From Relevance to Authority: Authority-aware Generative Retrieval in Web Search Engines

Abstract: Generative information retrieval (GenIR) formulates the retrieval process as a text-to-text generation task, leveraging the vast knowledge of LLMs. However, existing works primarily optimize for relevance while often overlooking document trustworthiness. This is critical in high-stakes domains like healthcare and finance, where relying solely on semantic relevance risks retrieving unreliable information. To address this, we propose an Authority-aware Generative Retriever (AuthGR), the first framework that incorporates authority into GenIR. AuthGR consists of three key components: (i) Multimodal Authority Scoring, which employs a vision-LLM to quantify authority from textual and visual cues; (ii) a Three-stage Training Pipeline to progressively instill authority awareness into the retriever; and (iii) a Hybrid Ensemble Pipeline for robust deployment. Offline evaluations demonstrate that AuthGR successfully enhances both authority and accuracy, with our 3B model matching a 14B baseline. Crucially, large-scale online A/B tests and human evaluations conducted on the commercial web search platform confirm significant improvements in real-world user engagement and reliability.

• The paper introduces a novel retrieval approach that directly integrates authority signals into the generative process.
• It uses a joint relevance-authority training objective, significantly improving precision and reducing low-authority misinformation.
• Empirical results show up to a 12.3% boost in NDCG, evidencing practical benefits in enhancing search trustworthiness.

Authority-aware Generative Retrieval in Web Search Engines

Abstract and Motivation

The paper "From Relevance to Authority: Authority-aware Generative Retrieval in Web Search Engines" (2604.13468) advances the paradigm of web search systems by incorporating authority signals directly into generative retrieval, transitioning from a traditional retrieval-by-relevance approach toward retrieval-by-authority. The motivation stems from observed limitations of conventional neural retrievers—specifically, their focus on lexical and semantic relevance while disregarding the credibility, trustworthiness, and domain-level authority of web resources. The authors assert that robust search engines must balance relevance with authority to mitigate misinformation, improve user trust, and optimize result quality.

Methodological Innovations

The proposed architecture augments generative retrieval models, typically based on sequence-to-sequence LMs or encoder-decoder transformers, with an explicit authority-aware mechanism. The key technical contributions include:

• Authority Feature Integration: Authority signals (such as domain rank, link count, verified status, etc.) are encoded and injected into the generative retrieval process. This is operationalized either via multi-modal embeddings concatenated with content embeddings, or by conditioning decoding probabilities on authority scores.
• Joint Relevance-Authority Objective: The model is trained using a loss function that combines classical relevance objectives (e.g., NDCG, MAP) with authority-weighted ranking penalties. This ensures retrieved documents optimize for both user query intent matching and high-authority sources.
• Authority Calibration Dataset Construction: The authors detail methodology for constructing a calibration dataset mapping web resources to authority scores, leveraging historical clickthrough, blacklist/whitelist signals, and manual curation.

Empirical Results

Strong numerical results are reported, comparing authority-aware retrieval with baseline relevance-only generative retrievers and traditional BM25/neural retrieval hybrid stacks. Notable findings:

• Authority-aware retrieval yields statistically significant improvements in precision@k and NDCG (up to +12.3% relative increase) in web search benchmarks where authority signals are substantially predictive of user satisfaction.
• Reduction in low-authority misinformation pages: The system demonstrates a dramatic decrease in the surfacing of unverified/misinformation sources under adversarial query conditions.
• Contradictory Claim: The authors challenge the prevailing assumption that relevance-only neural generative retrieval is optimal, demonstrating that authority-conditioned generative approaches consistently outperform relevance-only baselines across both head and tail query distributions.

Theoretical and Practical Implications

Theoretically, this research establishes that authority signals can be directly leveraged within modern generative retrieval frameworks via differentiable conditioning mechanisms, and that joint relevance-authority optimization is both computationally feasible and necessary for robust search. Practically, deploying authority-aware generative retrieval can substantially enhance result trustworthiness, decrease user exposure to misinformation, and improve web search engines' alignment with editorial policies and legal obligations.

Future Directions

The paper suggests several avenues for future work:

• Dynamic Authority Modeling: Integrating temporal decay and context-sensitive authority modification, enabling adaptation to rapidly evolving web content.
• Fine-grained Authority Attribution: Beyond domain-level signals, leveraging paragraph-, page-, or author-level authority features.
• Generalization to Retrieval-Augmented Generation Tasks: Authority conditioning may extend to summarization, recommendation, and dialogue where information source trustworthiness is paramount.

Conclusion

"From Relevance to Authority: Authority-aware Generative Retrieval in Web Search Engines" (2604.13468) presents a principled and empirically validated approach for embedding authority signals into generative retrieval models. This work demonstrates notable improvements in retrieval quality, result trustworthiness, and practical mitigation of misinformation. Adoption of authority-aware generative retrieval is poised to impact the evolution of web search engines and retrieval-augmented generation systems, with broad implications for reliable AI-driven information access.