Most modern search ranking systems are built around a compromise. One model retrieves a manageable set of candidate documents quickly, while a more computationally expensive model examines those candidates more carefully and reranks them.
A new research paper involving Google DeepMind asks whether a large language model could collapse more of that process into a generative ranking system. Instead of representing every document only as a vector and then scoring candidates with a separate reranker, Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders trains an LLM to generate document identifiers token by token and use those probabilities to produce an ordered ranking.
The approach is technically ambitious. The researchers provide a theoretical argument that Autoregressive Ranking, or ARR, can express rankings that become increasingly difficult for conventional dual encoders as the corpus grows. They also introduce a new rank-aware training objective called SToICaL that improves several experimental ranking metrics.
But this is research, not a Google Search launch. There is no evidence in the paper or accompanying reporting that ARR is currently deployed in Google Search, and the experiments use WordNet and Amazon's ESCI shopping-query dataset rather than the live web. The results are promising in several areas and mixed in others.
Autoregressive Ranking treats document IDs as something an LLM can generate
The central idea behind ARR is conceptually different from the standard retrieve-then-rerank pipeline.
Each document is assigned an identifier that can be represented as a sequence of tokens. Given a query, the language model learns to generate the identifiers of relevant documents autoregressively, one token at a time.
The probability assigned to those token sequences provides a way to rank the documents. Beam search can be used to explore likely document IDs and construct an ordered set of results.
In other words, the model is not generating a prose answer. It is generating identifiers that correspond to documents in the retrieval collection.
Why search systems commonly use dual encoders
Dual encoders solve an important efficiency problem in information retrieval.
A dual encoder separately converts the query and each document into vector representations. Because document vectors can be computed in advance, the system can compare a new query against a large collection relatively efficiently and retrieve promising candidates.
This architecture is highly practical at scale, but the separation between query and document encoding limits how richly the model can reason about their interaction.
The paper focuses on a deeper theoretical limitation: representing arbitrary rankings with a dual encoder can require the embedding dimension to grow as the number of documents being ranked grows.
Cross encoders trade efficiency for richer relevance modeling
Cross encoders approach relevance differently.
Instead of encoding the query and document independently, a cross encoder processes them together. This allows the model to examine detailed interactions between the query and the candidate document, which can improve ranking quality.
The cost is computation. Running a powerful cross encoder across an enormous corpus for every query would be prohibitively expensive, so these models are commonly used after a faster retrieval stage has already reduced the candidate set.
This is the architectural gap the ARR paper is trying to bridge.
ARR aims for more expressive ranking without a conventional cross-encoder pass
The researchers argue that autoregressive generation provides a different way to model relevance.
Rather than forcing every document into a fixed-dimensional vector space and then relying on a second model to recover richer query-document interactions, ARR learns a probability distribution over document identifiers.
That generative representation can, at least theoretically, express more complicated orderings than a fixed-dimensional dual encoder.
The promise is a ranking architecture with more expressive power than standard dense retrieval while avoiding the need to run an expensive cross encoder over every possible document.
The paper proves an expressivity advantage over dual encoders
One of the paper's main contributions is theoretical rather than empirical.
The authors show that if a dual encoder must be capable of representing any ordering of k documents, its embedding dimension must grow linearly with k.
Under the conditions defined in the paper, an autoregressive ranking model with multi-token document IDs can represent arbitrary rankings while maintaining a constant hidden dimension.
That is a meaningful mathematical distinction, but it should not be translated into the claim that ARR is automatically faster, cheaper or better on a production web search engine.
Expressive capacity is not the same as real-world search performance
A model can be theoretically capable of representing a ranking without being the best practical system for serving billions of queries.
Production search has to consider latency, memory, index construction, document updates, fault tolerance, spam, multilingual retrieval, freshness, safety and enormous corpus sizes.
The ARR paper establishes a formal advantage in ranking expressivity and demonstrates experimental gains on selected datasets.
It does not demonstrate a complete web-scale replacement for an operational search engine.
Standard language-model training is not designed for ranking
Once ranking is framed as token generation, another problem appears.
Large language models are normally trained with next-token prediction. That objective rewards the model for predicting the correct next token, but it does not naturally understand that one complete document identifier should rank above another because the underlying document is more relevant.
The authors describe conventional next-token loss as rank-agnostic for this task.
Their solution is a new fine-tuning objective called SToICaL.
SToICaL stands for Simple Token-Item Calibrated Loss
SToICaL is designed specifically to teach the autoregressive model about document-level relevance.
Instead of treating every correct document ID equally, the loss incorporates the relevance ordering supplied by the training data.
Higher-ranked items receive greater importance, while the model's token probabilities are calibrated so paths leading toward more relevant document IDs receive more probability mass.
This connects token-level generation with the item-level ranking objective the search system actually cares about.
Prefix-tree marginalization helps manage multi-token document IDs
Document identifiers can contain multiple tokens, which creates a structural challenge.
Different valid document IDs may share the same initial token sequence. Penalizing a shared prefix too early could incorrectly suppress multiple relevant documents.
SToICaL uses prefix-tree marginalization to distribute probability across valid continuations while respecting the relevance of the underlying items.
This is one of the mechanisms that lets the model learn ranking behavior rather than merely memorize isolated token sequences.
The researchers tested ARR on WordNet and ESCI
The experimental section uses two very different datasets.
WordNet provides a controlled setting that allows the researchers to investigate ranking expressivity and compare architectures under relatively clean conditions.
Amazon's ESCI Shopping Queries dataset provides a more applied retrieval problem involving product searches and graded relevance labels.
Neither dataset should be confused with Google's live web index, but together they allow the paper to test both theoretical behavior and practical ranking metrics.
SToICaL improves ranking beyond the first result
The authors report that their rank-aware objective improves several metrics that evaluate more than whether the single top result is correct.
This is important because search ranking is an ordered-list problem.
A model that retrieves one excellent result but places irrelevant documents immediately beneath it may be less useful than a model that produces a consistently relevant top ten.
The paper says SToICaL significantly improves ranking metrics beyond top-1 retrieval and helps suppress invalid document-ID generations.
ARR can perform similarly to a cross encoder in the WordNet experiments
In the WordNet comparisons, the autoregressive approach performs strongly against the architectures it is intended to bridge.
Search Engine Journal's analysis highlights that ARR performed similarly to the cross encoder in the reported WordNet setting while significantly outperforming the dual encoder.
That is an encouraging result because cross encoders represent the more expressive but computationally expensive side of the traditional architecture.
It remains a controlled experimental comparison rather than evidence that ARR would match a production reranking stack across the open web.
The results are not uniformly better
The paper is more interesting because it does not produce a simple “new model wins everything” conclusion.
On the ESCI shopping-query experiments, some SToICaL configurations improve broader ranking metrics such as MRR@10 and NDCG@10 while performance on Hits@1 can decline.
That means the model can become better at ordering the overall result set while becoming worse at placing the single most relevant document first.
For a search engine, that tradeoff matters enormously because the first result often receives disproportionate user attention.
Better NDCG does not automatically mean a better search experience
Ranking metrics capture different properties.
NDCG rewards placing highly relevant documents near the top while accounting for graded relevance across a list. MRR emphasizes how early the first relevant result appears. Hits@1 asks whether the top position is correct under the chosen definition.
A training method can improve one metric and hurt another because it is changing the shape of the entire ranking distribution.
That is why the mixed ESCI results should be treated as a research signal rather than a solved architecture.
Invalid document IDs are a unique generative-ranking problem
A conventional retriever selects from documents already present in an index.
A generative model can potentially produce a token sequence that does not correspond to a valid document identifier.
This creates an analogue of hallucination at the retrieval layer: the model may generate an impossible ID rather than selecting a real item.
The paper reports that the rank-aware training approach helps suppress these invalid generations, an essential requirement if autoregressive ranking is to become reliable.
Document identifiers become part of the retrieval architecture
ARR also changes the role of document IDs.
In a traditional index, an identifier is largely a reference used to locate a document after retrieval. In autoregressive ranking, the structure of the identifier becomes something the model must learn to generate.
That raises interesting engineering questions about how IDs should be assigned, how new documents enter the corpus and how deleted or changed documents are handled.
The paper focuses on ranking theory and training rather than solving every production indexing problem.
Beam search turns generation into a ranked list
Language models generate probabilities over possible next tokens.
Beam search keeps several promising token sequences alive during generation instead of greedily choosing only one path at each step.
For ARR, those candidate sequences correspond to possible document identifiers. Their accumulated probabilities can then support a ranked ordering of documents.
This is the mechanism behind the phrase “the LLM generates the ranking itself”: the ordered result list emerges from autoregressive generation rather than a separate scalar scoring pass over every candidate.
Could ARR unify retrieval and reranking?
That is the most provocative architectural possibility, but it needs careful wording.
ARR is designed to bridge the gap between efficient dual encoders and expressive cross encoders. If a generative ranker can retrieve relevant document IDs while also modeling enough relevance structure to order them well, some boundaries between retrieval and reranking could become less rigid.
That does not mean the paper has demonstrated a single model that can replace every stage of a web search stack at Google scale.
“Unification” is a research direction suggested by the architecture, not a production deployment claim.
This is not evidence of a new Google Search ranking algorithm
The paper includes researchers from Google DeepMind alongside academic collaborators, but publication affiliation does not establish deployment.
Nothing in the paper says Google Search currently uses ARR.
Search Engine Journal also frames the work as research into a potential future ranking architecture rather than confirmation of a live Google ranking change.
SEOs should therefore avoid adding “ARR” or “SToICaL” to lists of confirmed Google Search ranking systems.
The paper was submitted months before the September coverage
The research itself is not a September launch.
The arXiv record shows the first version was submitted on January 9, 2026, with subsequent revisions culminating in version four on February 11.
The September 10 Search Engine Journal article brought renewed attention to the work for the search-marketing audience.
This distinction matters because a news article discussing research does not indicate that Google deployed the system on the publication date.
The research comes from Google DeepMind and university collaborators
The authors include Benjamin Rozonoyer, Chong You, Michael Boratko, Himanshu Jain, Nilesh Gupta, Srinadh Bhojanapalli, Andrew McCallum and Felix Yu.
The affiliations span Google DeepMind, the University of Massachusetts Amherst and the University of Texas at Austin.
That combination places the paper at the intersection of industrial search research and academic information retrieval.
It also reinforces why the work deserves attention without turning it into a statement about Google's current production architecture.
ARR challenges a fundamental limitation of fixed embeddings
The theoretical result has implications beyond one benchmark.
Dense retrieval systems rely on fixed-size vectors to represent queries and documents. As the number of possible orderings grows, a fixed-dimensional geometry can limit which rankings the model can represent.
ARR moves the representation problem into a sequential generative process.
Because the model can use multiple tokens to identify documents, it gains a different kind of representational capacity that does not require the hidden dimension to grow linearly with corpus size under the paper's assumptions.
That does not make embeddings obsolete
It would be premature to conclude that autoregressive ranking will eliminate vector retrieval.
Embeddings are efficient, mature and widely deployed across search, recommendations and retrieval-augmented generation. They support fast approximate nearest-neighbor search over huge collections and have a large ecosystem of optimized infrastructure.
A new ranking paradigm has to beat not only a mathematical limitation but an enormous body of production engineering.
ARR should be viewed as a competing architectural idea, not a declaration that dual encoders are finished.
Latency will be one of the major practical questions
Autoregressive generation is sequential.
Each token depends on earlier tokens, and beam search can maintain multiple candidate paths. That creates different latency characteristics from retrieving nearest neighbors in a vector index.
The paper's expressivity analysis does not by itself establish that generating enough document IDs for a high-quality ranking will meet the latency requirements of a consumer web search engine.
Production adoption would require careful optimization of serving cost and response time.
Corpus updates create another challenge
The open web changes continuously.
New pages appear, old pages disappear and documents change meaning as their contents are updated.
A generative ranker tied to document identifiers needs a robust mechanism for incorporating that churn without expensive retraining every time the corpus changes.
How efficiently an ARR-style system handles rapidly changing indexes could be as important as its benchmark ranking quality.
Web search also has adversarial inputs
WordNet and shopping datasets are useful experiments, but the public web contains spam, manipulation, hacked content, duplication, misinformation and adversarial SEO.
A production ranker has to operate under active attempts to influence its decisions.
Generative ranking introduces additional questions about invalid IDs, memorization and robustness under distribution shifts.
The paper does not claim to have solved the full adversarial-search problem.
The SEO implications are therefore hypothetical
If autoregressive ranking eventually influenced production search, SEO professionals might need to rethink some assumptions about how retrieval and ranking stages interact.
But there is no actionable ARR optimization tactic in this paper.
The experiments do not identify page features that publishers can manipulate, and they do not reveal a Google Search scoring formula.
Trying to reverse-engineer content advice from the model's docID generation mechanism would go far beyond the evidence.
Relevance still remains the target
Despite the architectural novelty, the training objective is still fundamentally about relevance.
SToICaL uses ground-truth relevance information to teach the model which documents should receive more probability and which should be pushed downward.
For publishers, that reinforces a familiar principle: the retrieval architecture may evolve dramatically while the system's objective remains connecting a user's information need with useful documents.
There is no indication that ARR would reward arbitrary formatting tricks disconnected from relevance.
Generative ranking fits a broader shift in information retrieval
Search research is increasingly exploring models that generate retrieval targets, queries, answers and structured identifiers rather than only assigning independent relevance scores.
Large language models make these approaches more practical because they can model complex conditional distributions and sequences.
ARR belongs to that broader movement toward generative information retrieval.
The interesting question is whether generative models can deliver enough quality and flexibility to justify their serving complexity at scale.
Search and answer generation could eventually share more infrastructure
AI search products already combine retrieval with generative answer systems.
Today, those components can involve distinct models and stages: retrieve documents, rerank them, extract evidence and generate a response.
Research like ARR suggests a future in which generative models participate much earlier in the pipeline, potentially selecting and ordering evidence as part of the same family of modeling techniques used to produce the answer.
That architectural convergence is plausible, but the ARR paper itself focuses on document ranking rather than end-to-end answer generation.
SToICaL may be as important as ARR itself
The paper's practical contribution is not only the autoregressive architecture.
It also identifies a mismatch between standard language-model loss and ranking objectives, then proposes a way to make training explicitly rank-aware.
Even if future systems use a different generative ranking architecture, the broader lesson remains valuable: teaching an LLM to retrieve documents requires objectives that care about the order and relevance of complete items, not merely whether each next token is plausible.
That is an important distinction as LLMs move deeper into search infrastructure.
The mixed results are a feature of credible research
The ESCI outcome where broader ranking metrics improve while top-one performance can decline is not a reason to dismiss the work.
It is evidence that the researchers are exposing a real optimization tradeoff.
Ranking systems often balance multiple goals, and improvements in one metric can reveal weaknesses in another.
The next research step is not to declare ARR successful or unsuccessful, but to understand which training objectives and architectures improve the full set of outcomes that matter.
What would signal that ARR is moving toward production?
Search marketers should look for stronger evidence than another research citation.
Useful signals would include Google Search documentation naming an autoregressive ranking system, engineering papers describing web-scale deployment, public statements from Search leadership, production latency and cost studies, or patents and technical disclosures connected explicitly to live systems.
Until then, ARR belongs in the research category.
That does not make it unimportant. It simply keeps the claim proportional to the evidence.
The future of ranking may be more generative than today's pipeline
Autoregressive Ranking is compelling because it attacks a foundational search problem from a different direction. Instead of representing documents only as fixed vectors and depending on a separate expensive model to refine their order, an LLM learns to generate document identifiers in a sequence that encodes relevance.
The paper provides a theoretical argument that this approach is more expressive than dual encoders, while SToICaL gives the model a rank-aware objective that improves several experimental metrics and reduces invalid generations. On WordNet, ARR can approach cross-encoder performance in the reported setting; on ESCI, the picture is more mixed, including tradeoffs between top-one accuracy and broader ranking quality.
That is enough to make the research worth watching and nowhere near enough to call it Google's new ranking algorithm.
The important signal is architectural. Search researchers are investigating whether LLMs can move from generating answers after retrieval to participating directly in the generation of the ranked document list itself. If that idea survives the enormous practical demands of web-scale search, retrieval and reranking may eventually look very different. For now, ARR is a promising experiment showing what that future could look like—not evidence that Google Search has already arrived there.