform semantic search: given a query like "What is the country of citizenship of Mukul Roy?", the model should look for documents expressing the citizenship relation.
However, these retrievers are built on top of Large Language Model (LLM) backbones that already contain vast amounts of world knowledge. The authors hypothesize that instead of searching for the requested relation, the retriever "pre-calculates" the most likely answer (e.g., "India") and then steers the search toward any document containing that answer.
This is problematic because it leads to retrieval bias:
- Stale Information: The model might favor a "known" but outdated answer over a "new" correct one in the corpus.
- Lack of Diversity: It may ignore valid but less common alternative answers.
- Relation Neglect: It might retrieve irrelevant documents just because they mention a familiar entity.
Methods
The researchers evaluated eight popular retrievers (including DPR, Contriever, E5, and Qwen3-Embedding) using a two-step approach:
- Object-Knowledge Diagnostic: They first measured how much the model's internal backbone "prefers" a specific answer. Using the model's original token-prediction head, they calculated the probability difference between a "canonical" answer (the one the model likely knows) and an "alternative" answer for a given subject-relation prompt.
- Controlled Passage Contrasts: They created two experimental settings to see if this internal preference predicted retrieval behavior:
- Object Contrast: Two passages with the same relation but different objects (e.g., "Mukul Roy has citizenship in India" vs. "Mukul Roy has citizenship in Bulgaria").
- Relation-Object Conflict: A "tug-of-war" between relation matching and answer matching. One passage has the correct relation but the "wrong" (alternative) answer, while the other has a "wrong" relation but the "right" (known) answer.
Datasets
- LAMA/TREx: Used to construct the controlled factual triples (subject, relation, object) for the contrastive experiments.
- QAMPARI: A natural multi-answer retrieval benchmark used to validate the findings in a more realistic setting where questions have multiple valid answers.
Key Insights and Ideas
- The Bias is Pervasive: Every single retriever tested showed a positive correlation between its internal knowledge of an answer and its tendency to retrieve passages containing that answer.
- Knowledge Overrides Semantics: In the "Relation-Object Conflict" setting, models frequently ranked a passage with the wrong relation higher simply because it contained the "known" object. For example, E5-large preferred the wrong-relation passage over 56% of the time.
- Scaling Increases Bias: Perhaps most surprisingly, as models get larger and more "knowledgeable," the search-by-answer bias tends to get stronger. The Qwen3-8B model showed some of the highest bias scores, suggesting that more powerful backbones are more likely to "hallucinate" their preferred answers into the retrieval process.
Key Results
- Object Contrast Preference: Models preferred the "known" answer passage significantly more than the 50% baseline. Qwen3-8B reached a staggering 94.6% preference rate for known objects.
- Correlation: Pearson correlations between the internal knowledge diagnostic and retrieval score shifts ranged from 0.21 to 0.52.
- Natural Setting Validation: On the QAMPARI dataset, "known" answers (those ranked in the top 5 by the model's internal head) were retrieved much more successfully than "unknown" answers (ranked >20), even when both were equally valid in the corpus.
Limitations and Future Work
The authors acknowledge several limitations:
- Template-Based Passages: The primary experiments used synthetic, template-generated sentences to isolate variables. Real-world text is noisier and more complex.
- Single-Token Focus: The diagnostic was restricted to single-token answers to avoid tokenization artifacts, though the authors noted anecdotal evidence that the bias exists for multi-token entities as well.
- Alternative Relations: The "conflict" setting relied on manually specified alternative relations, which might not capture the full spectrum of semantic confusion in the wild.
Future work could explore ways to de-bias retrievers, perhaps by training them to ignore their internal priors or by using architectures that more strictly separate "world knowledge" from "retrieval logic."
Summary updated