LongRetriever: Ultra-Long Sequence Retrieval

• LongRetriever is a recommendation retrieval framework that exploits ultra-long user behavior sequences to overcome the limitations of short-history candidate filtering.
• It incorporates in-context training to mitigate information leakage, ensuring that training procedures align with the candidate-specific dual-tower serving architecture.
• The framework employs multi-context retrieval by generating multiple user vectors for distinct interests, thereby enhancing both retrieval diversity and relevance in large-scale deployments.

LongRetriever is a recommendation retrieval framework designed for the candidate retrieval stage of industrial recommender systems, with the explicit goal of exploiting users’ ultra-long behavior sequences rather than restricting retrieval to short recent histories. It combines in-context training and multi-context retrieval so that user sequence modeling becomes candidate-specific while remaining compatible with dual-tower approximate nearest neighbor serving. In the reported large-scale deployment, the framework was evaluated through online A/B testing on a major e-commerce platform and then fully deployed, with reported impact on billions of users (Qin et al., 21 Aug 2025).

1. Problem setting and design objective

LongRetriever is motivated by a structural asymmetry in modern recommender systems: ultra-long user sequences are commonly exploited in the ranking stage, but the candidate retrieval stage typically uses only very short histories, such as the latest 200 behaviors. The paper identifies two resulting limitations: loss of deep interest information, because short histories may overlook diverse and long-term interests, and limited user-item interaction, because the standard dual-tower paradigm computes user and item embeddings separately (Qin et al., 21 Aug 2025).

The framework targets the retrieval stage rather than reranking. Its objective is to make retrieval sensitive to lifelong or near-lifelong user behavior while preserving the efficiency properties required by production systems. In the formulation described in the paper, a user $u$ has a complete behavior sequence $B_u = [b^u_1, b^u_2, ..., b^u_T]$, with $T$ potentially exceeding 20,000 interactions. The central design problem is therefore not merely long-sequence encoding, but long-sequence encoding under retrieval-stage constraints: pre-indexable item representations, fast ANN search, and training-serving consistency (Qin et al., 21 Aug 2025).

A common misconception in this area is that ultra-long sequence modeling is mainly a ranking-stage concern. LongRetriever is explicitly framed as evidence against that assumption: it is proposed precisely because candidate retrieval had remained under-explored relative to ranking, even though retrieval quality constrains everything that follows in the recommendation pipeline (Qin et al., 21 Aug 2025).

2. Candidate-specific long-sequence modeling

LongRetriever does not feed the entire behavior history into a monolithic encoder for every retrieval decision. Instead, it introduces category-based filtering. For a candidate item context, usually its category, the system filters the user’s full history and retains the most recent $L$ behaviors matching that context, producing a subsequence $S_u = [s^u_1, ..., s^u_M]$. This reduces the computational burden while making the user representation explicitly conditioned on candidate context (Qin et al., 21 Aug 2025).

The user encoder is a Pre-Layer Normalization Transformer. The paper describes the user representation as

$$e_u = \text{Transformer}([CLS, p^u_1, ..., p^u_K, s^u_1, ..., s^u_M]),$$

where the input includes a special $[CLS]$ token, profile features, and the filtered behavior subsequence. Item representations $e_v$ are computed independently, preserving the dual-tower paradigm and enabling item-side pre-indexing for ANN retrieval (Qin et al., 21 Aug 2025).

This architecture is intended to reconcile two normally competing requirements. On one side, retrieval must remain search-based and scalable. On the other, the user representation should reflect behavior that is relevant to the candidate rather than a generic summary of the entire history. The paper characterizes this as enabling candidate-specific interaction between user sequence and candidate item while preserving the serving efficiency of decoupled towers (Qin et al., 21 Aug 2025).

A plausible implication is that LongRetriever occupies an intermediate point between standard dual-tower retrieval and fully interaction-heavy ranking models. It does not introduce item-conditioned cross-attention at serving time, but it uses context-filtered histories so that the retrieved user vector is still specialized to a candidate-relevant context.

3. In-context training and leakage control

A technical obstacle arises when context-matching subsequences are inserted directly into contrastive retrieval training. The paper emphasizes that this can cause information leakage, especially with in-batch contrastive losses. If a minibatch contains multiple categories, the model may learn to align the context label shared by the filtered subsequence and the candidate item, rather than learning genuine behavioral compatibility (Qin et al., 21 Aug 2025).

LongRetriever addresses this with in-context training. The core intervention is batch construction: all user-candidate pairs within the same minibatch share the same context, such as category. Under this scheme, in-batch negatives are not trivially separable by context identity, because positives and negatives come from the same context. The model is thus forced to discriminate the true target item from hard negatives within that shared context (Qin et al., 21 Aug 2025).

The significance of this choice is not only regularization. The paper explicitly frames it as a mechanism for training-serving consistency under the search-based paradigm. At serving time, retrieval is performed within context-specific search spaces; in-context training makes the learning problem structurally closer to that serving condition. The framework therefore treats batch design as part of the retrieval model itself rather than as an implementation detail (Qin et al., 21 Aug 2025).

The paper also reports that omitting this component is undesirable: skipping in-context results in leakage and inferior conversion despite superficially high exposure ratios. This is one of the clearest cases in the work where better offline-looking behavior is not treated as evidence of a better retriever (Qin et al., 21 Aug 2025).

4. Multi-context retrieval and interest decomposition

Because user and item towers are decoupled at serving time, candidate information cannot directly interact with user sequences when ANN search is executed. LongRetriever’s answer is multi-context retrieval, a serving-time procedure that generates several user vectors per user, each associated with a distinct inferred interest context (Qin et al., 21 Aug 2025).

The framework estimates several user interests, described in the paper as contexts such as categories, and selects the top $N$ categories based on a time-weighted engagement score. The documented procedure favors recent behaviors by weighting category evidence with the time difference in days. The specific interest selection policy highlighted in experiments is “Random in Top”, including the deployed setting “Random 5 in Top 20” (Qin et al., 21 Aug 2025).

Each selected context yields a separate user vector. On the item side, the repository is partitioned by category or context, so retrieval is executed independently in each sub-repository and the results are merged. This produces a multi-interest, multi-vector retrieval process rather than a single global nearest-neighbor lookup (Qin et al., 21 Aug 2025).

The paper attributes two properties to this design. First, it restores candidate-relevant specialization at serving time despite tower decoupling. Second, it provides interpretability and control: the interests governing retrieval are explicit and tunable by business goals. In that sense, LongRetriever is not only a long-sequence encoder but also a retrieval control mechanism (Qin et al., 21 Aug 2025).

5. Empirical evaluation and industrial deployment

The reported deployment environment is a large-scale e-commerce platform with 900 million users, 150 million items, 10+ billion interaction samples, and sequence lengths reaching 20,000 on average. The baseline deployed in production is MIND, described in the paper as a standard multi-interest model. The principal online comparison is a 7-day A/B test between LongRetriever and MIND, using Random 5 in Top 20 as the interest selection configuration (Qin et al., 21 Aug 2025).

The paper reports statistically significant gains with $p < 0.05$ on all listed business metrics.

Metric	Reported change
PV	+0.62%
UV CTR	+0.17%
UV CVR	+1.33%
Orders per User	+1.7%
Exposed Categories per User	+0.14%
Clicked Categories per User	+1.39%

The work also reports intermediate retrieval indicators that characterize exposure behavior rather than only downstream business outcomes.

Intermediate metric	LongRetriever	MIND
AER	18.72%	31.82%
UER	9.29%	4.58%
Unique Conversion Efficiency	7.92%	6.89%

The interpretation given in the paper is specific. LongRetriever’s lower AER is attributed to the fact that multi-context retrieval is limited to a few interest categories, whereas MIND retrieves globally. Its higher UER indicates that it retrieves more unique items than the baseline, and its higher Unique Conversion Efficiency indicates better conversion on those uniquely exposed items (Qin et al., 21 Aug 2025).

Ablation findings refine this picture. Random 5 in Top 20 improves diversity and UER relative to simply taking the top 5 interests, although pure top-5 yields better conversion efficiency alone. Excluding long sequence degrades all metrics. Skipping in-context produces leakage and inferior conversion despite superficially high exposure ratios. Taken together, these results support the paper’s claim that both ultra-long sequence usage and the training protocol are essential to the observed gains (Qin et al., 21 Aug 2025).

The framework is reported as fully deployed on the platform, with impact on billions of users. This deployment claim is notable because the paper presents LongRetriever not as a purely offline modeling exercise but as a production retrieval system validated in a live recommendation environment (Qin et al., 21 Aug 2025).

6. Limitations, open directions, and broader significance

The paper identifies a clear limitation in the current system: interest assignment is simple. Context selection is based on time-weighted engagement over categories, and the authors state that more sophisticated methods could yield further gains. The specific future direction named in the paper is the use of LLMs to process user information and generate more optimized, adaptive interest or category assignments (Qin et al., 21 Aug 2025).

This limitation is structurally important. LongRetriever’s effectiveness depends not only on its encoder but also on the quality of the context partition used to filter ultra-long histories and route retrieval. If the chosen contexts are too coarse, the model may under-express user intent; if they are overly fragmented, retrieval may become sparse or unstable. The paper does not claim to have solved this interest-assignment problem fully (Qin et al., 21 Aug 2025).

Within the recommendation literature, LongRetriever is best understood as a retrieval-stage framework that extends multi-interest retrieval with ultra-long sequence conditioning, candidate-specific context filtering, and explicit training-serving alignment. Its contribution is therefore not reducible to “using longer histories.” Rather, it specifies how long histories can be made operational in a dual-tower ANN system without giving up controllability, industrial latency constraints, or large-scale deployment feasibility (Qin et al., 21 Aug 2025).

This suggests a broader methodological point: in industrial recommenders, long-context modeling can be valuable before ranking, provided that the context is organized into search-compatible interests and the training regime is matched to the serving pathway. LongRetriever formalizes that design point in a concrete production framework (Qin et al., 21 Aug 2025).