recommendation system engineers

Abstract: Dynamic web interactions continuously evolve as new edges arrive, offering large language models (LLMs) great potential for reasoning over edge-level temporal semantics. However, continuously arriving edges are often repetitive, obscuring the key structures that truly reflect temporal changes and limiting the efficiency of LLMs. A key insight is that beneath the dense stream of edge arrivals lies a sparse evolution of temporal intent, enabling LLMs to be responsively triggered only when such intent shifts, while otherwise inheriting the previous state. To this end, we propose a hypergraph-based framework named HyperIntent, which enhances dynamic graph reasoning by parsing differentiated intents, quantifying their temporal discrepancies, and performing shift-triggered semantic updating. Experimental results on destination node retrieval and future link prediction demonstrate that HyperIntent consistently achieves the best overall performance, with average relative Hits@1 gains of 34.0% and 25.4% over vanilla TGAT and DyGFormer, respectively. Meanwhile, HyperIntent reduces LLM token consumption by 54.4%-95.9% across six datasets while achieving leading predictive performance. In-depth studies further confirm the necessity of shift triggering for identifying structural and semantic temporal changes. The code will be made publicly available upon completion of the review process.

Abstract: LLM-based generative recommendation reads long, text-rich interaction histories but generates only short semantic identifiers (SIDs). This asymmetry makes supervised fine-tuning inefficient: encoding the history is expensive, yet each history usually provides supervision for only one target. We present Merge and Branch (M&B) to reduce input cost and learn more from each encoded history. Merge compresses each historical item's SID and text into a small number of learned tokens. Branch uses the compressed history to predict multiple target SIDs. To share history computation while keeping these targets independent, M&B organizes each example as a shallow, wide tree. We develop tree FlashAttention to execute this layout efficiently, processing the shared history once and propagating gradients from every branch for end-to-end training. A tree-structured key/value cache extends history sharing to constrained decoding. Experiments on public benchmarks show that M&B outperforms strong recommendation baselines while substantially improving training and decoding efficiency.

Abstract: Recommender Systems Engineer Agents (RSE-Agents), which use LLM-based agents to autonomously improve recommendation algorithms, are emerging as a promising paradigm. However, their capabilities and limitations remain poorly understood due to the lack of a unified, standardized evaluation framework and systematic trajectory-level analysis. To address this gap, we introduce AutoRSE, an evaluation framework spanning six representative industrial-grade recommendation tasks, and conduct a large-scale empirical study of three frontier LLMs across three widely used agent harnesses. Our results show that RSE-Agents consistently enhance the performance of backbones on five of the six tasks, with an average relative improvement of 16.1%, indicating that they can already autonomously improve recommendation performance. However, trajectory-level analysis also reveals that current RSE-Agents function primarily as engineering optimizers rather than algorithmic improvers: their gains are largely driven by hyperparameter tuning, training strategies, and model ensembling, while genuine algorithmic innovations are rare. Since the largest gains in recommender systems often stem from algorithmic advances, this behavior underscores a fundamental limitation of current RSE-Agents. To probe this limitation, we perform idea-injection experiments and show that agents equipped with innovative ideas from frontier research or human experts achieve substantially larger improvements than those relying on self-generated ideas. These results suggest that idea generation is a central bottleneck of current RSE-Agents, highlights a key direction for the future recommendation system research. Our implementation is publicly available on GitHub: https://github.com/Ray-Tiann/AutoRSE.

Abstract: Existing approximate nearest neighbor search (ANNS) methods generally assume that each entity has a single vector. In many real-world applications, however, an entity is associated with more than one vector, and queries may jointly consider two attributes. With different queries assigning different weights to each attribute, Pareto-frontier-based indices provide an effective solution for dual-vector retrieval by connecting to approximate nearest neighbors under all possible query preferences. However, existing Pareto-frontier-based approaches suffer from substantial inefficiencies in both index construction and dynamic updates, limiting their practical applicability. To address these issues, we propose ACTIVE, a dynamic Pareto-frontier-based DVQ index that supports efficient construction and updates while maintaining high retrieval performance under varying query preferences. It employs angle-constrained active-range computation and connectivity-loss-driven adaptive repair strategies to accelerate node insertion and deletion, respectively. Extensive experiments show that, at comparable (or slightly below) search accuracy, ACTIVE improves insertion throughput by $1.20$--$1.43\times$ over DEG, and deletion throughput by $1.64$--$2.06\times$ over neighbor-reconnection-based repair and by $22.64$--$39.51\times$ over reinsertion-based repair.

Abstract: In stochastic multi-objective bandits, Pareto optimal arms may represent different trade-offs among objectives. Balanced sampling across them is important in practical settings such as online advertising. Yet, there seems to be little consensus on how to measure fairness, to our best knowledge. Achieving fairness is also tough as the set of Pareto optimal arms is unknown, and early sampling decisions may create imbalances that are difficult to erase afterwards. We measure unfairness by the KL divergence from the uniform distribution over the true Pareto set to the empirical measure induced by the sampling trajectory. Nicely, this measure admits an exact decomposition that separates the cost of sampling suboptimal arms from the imbalance among Pareto optimal arms. With a recursive form for unfairness, we develop \emph{adaptive top-two sampling}, which combines sampling counts and empirical evidence of Pareto membership to define a gain for each arm and guide sampling without committing to an estimated Pareto set. When the Pareto set is stable under small perturbations, we establish an $O(\log T/T)$ and $O(T^{-1})$ bound on expected unfairness and variance, respectively. Information-theoretic lower bounds show that the $\log T/T$ rate can be unavoidable for uniformly fair policies on stable instances, matching our upper bound. More surprisingly, stability is indeed crucial, as the expected unfairness can be $ω(\log T/T)$ without it. Finally, we create a variant that encourages exploration of undersampled arms to reduce the risk of persistent underallocation. At sufficiently large horizons, it preserves the rate on the expected unfairness while admitting a sharper bound of $O(\log^2 T/T^2)$ on variance. Numerical studies support these findings and suggest practical insights.

Abstract: Conversational bandits improve recommendations by actively querying users about key-terms to elicit their item preferences. However, existing methods typically assume homogeneous feedback for items and key-terms, using either absolute rewards or pairwise comparisons. In practice, item feedback often takes the form of clicks or ratings, while comparisons are more natural for key-terms. To address this gap, we study conversational generalized linear bandits with hybrid feedback, combining absolute item rewards and relative key-term comparisons. We propose Conversational UCB with Hybrid Feedback (HyConUCB), which constructs unified confidence bounds from heterogeneous feedback and adaptively selects key-term comparisons by balancing information gain and directional coverage. Through a refined regret analysis that jointly exploits both feedback sources, we establish a regret bound that improves upon reward-only and adapted dueling-only benchmarks. Experiments on synthetic and real-world datasets validate the effectiveness of our algorithm.

Abstract: Generative recommendation with semantic IDs (SIDs) represents each item as a sequence of discrete tokens and autoregressively predicts the next item's SID. Despite strong recommendation performance, repeated decoding over multiple SID positions incurs substantial inference cost. Knowledge distillation transfers knowledge from a powerful generative teacher to a lightweight student, yet conventional objectives do not fully account for multi-token SID generation. We identify two limitations: item-level distribution alignment may not adequately preserve the teacher's ranking of candidate groups sharing SID prefixes, while uniform alignment across SID positions may affect final item ranking differently. We propose MuSID-KD, which combines Hierarchical Candidate Distribution Alignment (HCDA) and Ranking Impact Guided Levelwise Distillation (RIGD). HCDA aggregates item probabilities according to shared SID prefixes and aligns candidate distributions at successive levels, from candidate groups to individual items. RIGD evaluates the reduction in pairwise ranking surrogate loss when one student score contribution is replaced with the teacher's contribution, and allocates supervision accordingly. Experiments on real-world recommendation datasets using three lightweight student architectures demonstrate improved recommendation accuracy and better alignment of prefix marginal distributions and pairwise item rankings, without modifying the student's online inference architecture. Code is available at https://anonymous.4open.science/r/m7k2-review-83f9/.

Abstract: As a core component of modern RAG systems, approximate nearest neighbor (ANN) search has been extensively studied for decades. In practice, deploying ANN often requires symmetric privacy, where the querier's query and the database owner's data must remain hidden from each other, necessitating a secure multi-party computation (MPC) solution. So far, most MPC-based ANN protocols follow a cluster-based approach: they evaluate distances for the query over a large candidate set and rely on heavyweight primitives such as distributed ORAM to hide access patterns. As a result, each search typically requires $O(\sqrt{N})$ time. We propose HNSecW, a novel MPC protocol for ANN search that achieves the $O(\log N)$ search time for the first time. We first study the single-query setting. Starting from HNSW, the state-of-the-art ANN algorithm, we redesign its control flow and data structures to achieve fully data-independent execution. To further achieve logarithmic search time, we employ offline reshuffling and partial database duplication. Furthermore, we extend our protocol to support multiple sequential queries by introducing a "double visited" caching structure that eliminates redundant cross-query memory accesses, thereby avoiding per-query online shuffling. In addition, we develop a batch-query protocol that exploits higher parallelism to reduce both round complexity and the circuit's critical path. Our experiments show that HNSecW preserves the high accuracy of plaintext HNSW ($\geq 90\%$ recall) and delivers over a $20\times$ efficiency improvement over prior work on million-scale datasets.

Abstract: Large Language Model (LLM) agents have enabled automated code evolution for algorithm discovery. However, current recursive self-improvement (RSI) research for sequential recommendation faces a dilemma: general-purpose frameworks lack domain-specific feedback, while recommender-specific frameworks rely on linear evolution. Crucially, naively integrating them fails when applied to Generative Recommendation (GR). Unlike end-to-end models, GR relies on a module-cascaded pipeline (tokenization, pre-training, and alignment). In this paradigm, existing evolution methods suffer from: i) cost blindness, squandering budgets by ignoring stage-asymmetric retraining overheads; and ii) short-sighted evaluation, discarding promising upstream innovations that temporarily regress metrics before downstream adaptation. To bridge these gaps, we propose TGR-EvoDAG, a multi-agent code evolution framework tailored for GR. TGR-EvoDAG models code evolution as a Directed Acyclic Graph (DAG) to maintain multiple evolution branches and merge complementary innovations, unifying advanced search with rich GR diagnostic feedback. We further introduce stage-aware cost planning to optimize GPU-hour budgets, and active downstream co-adaptation to navigate temporary regressions and unlock cross-stage gains. On public Amazon datasets, TGR-EvoDAG enables the first autonomous evolution of a complete GR pipeline, improving offline metrics by up to 37.4% and Agent4Rec simulated user satisfaction by 19.7% over the strongest baseline. Furthermore, in industrial deployment on Tencent's large-scale commercial media platform, the evolved GR achieves a 12.16% relative improvement in offline Hit@1 and a statistically significant 0.40% increase in online click-through rate (CTR).

Abstract: Large language models (LLMs) increasingly operate in multi-agent systems where they collaborate by responding to others' choices, creating endogenous feedback loops with vulnerability to social influence, herding behavior and runaway cascades. Following the design of the classic Music Lab experiment, we tested this vulnerability in large populations of LLM agents that recommended unfamiliar songs, with or without exposure to cumulative recommendation counts. Exposure led to runaway cascades that generated extreme and unpredictable collective recommendations increasingly divergent from a song's intrinsic appeal. Compared to humans, LLM populations were generally far more susceptible to herding behavior, and unlike humans, LLM cascades were less able to self-correct. This vulnerability has implications for multi-agent collaboration in recommender systems, automated evaluation, complex problem-solving, and unauthorized conspiracies.

Abstract: Fairness-aware recommender systems often encode one fairness objective during training or apply a fixed re-ranking rule after scoring. In deployment, however, the appropriate fairness action can vary across requests because the candidate pool, item metadata, exposure debt, user preference history, and utility budget all change at runtime. We propose FAIRGate, a model-agnostic delivery controller that places an agentic policy-selection layer between a base ranker and the final top-K list. The agent does not define fairness objectives from scratch. It chooses from a predefined policy library covering long-tail exposure, item-group exposure balance, concentration control, calibration protection, and abstention. A deterministic executor then applies the selected policy through constrained repair, and a validator accepts the result only when relevance, feasibility, metadata, calibration, and audit constraints are satisfied. The final language-model component receives only the verified audit trace and produces evidence-linked narration after the ranking decision is fixed. This design gives the agent runtime responsibility while keeping fairness enforcement measurable, bounded, and replayable.

Abstract: Personalized reward models require user-specific preference comparisons, but collecting such feedback is costly and provides only limited supervision for each user. At the same time, large pools of responses can be generated without user identities or preference labels, yet existing personalized reward models leave this data unused. Directly learning from these responses is difficult because pairwise objectives constrain relative scores within a prompt, whereas pseudo-labeling individual responses requires scores that are comparable across prompts for the same user. We propose \textbf{PerSPL} (Personalized Semi-supervised Preference Learning), a backbone-agnostic framework that uses pair-mean score centering to stabilize this cross-prompt score coordinate and iteratively turns unlabeled responses into user-specific training signals. Against seven representative baselines across four benchmarks---PersonalLLM, HelpSteer2, Reddit TLDR, and Chatbot Arena---PerSPL achieves the highest reported mean accuracy among the evaluated methods for both seen and unseen users. The two latter benchmarks provide real-user preference comparisons in summarization and open-domain chatbot interaction, respectively. These results indicate that unlabeled responses can provide additional supervision for personalized reward modeling when per-user preference labels are scarce.

Abstract: Modeling lifetime user behavior sequences is essential for capturing long-range preferences in industrial recommendation, yet 100K-length user sequences make target-aware ranking prohibitively expensive. Search-based methods that retrieve from raw sequences incur candidate-dependent costs that grow with sequence length, whereas latent compression methods reduce computation at the expense of item-level addressability. We propose ECHO (Efficient Cascade with Homogeneous Operators), a framework that balances efficiency and fine-grained retrieval by retaining a subset of original event records per user via a Distilled Selector (DS) and retrieving a smaller set of target-relevant records per candidate via a Target-aware Fine-grained Retriever (TFR). DS is supervised by a candidate-aware full-history teacher via per-target distillation, enabling target-aware user-level selection without incurring full-history computation for every candidate. The retriever and target-attention scorer share the teacher's learned relevance parameters, ensuring consistent relevance estimation across the cascade. At a sequence length of L = 100K, ECHO reduces online sequence-branch FLOPs by approximately 160x compared with a representative end-to-end model. Three deployment optimizations, including common user inputs, compact record encoding, and key-side projection absorption, further reduce memory, storage, and data-movement overhead. Experiments on KuaiRand-1K and an industrial-scale short-video platform demonstrate relative GAUC improvements of 0.113%-1.520% over state-of-the-art (SOTA) baselines across multiple tasks. Furthermore, online A/B testing shows a 0.613% improvement in app usage time.

Abstract: Existing continuous-time graph learning methods rely heavily on complex neural architectures to encode ever-expanding interaction histories. However, the future conditional distribution induced by complex histories can often be characterized by a low-dimensional predictive state manifold, implying that only a few predictive degrees of freedom are essential for future interactions. We therefore argue that continuous-time graph learning should identify low-dimensional states sufficient for future prediction rather than reconstruct the full history. To make this perspective measurable, we define and independently estimate the intrinsic predictive dimension, which quantifies the effective predictive degrees of freedom required by a given task. Specifically, we remove the history-independent marginal from the history-conditioned future distribution, decompose the resulting predictive residual operator, and estimate the minimum effective rank needed to preserve the predictive power of the full history. Experiments across datasets show that model performance saturation is largely determined by the intrinsic predictive dimension. Based on this finding, we propose PreDyG, a lightweight model whose capacity is matched to the intrinsic predictive dimension. PreDyG achieves superior predictive performance with only 0.07\% of the parameters of a representative complex baseline.

Abstract: Recommender Systems (RSs) can be evaluated for many aspects, two of them being recommendation relevance and fairness. These two aspects can be evaluated either in isolation, using a measure for each aspect; or jointly with a single`relevance-aware' fairness measure that quantifies fairness with respect to relevance. Yet, neither approach provides a simultaneous evaluation of both relevance and fairness: the first approach may result in two different rankings of models, each corresponding to an evaluation aspect; the second approach often provides an imbalanced evaluation of the two aspects. To address this gap, we propose Distance to Pareto Frontier (DPFR), a new approach to jointly evaluate RS fairness and relevance. DPFR is based on a Pareto frontier of relevance and fairness scores from existing measures. The frontier is generated from user-item interaction data and represents various Pareto-optimal trade-offs between the two aspects. Thus, the distance from a point in the frontier provides a joint score of fairness and relevance. As DPFR can be used with existing measures, it is modular and intuitive. Our experiments with 4 recommenders and 3 fairness-aware re-rankers on 6 real-world datasets reveal inconsistent associations between existing measures and DPFR, showing that they are not reliable proxies for evaluations based on Pareto optimality. We further examine the robustness of DPFR as a joint evaluation approach for fairness and relevance by conducting experiments with artificial recommendations and with various fairness-relevance trade-offs, as well as by analysing DPFR's agreement to simple measures of fairness and relevance. These analyses show that DPFR is more robust than existing joint evaluation approaches. Finally, we provide guidelines for evaluating fairness and relevance jointly with DPFR. Our code: https://github.com/theresiavr/robustness-DPFR-recsys-evaluation.

Abstract: Large-scale search and recommendation systems derive much of their retrieval quality from rich, heterogeneous features, yet existing generative retrieval (GR) methods struggle to exploit such feature ecosystems efficiently. Static feature injection either encodes all available signals or places them according to predetermined schemes, increasing serving cost and introducing irrelevant or conflicting evidence. More fundamentally, existing methods cannot adaptively determine when a feature is needed, which feature to select, or where it should influence hierarchical Semantic ID (SID) generation. We propose Agentic Generative Retrieval (AGR), a tool-augmented paradigm that interleaves SID generation with on-demand feature acquisition. Conditioned on the request and decoded SID prefix, AGR decides whether to invoke a tool, which feature to acquire, and where its observation should influence subsequent decoding, thereby steering generation toward feature-consistent SID branches. To supervise this behavior, we construct tool-augmented trajectories based on feature applicability and information gain along the SID hierarchy, and train the model through SID pretraining followed by tool-augmented fine-tuning. Experiments on an industrial dataset and three public benchmarks, together with an online A/B test, consistently demonstrate the effectiveness of AGR.

Abstract: Given a vertex $v$ in a large bipartite graph $G$, can one quickly find a small densely connected region near $v$? This is the algorithmic core of many fraud detection, community detection, and recommender systems. One would like a local algorithm, i.e., an algorithm that probes $G$ in only a small number of locations to find such a region. Our main contribution is the first constant-factor local approximation, independent of the maximum degree of the graph, for the classical notion of bipartite density: if the sides of a bipartite subgraph are $S$ and $T$, its density is $d(S,T):=e(S,T)/\sqrt{|S||T|}$, where $e(S,T)$ is the number of edges between $S$ and $T$. We consider cohesive dense subgraphs, since every dense subgraph contains a cohesive core of at least the same density. For such a cohesive dense subgraph $S\mathbin{\dot\cup}T$ with at most $k$ vertices per side, our algorithm, starting from any vertex in $S\mathbin{\dot\cup}T$, uses $O_ε(k^{4+ε})$ neighbor queries to find a subgraph $X\mathbin{\dot\cup}Y$ of density $Ω_ε(d(S,T))$ on $O_ε(k^{2+ε})$ vertices, for every fixed $ε>0$. Our result improves upon the result of Andersen (2010), whose approximation guarantee depends logarithmically on the maximum degree. We complement our positive result with information-theoretic and computational limitations showing that the quadratic dependence on $k$ in the output size is essentially unavoidable for constant-factor approximation. Our main technical innovation is an adaptation of the Lovász-Simonovits curve to the bipartite density problem, based on weak-$\ell_p$ norm bounds.

Abstract: Arrays of modern NVMe SSDs are approaching memory in terms of bandwidth while offering much lower cost per gigabyte. Hence, approximate nearest neighbor search (ANNS) for billion-scale vectors in applications such as recommendation, search, and retrieval-augmented generation increasingly runs on SSD-based systems. Unfortunately, due to the limited update throughput of existing SSD-based solutions, the indexed corpus cannot keep pace with continuous modifications to data and embedding models, thus hindering broader deployment across more scenarios and larger scales. The bottleneck lies in CPU-intensive update paths, where foreground insertions incur frequent SSD accesses and background reassignments repeatedly relocate each vector, making high update throughput difficult to achieve even by exhausting available hardware. In this paper, using buffered updates and batched local rebuilds, we develop PrairieFire with cascaded adjustments, concurrency control, and CPU-GPU-SSD cooperation. These co-designs avoid frequent SSD accesses in foreground updates and heavy background reassignments, achieving high update throughput while maintaining excellent search performance. Evaluations show that, compared with state-of-the-art systems, PrairieFire can sustain approximately 3-35x update throughput while maintaining approximately 3-5x search throughput simultaneously. It also achieves approximately 1.5-8.7x cost efficiency. PrairieFire has been deployed in RedNote's production environments for two months, improving business metrics by higher freshness without frequent index rebuilds.

Abstract: We study how to evaluate $K$ target policies simultaneously in stochastic contextual bandits with linear reward structure. Given a known $d$-dimensional feature map and reward-free access to policy queries, we propose PRISMA (Policy evaluation through Regularized Information Sharing and Mixture Allocation), an algorithm that estimates every target policy's value to accuracy $\varepsilon$ with probability at least $1-δ$. The algorithm uses $\widetilde{\mathcal{O}}\!\left( \frac{(L^2S^2+β^2)(1\veeρ^\star)}{\varepsilon^2} \cdot \log(K) \right)$ context draws and reward observations, where $ρ^\star\le d$ is an instance-dependent measure of the difficulty of designing a shared sampling distribution. Here, $L$, $S$, and $β$ bound the feature norm, reward parameter norm, and reward noise, respectively, and the suppressed factors are logarithmic in $K$, $d$, and $1/δ$. Our method combines reward-free planning, a regularized sampling design, and a single ridge regression estimator shared across all target policies. An adaptive doubling procedure selects the regularization scale without knowledge of $ρ^\star$. The guarantees require no numerical lower bound on second-moment eigenvalues and have no explicit dependence on the cardinalities of the context and action spaces. Policy-query costs are accounted for separately, with $K$ queries per shared planning context.

Abstract: Disk-based approximate nearest-neighbor search (ANNS) supports retrieval-augmented generation (RAG), search, recommendation, and advertising over corpora that exceed memory capacity. Existing ANNS research often focuses on vector search at small top-K, whereas industrial-scale applications require larger retrieval sets. Fetching the corresponding content from storage remains comparatively underexplored. Our profiling across million- to billion-vector corpora shows that, at large K, content fetch latency can match or exceed vector-search latency and cannot be ignored in end-to-end retrieval. We present ConStore, a content storage engine combining two complementary techniques to reduce end-to-end retrieval latency. Speculative Inline Content Prefetch (SICP) commonly abstracts graph-based and cluster-based ANNS as progressive SSD I/O that refines an in-flight candidate queue. It prefetches content for stable leading candidates under a bounded I/O budget, overlapping content reads with search without changing the final top-K results. Co-Retrieval-Pattern-Aware Grouping (RPAG) uses query logs to place frequently co-retrieved documents in contiguous segments. Workload-Adaptive Segment Reorganization (WASR) adapts this layout to workload shifts by selectively regrouping segments under a bounded rewrite budget. Group Content Fetch (GCF) exploits this layout to coalesce per-document random reads into batched, page-aware I/O requests. SICP works independently of the content layout, while RPAG and GCF reduce fetch cost even without prefetching. Across three multi-hop QA workloads with K from 10 to 1600, ConStore reduces end-to-end retrieval latency over the best per-document-read baseline by up to 48% on DiskANN and 25% on SPANN. Separate experiments on 100M- and billion-vector indexes demonstrate SICP's applicability at larger scales.

Abstract: Transfer learning in stochastic multi-armed bandits is typically studied assuming that prior data forms a fixed offline dataset or is generated by a known source. In practice, historical observations come from a previous adaptive decision process whose policy is unknown. We study this setting for bounded rewards in $[0,1]$. Before the online game starts, the learner receives, for each arm $k$, $N'_k$ observations drawn i.i.d.\ from an unknown distribution $ν'_k$, collected by an arbitrary and possibly data-dependent strategy, so the sample sizes $N'_k$ may be random. Source and target means satisfy $|μ_k-μ'_k|<L_k$ for known constants $(L_k)_k$. We derive an asymptotic problem-dependent lower bound and propose KL-UCB-Transfer, which combines online and historical data and achieves logarithmic regret with an improved leading constant whenever $μ'_a<μ_1-L_a$, and recovers the classical KL-UCB guarantee otherwise. Since bounding the source-target discrepancy is often unrealistic, we relax the assumption to the one-sided condition $μ_k-μ'_k<L_k$ and show that logarithmic regret can survive under misspecification, characterizing three regimes ranging from improvement to a degraded logarithmic bound to a loss of guarantee. Simulations illustrate the practical benefits of exploiting process-generated historical data.

Abstract: Large language models increasingly generate multiple retrieval actions jointly, while downstream systems often provide only one reward for the resulting set. Broadcasting this reward provides a valid score-function gradient estimator for the expected set-level objective, but does not determine a coalition-dependent within-set attribution. We introduce CASI (Credit Assignment via Set Intervention for Agentic Retrieval), which constructs the missing counterfactual information through snapshot-grounded set intervention: generate and retrieve once, hold replayable snapshot components fixed, and evaluate alternative subsets under common conditions. Shapley aggregation converts these marginals into within-set credit, which bounded zero-sum advantage shaping incorporates into policy optimization. For m >= 4, singleton, leave-one-out, and full-set endpoints do not generally determine the Shapley allocation; under additive set values, extra coalition evaluations are redundant. In a five-action sponsored-search setting, exhaustive auction replay over 1,200 requests reveals substantial coalition dependence. Under the training-time evaluator, CASI achieves the highest median cosine similarity to a replay-defined Shapley reference among evaluated credit rules; under a sensitivity backend, CASI does not outperform Singleton. Across five independent training runs, CASI-trained policies achieve higher held-out auction-replay set value than Singleton- or Broadcast-trained policies. In a randomized seven-day production A/B test on 10% of traffic, the CASI-trained policy yields an observed 3.51% lift in the primary advertiser-value metric over production control.

Abstract: We study stochastic contextual bandits with an unknown context-lumpable structure. The $S$ contexts are partitioned into $r$ blocks, and every arm has the same mean reward across contexts in the same block. Under arbitrary context distributions and without separation assumptions, we design an algorithm with expected regret $\widetilde{O}(\sqrt{r(S+K)T})$, matching the lower bound of Lee et al. (2023) up to logarithmic factors and improving their previous upper bound by a factor of $r$. The algorithm operates over dyadic accuracy levels, using a baseline policy and confidence-based early stopping to control exploration. It then performs cross-context elimination: since rewards are identical within a block, one selected arm eliminates all surviving records from that block, requiring at most $r$ selections at each level. Multi-scale exploration assigns finer accuracies to frequent blocks and coarser accuracies to rare blocks, yielding a near-optimal arm for every block while keeping the weighted policy error small.

Abstract: Recent advances in self-improving LLM agents have demonstrated the potential of AI systems that drive their own intelligence. However, many domain-specific applications rely on specialized foundation models that possess strong domain expertise but remain dependent on human-designed improvements. Extending recursive self-improvement to these models presents a key challenge: as a specialist improves, the strategies used to advance it must also adapt. To address this challenge, we propose Heterogeneous Agentic CoEvolution (HACE), a framework that jointly evolves specialized foundation models and LLM agents through reciprocal learning. HACE improves the specialist through supervised learning while training the agent with on-policy reinforced self-training, guided by verified performance gains on the current specialist. The agent's exploration, in turn, informs subsequent specialist training, enabling both models to adapt to each other's progress. To sustain this reciprocal learning, HACE further employs evolutionary search to optimize their interaction harness, adapting how the models exchange information and select training experiences as their capabilities evolve. We evaluate HACE across three challenging domains: recommendation systems, protein function prediction in biology, and crystal stability prediction in material discovery. We show that HACE outperforms the evaluated self-improving methods on the primary task metrics across all three domains. These results shed light on heterogeneous coevolution as a universal approach to improving specialized foundation models through reciprocal interaction with a learning agent.

Abstract: In sequential learning, a group of agents sequentially receives a noisy private signal and observes the predictions of prior agents to estimate a hidden ground truth. While extensively studied for binary signals, we formalize and analyze sequential learning when the hidden ground truth is an ordering of $n$ items, and both private signals and inferences involve pairwise comparisons. A key focus in sequential learning is herding, where agents disregard their private signals to follow the crowd. We show that, while local herding in our model may lead agents to make incorrect pairwise inferences, the global hidden order can still be approximated. Specifically, by aggregating all pairwise comparison outputs, the highest-ranked items can be retrieved. We investigate top-$k$ retrieval across three scenarios: (1) passive aggregation, where information is gathered from the social learning process without intervention; (2) strategic selection, where agents choose their pairwise comparisons dynamically; and (3) aggregation with stubborn nodes, where a very small set of stubborn agents, who base their inferences solely on their private signals, join the mechanism. With query complexity as the number of agents needed for a high-probability estimation, we propose top-$k$ retrieval algorithms with query complexities sublinear in $O(n^2)$, the total number of pairs.

Abstract: Multimodal recommender systems effectively alleviate data sparsity and cold-start problems by integrating heterogeneous information, yet complex cross-modal representations and model decision-making processes also introduce new interpretability challenges. Despite the rapid growth of research on the interpretability of multimodal recommendation in recent years, relevant methods and evaluation criteria remain relatively fragmented, lacking a unified conceptual framework. To address this gap, this paper categorizes existing methods into modality explanation and post-hoc explanation based on the timing and manner of explanation generation, and further constructs a progressive system for each category. Building on this foundation, this paper synthesizes four categories of evaluation metrics from both objective quantitative and subjective perceptual perspectives, and reviews relevant datasets along the dimensions of application domain and data source. Through a comprehensive analysis of existing research, we further propose future research directions. This paper aims to establish a unified analytical perspective across multimodal representation, recommendation decision-making, and user preferences, providing a research foundation for building transparent and trustworthy multimodal recommender systems.

Abstract: In off-policy evaluation, merging decision-time histories that matter differently for a specified policy value or comparison can make that evaluation unidentified, even with observed actions and outcomes and known behavior and target policies. We characterize which distinctions must be preserved: identification holds if and only if the action-specific evaluation weights are constant within each released record, and the coarsest exact release groups histories by these weights. For normalized outcomes, the largest sharp width induced by the release equals a behavior-weighted within-record radius of these weights. The same radius becomes a release-design objective when each record may be refined into at most $m$ categories. The optimum has an exact $m$-center characterization. When behavior does not vary within a released record, we give a finite-data repair algorithm whose maximum evaluation-weight radius is at most twice the finite-data optimum. For a fixed action-outcome distribution within a released record, a distinct sharp-identification problem remains. Under the same condition, the action-specific outcome distributions are identified, and optimizing over their couplings gives the sharp interval. We construct an AIPW envelope containing this sharp interval; it matches the interval for all such outcome distributions exactly when the extreme within-record action-coefficient vectors take at most two values in each action coordinate. In 24 KuaiRec tasks, the repair algorithm yields a 93% median reduction in mean pairwise sharp width at $m=64$ relative to the unrepaired release.

Abstract: Generative recommendation models based on autoregressive decoding have been adopted as alternatives to classical ranking-based systems. While their modeling flexibility has been widely studied, their resilience to data poisoning remains largely unexplored. In this work, we ask: how susceptible are generative recommenders to conditional targeted poisoning? Under a realistic setting where an adversary can inject limited malicious interactions and query a recommendation API as a black box, our empirical study shows that generative recommenders are vulnerable: injecting fewer than 1% synthetic interactions significantly increases the likelihood of promoting malicious targets under the trigger. To characterize this vulnerability, we analyze the synthesis of malicious interactions from two perspectives: (i) decoding alignment and (ii) prefix popularity. We observe that short, context-aligned prefixes that are both common under the model's induced transition dynamics and strongly linked to a frequent trigger consistently amplify targeted promotion. Motivated by this analysis, we develop AlignPoison, a black-box framework that discovers and samples context-aligned prefixes from API outputs. We conduct experiments on benchmarks in unimodal and multimodal generative recommendation settings. Results show that these prefixes outperform competitive baselines, improving ASR lift by up to 16.3 percentage points and highlighting a decoding-dependent, context-sensitive poisoning vulnerability in generative recommendation models.

Abstract: Universal multimodal retrieval typically encodes a query once and ranks independently indexed items by embedding similarity. This design supports efficient search, but leaves the query representation unchanged even when retrieved items could help clarify the information need. We introduce the AutoRegressive Retriever (ARR), a multimodal retrieval model that learns both to select informative items and to use their content to refine subsequent retrieval. ARR alternates between retrieving an item and updating the query embedding, then uses the final embedding to rank the collection. Supervised fine-tuning teaches the encoder to use feedback through stepwise contrastive supervision. Reinforcement learning treats feedback items as actions and optimizes their selection using the final reciprocal rank of a relevant item. A query-side adapter enables this optimization against a fixed item index. ARR demonstrates strong retrieval performance on both in-domain and zero-shot benchmarks, outperforming the compared baselines on average. Further analyses show that feedback improves retrieval at inference time and that training with feedback also improves the initial query embedding, before any item is observed.

Abstract: We study ranking models by population-average utility from pairwise comparisons when preferences vary across users and tasks. Prior work shows that a single comparison per user can be insufficient to identify the alternative with the highest average utility, even with arbitrarily many users (Golz et al., 2025). We investigate how many repeated comparisons within each user-task context are necessary and sufficient for ranking recovery. Under a heterogeneous Bradley-Terry model with fixed inverse temperature, we start with a naive MLE-based algorithm that requires $Ω(1/Δ^2)$ repeated comparisons per context to ensure ranking recovery. We then present two MLE-based variants and a randomized Russian Roulette-style algorithm that recover the ranking using $O(\log(1/Δ))$ repeated comparisons per context, and we prove that this logarithmic dependence is optimal. Despite this worst-case requirement, our Russian Roulette algorithm uses only $O(1)$ comparisons per context in expectation. Synthetic experiments and semi-synthetic experiments based on Arena data compare the four algorithms in settings with varying levels of preference heterogeneity and under varying context distributions.

Abstract: Long-history recommenders often compress each user's history into a compact, candidate-independent memory that is cached and reused to score large candidate pools. We show that real user histories exhibit multi-scale semantic structure, with short-lived intent, medium-term interests, and long-term preferences coexisting in one sequence, and that monolithic cached memories preserve these scales unevenly: linear probes recover recent and mid-range content far worse than long-range content. We call this failure mode \textit{temporal aliasing}. We propose \textbf{MARS}, a multi-resolution user memory that writes the full history into recurrent state tracks anchored to different half-lives, and a sparse routing reader that materializes compact seed memories by selecting the relevant temporal resolutions for each seed, preserving fixed-size candidate scoring. MARS outperforms strong baselines on three public datasets, with gains that grow with history length. Component-matched ablations with paired tests show that temporal diversity and selective routing each contribute beyond what hard-window memories or added capacity provide. The advantage of MARS over its interface-matched baseline also widens after within-user behavioral shifts, at about $1.02\times$ that baseline's warm-cache serving latency for $1{,}000$ candidates per user.

Abstract: Semantic ID-based generative recommendation represents each item as a sequence of discrete tokens, enabling structured modeling of item semantics. A critical challenge is constructing semantic IDs that are both semantically expressive and computationally efficient. While recent approaches favor complex learned quantization, simple hashing-based methods such as SimHash are widely regarded as fundamentally inferior. In this work, we challenge this consensus by showing that the apparent performance gap does not stem from inherent limitations of hashing, but rather from a structural mismatch with autoregressive decoding, coupled with the inevitable information loss during rigid discretization. Based on this insight, we propose FLASH, a two-stage framework that revitalizes training-free SimHash tokenization through parallel decoding and explicit semantic alignment. Despite its simplicity, FLASH achieves state-of-the-art performance across multiple datasets without requiring any tokenizer training, while exhibiting stronger generalization in cold-start scenarios. Notably, we demonstrate that semantic alignment acts as a universally effective mechanism across diverse paradigms. Our findings suggest that, with compatible decoding and semantic grounding, simple and efficient tokenizers can achieve performance comparable to complex learned counterparts in generative recommendation. Our code is available at https://github.com/KevinC2015/Flash.

Abstract: We study contextual bandits in which a surrogate is observed after the action but before the learner decides whether to acquire the primary outcome that defines action value and regret. The value of acquiring the primary outcome depends on both decision relevance (how much the current outcome matters for comparing policies) and the residual uncertainty after observing the surrogate. The Audited Surrogate Bandit (ASB) learns a contextual policy while allocating a budget of $B$ primary-outcome acquisitions over $T$ rounds. ASB sets a pre-surrogate acquisition level from current decision relevance and, after observing the surrogate, redistributes that level using an estimate of that residual uncertainty. For a finite class of $N$ policies over $K$ actions, ASB incurs $\widetilde O[\sqrt{KT\log N}\{1+\sqrt{T/B}\}]$ regret relative to the best policy in the class. In a two-action family where the surrogate does not reveal the better action, a learner that observes the surrogate before deciding whether to acquire can achieve bounded regret, whereas any learner that must decide before seeing the surrogate incurs $Ω(T/B)$ worst-case regret under the same budget. Synthetic experiments show that both acquisition factors matter: ASB has lower regret than variants using only decision relevance or only residual uncertainty. On a KuaiRec benchmark of user-video interactions, the regret gap relative to relevance-only acquisition widens and then narrows as the budget grows.

Abstract: Sequential recommenders can generate synthetic interaction sequences and retrain on the augmented corpus in a recursive self-improvement loop. To limit error accumulation, current methods verify each generated sequence remains predictive of the user's real interactions and discard those that drift away from it. Verification does not, however, determine which verified sequences should train the next model. With every verified sequence used for training, source sequences yielding more verified sequences or longer continuations have more influence, although neither quantity indicates how much those sequences will help the next model. We formulate the decision of which verified sequences are used to train the next model as \emph{post-verification acquisition} and introduce {\bf Disagreement-Aware Recursive Self-Improving Recommendation (DA-RSIR)}. DA-RSIR caps each source sequence's contribution and ranks its verified sequences by how much the model's predictions disagree over their augmented interactions. It uses a score derived from Bayesian Active Learning by Disagreement (BALD) and estimated with Monte Carlo (MC) dropout. DA-RSIR requires no extra labels, teacher model, or quality scorer. Across four datasets, three recommender models, and two metrics, it improves on the retain-all approach in all $24$ comparisons and attains the highest mean in $23$ of $24$ overall; the aggregate improvement is statistically significant on both metrics. A single DA-RSIR round exceeds the retain-all approach's best gain over five recursive rounds. These findings establish post-verification acquisition as a separate control point in recursive self-improvement, separating which sequences pass verification from which verified sequences are used to train the next model.

Abstract: Large language models (LLMs) can infer user preferences from interaction histories and reviews, yet the rationales they generate may not reflect the information actually used for recommendation. A preference claim may be weakly supported by its selected evidence, or may have little effect on the final ranking. We refer to these two failures as the grounding-influence gap. We introduce PROVE-REC, a general framework for verifiable preference reasoning in LLM-based recommendation. Pass A converts the complete pre-target history into a compact preference proof consisting of positive and avoidance claims linked to selected evidence entries. Pass B predicts the next item using only the proof and its selected evidence, preventing the recommender from bypassing the reasoning path. To verify evidence-to-proof grounding, we compare the effect of masking selected evidence with masking a comparable control entry. To verify proof-to-recommendation influence, we remove a preference claim and measure the resulting decrease in the target item's ranking margin. A ranking-preservation objective further retains useful information from the complete history. Comprehensive experiments on wide-ranging real-world datasets demonstrate that PROVE-REC consistently outperforms strong sequential, generative, and LLM-enhanced baselines, with improvements of up to 7.45%. Controlled ablations confirm the effectiveness of the two-pass architecture and verification objectives. Moreover, PROVE-REC produces claims that are more strongly grounded in historical evidence and more influential to recommendation while preserving ranking quality.

Abstract: In structured rating prediction, does a personal AI use historical item-rating associations, or mainly the user's rating tendencies? We audit this distinction by permuting historical ratings within each user while preserving the exact rating distribution, item support, and metadata. We combine this control with full history, native memory extraction, and matched numerical readers in a publicly frozen evaluation of 400 held-out user profiles and 6,160 target ratings across Coat and MovieLens. On Coat, the tested Qwen-written Mem0 pipeline increases user-macro mean absolute error relative to full history by 0.084 for Qwen and 0.149 for Phi; both family-adjusted bootstrap intervals exclude zero. Correct historical assignments help both readers on Coat, but the corresponding MovieLens effects are smaller and inconclusive after adjustment. A history-only ridge reader outperforms Qwen in both domains and Phi on MovieLens, while the Coat Phi comparison is unresolved. All 2,800 reader calls, including 150 invalid outputs, are retained under a fixed fallback rule. A separate implementation verifies inputs, metrics, and all ten primary contrasts. The contribution is a reproducible diagnostic study showing why extraction, association use, output reliability, and reader choice require separate evaluation.

Abstract: Personalization encoders compress evolving interaction histories into preference states used to rank items or condition text generation. A task head operating only on this state can miss useful evidence that remains in the frozen encoder's cached representations for individual timesteps. We study this recoverability gap and propose REPAIR, which compares cached representations with the current preference state in a compact learned coordinate space. It resolves corrective evidence over extended history, recent interactions, and localized bursts. It then selects which patterns at which timesteps contribute and adds their aggregate correction to the state before the task head. Encoder-host repair reuses representations from the existing forward computation without re-encoding the history. Across MovieLens, PENS, MIND, and Amazon Reviews 2023, training only REPAIR improves MRR and nDCG@10 for all twelve representative recommendation hosts while both encoder and task head remain frozen. Head-only finetuning of the same hosts yields smaller gains. For example, Mamba4Rec on MovieLens gains 3.96 MRR points, compared with 0.19 from head-only finetuning. Rank and temporal diagnostics support a compact, host-dependent corrective structure. In personalized generation, IMPerSumm improves the two reported weighted PerSEval variants, which assess responsiveness to user preference, by up to 25.23%. These results support post-compression state correction and distinguish the availability of preference evidence from its downstream use.

Abstract: Generative retrieval has emerged as a general retrieval paradigm, representing items with discrete Semantic IDs (SIDs) and retrieving them through autoregressive identifier generation. When SIDs are constructed with residual quantization (RQ), standard retrieval training supervises only the selected codes and discards the residual trajectories that produce them. The same hard code can nevertheless arise from different preferences over competing codewords, while the residual trajectory also contains information about subsequent quantization decisions. As a result, hard SID supervision collapses distinct quantization behaviors into identical targets and leaves information available during indexing unused in retrieval training. We introduce ResTD, a Residual Trajectory Distillation framework that transfers this discarded indexing information into retrieval training. Treating the frozen RQ indexer as a process teacher, it distills residual-induced codeword preferences into SID-decoding states. This supervision recovers distinctions hidden by hard assignments and allows earlier decoder states to capture information about subsequent quantization decisions before the corresponding SID suffix is generated. In this way, richer information from SID construction is incorporated into retrieval learning while preserving the original retrieval index and inference procedure. Experiments on multilingual e-commerce retrieval show consistent improvements over strong baselines and matched training controls. Controlled comparisons show that residual-derived targets outperform the tested codebook-only soft targets. Representation probes further show that future codebook preferences become more recoverable from earlier decoder states. ResTD can also be readily extended beyond retrieval to generative recommendation. Code is available at: https://github.com/Nevaeh7/iclr2027_ResTD.git.

Abstract: Knowledge Graph Embedding models have been extensively used to learn representations of entities and relations in Knowledge Graphs for predicting missing links. However, the quality of the learned representations varies a lot across different areas of the graph. If previous research has loosely linked the problem to relation types or degree bias, we show that it is more widespread and it correlates with the degree imbalance of the entities in test triples. In particular, the prediction of a target entity that has a degree much smaller than the degree of the anchor entity is extremely problematic. This is critical in recommender systems and other use cases, where these triples represent important corner cases. To address this issue, we propose an inference-time latent search optimization method capable of significantly improving model predictions on the most imbalanced triples. Built on top of a pre-trained model, it explores the embedding space at evaluation time, blending known and out-of-band information to mitigate the degree imbalance bias. We show the value of our approach on imbalanced triples from common benchmark datasets, where we outperform conventional methods, opening the door to the successful adoption of Knowledge Graph Embedding models on these critical corner cases.

Abstract: Many recommender systems such as for e-commerce and news platforms aim to provide users with rankings they are likely to interact with. Off-Policy Learning (OPL) of ranking policies enables us to learn new ranking policies using only historical logged data. However, ranking settings make OPL remarkably challenging because their action spaces consist of permutations of unique items, being extremely large. Existing methods primarily use either policy- or regression-based approaches. The policy-based approach, which typically uses importance-weighted policy gradients, can suffer from high variance due to large action spaces. The regression-based approach, on the other hand, estimates the expected reward using conventional machine learning methods, avoiding variance issues but potentially suffering from severe bias. To circumvent these issues of existing methods, we propose a new OPL method for ranking, named Ranking Policy Optimization via Top-$k$ Policy Decomposition (R-POD), which combines the policy- and regression-based approaches in an effective fashion. Specifically, R-POD decomposes a ranking policy into a first-stage policy for selecting top-$k$ actions and a second-stage policy for choosing the bottom actions given the top-$k$ actions. It learns the first-stage policy using a new policy gradient estimator and the second-stage policy via the regression-based approach. This method can substantially reduce variance, since it applies importance weighting only to the top-$k$ actions. We also demonstrate that our policy-gradient estimator for the first-stage policy is unbiased under a conditional pairwise correctness condition, which only requires that the expected reward differences of pairs of rankings sharing the same top-$k$ actions can be estimated correctly.

Abstract: Aligning large language models (LLMs) to diverse user preferences is fundamentally hindered by standard alignment paradigms that optimize for monolithic users. In this work, empirical studies are first used to reveal the existence of a massive, untapped performance headroom for personalized generation through test-time alignment. We demonstrate that personalized generation is uniquely suited for test-time scaling methods like Best-of-N (BoN) because it can be viewed primarily as a candidate matching problem rather than a generator capability bottleneck. While reward models could in principle exploit this headroom, they are poorly calibrated for personalization, and their billion-parameter scale makes scoring large candidate pools prohibitively expensive. To overcome this limitation, we propose a parameter-efficient framework utilizing million-parameter scale multi-layer perceptron (MLP) ranking models. Our personalized ranking model directly reuses the internal embeddings of the base generator with minimal overhead. By scaling train-time data to provide fine-grained personalized preferences, this million-parameter ranking model accurately scores large candidate pools and can seamlessly guide generation to reduce the cost of materializing N candidates. Extensive experiments on nine datasets spanning three personalized generation settings show that our personalized ranking model effectively exploits the discovered headroom, outperforming billion-parameter generalist reward models on every dataset, with under 0.4% of their parameters and four orders of magnitude lower scoring latency.

Abstract: Reinforcement learning from human feedback (RLHF) aligns large language models (LLMs) with human preferences, yet most pipelines learn a single reward model that overlooks individual differences in preferences. Personalized reward models (PRMs) address this by conditioning rewards on user-specific feedback, most commonly through in-context learning (ICL), where a user's historical comparisons are supplied as contextual preference pairs. However, we identify a key limitation of ICL-based PRMs: they fail to capture the preference relations conveyed by contextual pairs. To address this, we propose Preference-Aligned Test-Time Training (P-TTT), which explicitly encodes these relations into user-specific fast weights for personalized reward prediction. P-TTT introduces sequence-level update and apply operations to match the response-level granularity of preference feedback, together with a preference-aligned objective that directly uses pairwise preference relations to guide fast-weight adaptation. Notably, P-TTT is simple to implement and computationally efficient, updating fast weights within a single forward pass without inference-time backpropagation. Extensive experiments show that P-TTT more effectively captures historical preference relations and outperforms state-of-the-art methods by a large margin.

Abstract: In large language model-based recommender systems, identity cues embedded in prompts can steer recommendations toward group-level patterns even when the underlying behavioral evidence remains unchanged. We introduce PromptShift, an interpretable, training-free framework for quantifying and mitigating such identity-cue preference drift. We define Drift as the divergence, in both item membership and ranking order, between a recommendation list generated under an identity-cued prompt and the reference list produced from the same user's interaction history alone. SliceShift then measures the extent to which a cued list gravitates, relative to the history-only reference, toward items that are more popular within the cued slice than among the global user population. Beyond conventional accuracy, we propose DifHitRate, a difficulty-weighted hit metric that credits only relevant items, assigning higher credit to hits that are less popular within the cued slice and ranked higher in the list. All components are supported by an identity-slice-by-item table constructed from positive interactions, which further enables an adaptive post-hoc reranking strategy: the reranker interpolates between the original LLM ranking and inverse slice-popularity, with personalized interpolation weight. Experiments on two datasets with three LLMs show that identity-cued prompts incur higher mean Drift than identity-free paraphrase controls, an effect beyond generic wording sensitivity, and that SliceShift is positive across all six dataset-model settings. PromptShift consistently reduces both Drift and SliceShift, lowering macro-mean SliceShift by 62.42%, while improving DifHitRate, HitRate and MRR. These results demonstrate that identity-cue preference drift can be measured and mitigated without any model training, albeit with a modest, metric-dependent utility cost.

Abstract: We consider contextual binary prediction with i.i.d. contexts from an unknown distribution and adaptively chosen losses. We show that a simple Follow-the-Perturbed-Leader algorithm with Gaussian perturbation for each observed context achieves the optimal $\widetilde O(\sqrt{T\log N})$ expected regret for a class of $N$ experts, while requiring one optimization-oracle call per round and no explicit enumeration of the class. For an infinite hypothesis class $\mathcal H$, the algorithm attains $\widetilde O(\sqrt{T\operatorname{VC}(\mathcal H)})$ regret. This resolves an open problem posed by Lazaric and Munos (2012), showing that hybrid classification is computationally as easy as statistical learning.

Abstract: A pricing page can walk the posted price up to the last amount a buyer still accepts, a recommender can hold back a better item for a barely acceptable promoted one, and a classifier can shift its boundary once applicants change their features. The system predicts the response and then picks the menu that serves its own objective, so the surplus above the user's cutoff is taken. Playing the single best action publishes that cutoff, while noise on actions the user would never take throws away payoff and teaches the platform that a worse menu is still acceptable. We study unpredictable near-optimal policies (UNOP), which mix uniformly on near-best actions that remain individually rational. The mixture is a commitment about the response. On a finite price grid, when the best sure-demand price strictly out-earns the randomized band, a seller who already knows the curve posts below the band, and the purchase that occurs is deterministic. Knowing that curve is not the same as predicting the next draw. The mixture can be learned and the optimizer can match its best response, while the user's payoff stays higher because the mixture changes which action is targeted. In pricing and in policy-aware recommendation this leaves more surplus than greedy play when the platform optimizes against the curve and more than one action is acceptable. The gain goes away under quality ranking, a singleton near-optimal set, a wrong utility estimate, or a short-horizon explorer. That is also where mixing should be turned off if the other side is trying to cooperate.

Abstract: Transformers have redefined sequential recommendation by effectively modeling dynamic user behaviors and long-range dependencies. However, they remain inherently inefficient: standard architectures operate at a fixed rate, allocating comparable computation to every item in a user's history regardless of its information content. This leads to prohibitive computational overhead on long sequences and increased sensitivity to behavioral noise. To address this, practitioners often resort to lossy sequence compression, staged modeling, or truncation. This limits the model's ability to leverage the full context of long histories during inference. Inspired by the recent success of Byte Latent Transformer, we propose DP-Rec, a dynamic latent patching architecture for recommendation. DP-Rec shifts from item-level modeling to patch-level modeling by segmenting interaction sequences using contrastive entropy surprise to identify informative behavioral boundaries. A lightweight patch encoder compresses these temporally contextualized segments into a reduced set of dynamic latent behavior vectors, which are then processed by a larger latent transformer and decoded for next-item prediction. Extensive experiments show that, under constrained computational budgets, DP-Rec scales effectively to long sequences and achieves a superior efficiency-accuracy trade-off over both non-compressed and fixed-size compression baselines.

Abstract: We formalize slate recommendation as a randomized score learner followed by deterministic selection. First, an appropriately scoped differential-privacy guarantee passes through selection and its audit trace by post-processing. End-to-end privacy holds only when selector inputs are public or independent, previous private outputs, or separately privacy-accounted; fixing raw state or candidate information instead yields only a conditional guarantee. Second, we derive a logged margin certificate: bounded score-induced objective movement below half the smallest greedy decision margin guarantees that the ordered slate is unchanged. Controlled fixed-margin tests show near-linear exponent scaling, with an empirical slope of $-0.220$ (95% CI $[-0.231,-0.210]$) against the independent-noise reference $-1/4$. Real-anchor experiments on OULAD, MovieLens-25M, and Amazon Musical Instruments show that greater anchor weight reduces score-noise-induced ranking churn. OULAD and EdNet certificate checks validate the implementation of the logged inequality, while closed-loop simulations show bounded target drift and setting-dependent downstream utility. The contribution is therefore a privacy-scope contract and a certifiable score-to-slate stability mechanism, not a universal utility claim.

Abstract: Recent advances in generative modeling have reshaped recommender systems by formulating recommendation as a next-item generation problem. Existing retrieval approaches primarily follow two paradigms: user-to-item (U2I) methods represent user context using one or a few deterministic embeddings, which limits the ability to capture diverse and multi-mode interests, while semantic-ID-based autoregressive (SID-AR) methods model more expressive distributions but suffer from quantization errors and the low throughput of sequential decoding. To address these limitations, we propose X-Rec to directly learn the recommendation distribution in the continuous item embedding space through flow matching and generate embedding triggers for approximate nearest neighbor retrieval. X-Rec incorporates three key designs to make this formulation effective and efficient. First, we introduce anchor conditioning to decompose generation into coarse semantic-region selection and fine-grained refinement. Second, we adopt Riemannian flow matching to align generative trajectories with the hyperspherical geometry of item embeddings. Third, we design a late-interaction diffusion Transformer that restricts repeated velocity-field estimation to the final Transformer layer. On a streaming benchmark, X-Rec substantially outperforms U2I baselines, matches the retrieval quality of SID-AR methods, and delivers 3.46x higher inference throughput than SID-AR. X-Rec has also been deployed as a new retrieval source for a specific vertical content on TikTok, where two consecutive launches have yielded significant improvements in both vertical engagement (+4.1484%) and general engagement (+0.0111%).

Abstract: We study the problem of computational efficiency for linear bandits in high-dimensional settings with a finite arm set. In linear bandits, the increase in the dimension $d$ of the feature vectors leads to growing computational costs of $O(d^2)$ at each round of update. Traditional sketching-based methods such as SOFUL reduce computation via fixed-size matrix sketching, yet run the risk of incurring vacuous linear regret when the spectral tail of the data is heavy and the sketch size is inadequately selected. To guarantee regret convergence and effectively reduce computational costs, we introduce a clustering mechanism and propose the Cluster Sketch Linear Bandit (CS-LB) algorithm. Our method preserves the full covariance information in each cluster to guarantee robust sublinear regret without spectral-tail vulnerabilities, performs cluster switching by assigning a sentinel for each cluster, and reduces per-round update computation to $O(l^2d)$ via a tunable sketch size $l<d$. Experiments on synthetic datasets demonstrate that our method consistently maintains a favorable trade-off between efficiency and regret.

Abstract: Reproduction studies can produce mixed outcomes. Reported values may differ while the ordering of the compared methods remains the same, a result may hold only under some experimental conditions, or a released implementation may fail to reproduce a result that the model can still reach. The terms repeatability, reproducibility, and replicability describe how a follow-up study relates to the original experiment, but not which parts of the original claim are supported by the new results. We introduce \emph{claim calibration} as a way of stating the strongest claim supported by a follow-up study, together with the conditions under which it holds and the parts that remain untested. We apply this perspective to five original--follow-up paper pairs from recommender-systems research. The cases show that agreement in numerical values, method rankings, statistical results, and overall conclusions does not always coincide, and that follow-up studies often support only part of the original claim. Based on these observations, we propose a Claim Evidence Profile for reporting the original claim, its scope, the reproduction target, the reported results, the calibrated claim, and the parts of the original claim that remain unresolved.

Abstract: Given a user-item graph $G$, a query item $v_q$ and a target item $v_t$, the Swing score $sw(v_q, v_t)$ of the item pair $(v_q, v_t)$ leverages the user-item-user interaction structure to evaluate their similarity. This measure is found to be highly effective in item-to-item (i2i) retrieval task and finds extensive applications in industrial-scale recommender systems. However, existing solutions towards computing Swing scores are either prohibitively expensive due to their quadratic time complexity w.r.t. the item degree, or rely on truncation heuristics that yield unsatisfactory quality, rendering them impractical particularly on graphs with billions of interactions. In this paper, we present ASC and $K$-ASC, two novel and efficient algorithms for approximate and top-$K$ Swing queries, to address the aforementioned limitations. Specifically, these algorithms provide rigorous theoretical guarantees in probabilistic relative and additive errors of Swing values. The basic idea of ASC is to combine two randomized algorithms, GNS and USS, in a simple yet non-trivial way to adaptively process high- and low-degree query items with minimal runtime cost. In particular, $K$-ASC offers practical efficiency and effectiveness for top-$K$ queries through a filter-refinement paradigm with carefully-designed heuristics. Extensive experiments over eight real datasets demonstrate that ASC and $K$-ASC can achieve orders of magnitude speed-up over competitors in terms of computational time while offering the same approximate and top-$K$ query result quality, and in particular, $K$-ASC is highly efficient on massive graphs including the billion-edge Yambda and MAG datasets.

Abstract: Decentralized bandit systems often contain heterogeneous agents: rewards can change at individual agents even when the best action for the network stays the same. These local changes may cancel when rewards are averaged across agents, so the number of local changes $\Stloc$ can be much larger than the number of changes in the best common arm $\Stdec$. We introduce Decision-Relevant Fresh Comparison (DRFC), which uses new, balanced samples from all agents to compare arms at the network level and switches only when fresh global evidence indicates that the common best arm has changed. We prove a high-probability dynamic regret bound with no adaptation term depending on $\Stloc$, and show that every algorithm must still pay for identifying genuine decision switches and propagating them through the communication graph. Under a distinct time-average benchmark, an anytime-valid sliding-window extension handles gradual drift; experiments on synthetic, semi-real, and MovieLens-1M replays show that DRFC ignores decision-irrelevant local changes while the extension avoids false switches.

Abstract: Approximate nearest neighbor search (ANNS) retrieves the most similar vectors to a query vector in high-dimensional space. Label-filtering ANNS (LFANNS) extends ANNS with a label filter that the labels of base vectors must satisfy a set relation (e.g., equality, containment, or overlap) with the query labels. Existing LFANNS indices suffer from inconsistent performance across different filter types and degraded scalability under varying label scale and distribution. In this paper, we define label-stratified similarity graph (LSSG), where edges connect neighboring vectors whose label sets fall within stratified similarity thresholds. To implement LSSG efficiently, we design an incremental insertion algorithm to prune redundant edges in both vector and label spaces, and leverage a MinHash structure to ensure scalability for large-scale labels. We analyze stepwise probabilities under explicit label models and explain why stricter label tiers reduce ineffective in-filtering expansions. Benchmark experiments show that LSSG achieves ideal optimality for equality queries, and 1.06x-92.9x and 1.08x-84.1x faster than the best competing index for containment and overlap, respectively, in query speed with identical accuracy and 0.35x index size.

Abstract: Large language models (LLMs) exhibit strong semantic understanding and preference reasoning capabilities, offering new opportunities for user modeling in recommender systems. Existing LLM-as-Enhancer methods typically distill LLM-derived preference knowledge into lightweight recommenders to avoid costly online LLM inference. However, they often construct distillation knowledge from only one perspective. Prior preference captures users' stable and consistent interests but provides limited guidance for the current decision, whereas posterior preference reveals target-relevant fine-grained interests but may rely excessively on target clues. To address these limitations, we propose P$^3$Rec, a framework that jointly extracts and internalizes complementary prior and posterior preference reasoning knowledge. Specifically, P$^3$Rec first derives target-agnostic prior preferences and target-conditioned posterior preferences from the user side, while further extracting item-centric preference representations from item semantics and predecessor interactions. It then progressively internalizes prior and posterior knowledge into behavioral representations through prior preference absorption and posterior-guided preference distillation. Since the resulting comprehensive preference representation may not always provide an equally decisive retrieval direction, P$^3$Rec further characterizes historical interest dispersion with interest entropy and adaptively calibrates the user representation before contrastive retrieval optimization. In this way, P$^3$Rec achieves more complete preference reasoning while preserving efficient recommendation. Extensive experiments on multiple public datasets demonstrate its effectiveness.

Abstract: While Reinforcement Learning from Human Feedback (RLHF) is the standard paradigm for aligning large language models with human preferences, its effectiveness in pluralistic settings has been called into question. Notably, recent work by Gölz et al. (2025) demonstrated that the \textit{distortion} -- defined as the multiplicative gap between the average user utility of the RLHF policy and the optimal average utility -- can scale exponentially with the Bradley-Terry temperature parameter $β$ when users have heterogeneous preferences. In this work, we present a fine-grained analysis of the distortion of RLHF with reward clipping and demonstrate that such exponential degradation is not a fundamental property of the algorithm but rather a consequence of distribution mismatch between the distribution generating preference data ($μ$) and the KL reference policy ($π_{\mathrm{ref}}$). To this end, we establish tight upper and lower bounds on the distortion of RLHF across multiple regimes of the KL regularization strength. We show that in a representative regime, under the Bradley-Terry model, the distortion is $\tildeΘ(βB + β)$, where $B$ is an upper bound on the log density ratio between $μ$ and $π_{\mathrm{ref}}$. In particular, when there is no distribution mismatch (i.e., $μ= π_{\mathrm{ref}}$), RLHF achieves the optimal distortion of $O(β)$ up to a constant. Our results suggest that, to reasonably maximize average utility with RLHF, it is preferable to use on-policy sampled preference data or to fine-tune before RLHF on data from a source close to $μ$.

Abstract: Generative recommendation (GR) relies on large-beam decoding to generate hundreds of candidate items, creating a new scaling challenge for recurrent linear attention. Existing linear attention serving systems either materialize a full recurrent state for every beam or repeatedly replay shared history, incurring substantial memory and traffic overhead. To address this, we present OneLA, a linear-attention decoding framework that exploits the shared prompt and short divergent suffixes of GR workloads. Specifically, OneLA represents all beam states using a single shared prompt-derived state and compact, append-only records of their divergent transitions. Using this representation, OneLA computes only the state information required at each decoding step, without reconstructing a full recurrent state for every beam. Furthermore, OneLA uses a lightweight ancestry index to track the transition records that make up each beam's history, allowing beams to be updated without moving or copying existing records. A fused GPU kernel further reuses the shared state across beams. Our analysis shows that OneLA achieves 1.54-2.46x end-to-end decode speedups while substantially reducing recurrent-state memory use and data movement.

Abstract: Universal multimodal embedding (UME) increasingly demands encoder's capacity for handling a broad range of tasks and modalities with increased complexity. Prior scaling methods either increase the representation size, retrieval effort, or scales the encoder into a heavy multimodal LLM. Recent works, such as Think-Then-Embed (TTE), explore scaling via reasoning tokens. However, embedding models are hard to scale up: increasing parameters directly tradeoffs for the large training batch size that contrastive learning needs, and retrieval has to be served under tight latency. Moreover, UME tasks are diverse in complexity, where scaling up embedders can bring significant redundant computation. In this work, we propose MOEMB, which instead scales UME along the expert axis through mixture-of-experts (MoE), growing encoder capacity while preserving single-vector, non-autoregressive encoding. Through a systematic study of the design space and training recipes for MoE-based UME, MoEMB sets a new state of the art on both MMEB-V2 and MRMR among models trained on public MMEB-family data: with only 3B active parameters, MoEMB surpasses TTE-based methods with >4x active parameters, using significantly less computes. To further improve the scalability and efficiency, we conduct the first comprehensive study of adaptive computation for MoE-based embedding, spanning diverse strategies across training-based and inference-only methods. Together, these results support expert scaling as an effective and efficient direction for UME, with adaptive computation further improving efficiency for MLLM-based embedding models towards large-scale retrieval and recommendation systems.

Abstract: Retrieval-augmented personalization enables large language models to produce more accurate and preference-aligned outputs using relevant records retrieved from user histories. Personalized language models typically prepend a fixed number of retrieved user records, even when additional history is redundant, harmful, or unrelated to a user's distinctive behavior. We study minimal sufficient personalization: constructing the least costly ordered profile for each input while preserving the utility achievable from a retrieved candidate pool. We introduce ENOUGH, a method that iteratively appends behavioral records or emits STOP to construct profiles with adaptive lengths. Offline, bounded counterfactual search evaluates profile prefixes by jointly considering downstream gains, user specificity, and token costs. The resulting long-horizon targets are distilled into a multi-head value controller with explicit ranking and stopping supervision. At inference, the controller selects and orders records through lightweight decisions, and the frozen generator is invoked once after stopping. Extensive experiments on six personalized tasks demonstrate that ENOUGH consistently outperforms strong heuristic and retrieval-augmented baselines in both effectiveness and efficiency, achieving minimal sufficient profiles that preserve personalization utility while reducing unnecessary context costs.