💻 computer science

Do Neural Retrievers Prefer Certain Documents? Evidence of Learned Relevance Priors

This paper reveals that supervised neural retrievers implicitly learn and encode query-independent document relevance priors from biased annotation protocols, causing them to systematically favor comprehensive, mainstream content over niche or technical documents and creating a "findability gap" that hinders the retrieval of genuinely relevant but less favored materials.

Original authors: Francisco Valentini, Edgar Altszyler, Martin Fajcik

Published 2026-06-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Francisco Valentini, Edgar Altszyler, Martin Fajcik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a librarian to find books for you. You train this librarian by showing them thousands of examples of "good matches" (a question and the perfect book) and "bad matches" (a question and the wrong book). The goal is for the librarian to learn what makes a book relevant to a question.

This paper argues that our current "neural librarians" (AI search engines) are learning a secret, unfair trick. They aren't just learning what matches the question; they are also learning which types of books are most likely to be labeled as "good" by their trainers.

Here is the breakdown of what the researchers found, using simple analogies:

1. The "Favoritism" Bias

When humans create training data for these AI search engines, they don't label every single document in the world. They pick a subset.

The Problem: The humans tend to pick documents that are comprehensive, well-written, and about popular topics (like a full encyclopedia entry on "Climate Change"). They often skip documents that are short, technical, niche, or fragmented (like a quick forum post on "How to fix a specific error code" or a paragraph from the middle of a long article).
The Result: The AI learns a hidden rule: "If a document looks like a polished, comprehensive summary of a popular topic, it's probably relevant." The researchers call this a "Relevance Prior." It's a bias the AI picks up on its own, even though no one explicitly told it to prefer those documents.

2. The "Findability Gap"

Because the AI has this hidden rule, it creates a "findability gap."

The Analogy: Imagine two people are looking for a specific tool in a massive warehouse.
- Person A has a tool that looks like a shiny, new, branded box (a "high-prior" document). The AI librarian immediately spots it and puts it at the front of the line.
- Person B has the exact same tool, but it's wrapped in plain brown paper and looks a bit messy (a "low-prior" document). Even though it is the correct tool, the AI librarian ignores it or pushes it to the back of the line because it doesn't look like the "good" examples it was trained on.
The Finding: The paper shows that documents with these "messy" or "niche" features are systematically harder to find, even if they are genuinely the right answer.

3. The "Toy Experiment" Proof

To prove this wasn't just a coincidence, the researchers ran a controlled experiment.

The Setup: They took a bunch of documents and secretly added a hidden marker (a code like [X]) to half of the "good" documents and half of the "bad" documents. This marker had nothing to do with the actual content; it was just a random tag.
The Outcome: The AI learned to associate the [X] marker with "relevance." When they tested it later, the AI ranked documents with the [X] marker much higher, even when the marker was removed or when the document was actually irrelevant.
The Lesson: The AI is a "pattern matcher" that will latch onto any signal that correlates with being labeled "good," even if that signal is fake or irrelevant to the actual question.

4. What Does "Good" Look Like?

The researchers used AI to explain why certain documents get labeled as "good" and others don't. They found a clear pattern:

High-Prior (Easy to find): Documents that are like encyclopedia introductions. They are self-contained, explain the "big picture," cover mainstream topics, and are written in a polished, formal style.
Low-Prior (Hard to find): Documents that are like sticky notes or technical manuals. They might be short, jump straight into details without context, focus on very specific technical problems, or look like raw data.

5. Why This Matters

The paper concludes that supervised AI retrievers (the kind trained on human-labeled data) are not just learning "relevance." They are also learning the preferences of the people who labeled the data.

The Consequence: If you ask a question, the AI is more likely to show you a broad, well-written summary than a specific, practical guide, even if the specific guide is exactly what you needed.
The Comparison: Older search methods (like BM25) rely on word matching and don't have this specific "style bias." They are less consistent, but they don't systematically penalize "niche" documents in the same way the neural networks do.

In short: The AI has learned to judge a book by its cover, preferring the "shiny, comprehensive" covers because that's what its teachers showed it most often, leaving the "plain, practical" books hidden on the bottom shelf.

Technical Summary: "Do Neural Retrievers Prefer Certain Documents? Evidence of Learned Relevance Priors"

Problem Formulation
Standard information retrieval (IR) models are trained to estimate the posterior probability of relevance, $P(R | q, d)$ , given a query $q$ and a document $d$ . However, training data for supervised retrievers is constructed via annotation protocols that select only a subset of documents for labeling, often leaving many others unjudged. The authors hypothesize that this selection bias leads supervised bi-encoder retrievers to implicitly learn a document-level relevance prior, $P(R | d)$ . This prior is a query-independent signal encoded in the model's representation space, reflecting the likelihood of a document being deemed relevant based solely on its content properties (e.g., topic, style, format) rather than its actual compatibility with a specific query. The core concern is that in general-domain retrieval, these learned priors may create a "findability gap," where documents with low relevance priors are systematically harder to retrieve, even when they are genuinely relevant.

Methodology
The investigation proceeds through three main phases:

Toy Experiment (Causal Validation): The authors injected a spurious, query-independent signal (a text prefix [X]) into a fraction $M$ of relevant documents in the LoTTE dataset. They fine-tuned bi-encoders (E5-small and Qwen2.5-0.5B) on this biased data. They then trained a simple logistic regression classifier (the "prior model") on the frozen embeddings to recover the signal. This setup tested whether retrievers could learn and encode a known bias and whether this encoding affected document findability.
General-Domain Analysis: The study evaluated three state-of-the-art supervised bi-encoders (BGE, NV-Embed, GTE) across multiple IR benchmarks (e.g., FEVER, Climate-FEVER, MIRACL, Natural Questions).
- Prior Estimation: A binary classifier was trained on the frozen embeddings of a retriever (using its training set of relevant vs. hard-negative documents) to predict $P(R | d)$ .
- Generalization: The prior model's ability to distinguish relevant from unjudged documents in held-out datasets was measured using AUC.
- Consistency: Pearson correlations were computed between prior scores derived from different retrievers to determine if the learned priors were model-specific or shared.
- Findability Correlation: Document findability ( $F_R(d)$ ), defined as the average reciprocal rank over queries where the document is relevant, was correlated with the estimated prior scores.
Controlled and Explanatory Analysis:
- Matched Comparisons: To isolate the prior effect from confounding factors (e.g., document length, named entity density), the authors created matched pairs of high-prior and low-prior documents from the MIRACL dataset, controlling for ten specific features. Synthetic queries were generated via LLMs to evaluate retrieval performance on these pairs.
- LLM-Based Explanations: A two-stage framework using Large Language Models (LLMs) was employed. First, an LLM ( $M_1$ ) analyzed pairs of documents (e.g., relevant vs. unjudged, or high-prior vs. low-prior) to generate natural language explanations of their differences. Second, a second LLM ( $M_2$ ) used these explanations to classify held-out documents, validating the reliability and generalizability of the identified features.

Key Results

Existence and Generalization of Priors: Supervised bi-encoders consistently encode relevance priors that generalize to unseen documents. In datasets like FEVER and Climate-FEVER, the prior models achieved high AUC scores (>0.8), indicating that systematic differences between relevant and unjudged documents are captured in the embeddings.
Cross-Model Consistency: Priors are not model-specific; different retrievers trained on similar data exhibit moderate to high correlation (0.5–0.7 Pearson $r$ ) in their prior scores, suggesting they internalize the same annotation biases.
Impact on Findability: There is a strong positive correlation between a document's relevance prior and its findability for supervised neural retrievers. Documents with lower priors are systematically harder to retrieve, even when they are genuinely relevant. This effect is absent or significantly weaker in unsupervised lexical methods like BM25.
Causal Confirmation: In matched-pair experiments controlling for confounders (length, entity density, etc.), supervised retrievers still favored high-prior documents over comparable low-prior ones, confirming that the prior itself drives the findability gap.
Drivers of Bias: LLM-based explanations revealed that annotation protocols favor comprehensive, self-contained summaries of mainstream topics (e.g., Wikipedia lead paragraphs, encyclopedic entries). Conversely, niche, fragmentary, context-dependent, or highly technical content (e.g., practical how-to guides, mid-article paragraphs, technical abstracts) is frequently left unjudged. Retrievers internalize these biases, ranking documents with "encyclopedic" features higher regardless of actual query relevance.

Significance and Claims
The paper claims to expose a structural limitation of supervised retrieval: models trained on annotated data do not merely learn relevance; they also learn the implicit document preferences inherent in the training data's annotation process.

Reframing Priors: While prior work has utilized query-independent signals for index pruning to save compute, this work reframes such signals as a potential source of bias that penalizes genuinely relevant but "low-prior" documents.
Findability Gap: The study demonstrates that the "findability gap" is not just a function of query complexity but is structurally embedded in the retriever's representation space due to annotation artifacts.
Diagnostic Nature: The authors position this work as diagnostic. They demonstrate the existence and mechanism of the problem but do not propose specific mitigation strategies, noting that such solutions would require trade-offs and further investigation. They also highlight the risk of adversarial misuse, where knowledge of these priors could be exploited to artificially inflate rankings.

The findings suggest that the "relevance" learned by current neural retrievers is a conflation of true query-document compatibility and the statistical biases of how relevance judgments are collected.