Large Language Models (LLMs) have been widely adopted in ranking systems such as information retrieval (IR) systems and recommender systems (RSs). To alleviate the latency of auto-regressive decoding, some studies explore the single (first) token decoding for ranking approximation, but they suffer from severe degradation in tail positions. Although speculative decoding (SD) methods can be a remedy with verification at different positions, they face challenges in ranking systems due to their left-to-right decoding paradigm. Firstly, ranking systems require strict latency constraints, but verification rounds in SD methods remain agnostic; Secondly, SD methods usually discard listwise ranking knowledge about unaccepted items in previous rounds, hindering future multi-token prediction, especially when candidate tokens are the unaccepted items. In this paper, we propose a Reinforcement Speculative Decoding method for fast ranking inference of LLMs. To meet the ranking systems' latency requirement, we propose an up-to-down decoding paradigm that employs an agent to iteratively modify the ranking sequence under a constrained budget. Specifically, we design a ranking-tailored policy optimization, actively exploring optimal multi-round ranking modification policy verified by LLMs via reinforcement learning (RL). To better approximate the target LLM under the constrained budget, we trigger the agent fully utilizing the listwise ranking knowledge about all items verified by LLMs across different rounds in RL, enhancing the modification policy of the agent. More importantly, we demonstrate the theoretical robustness and advantages of our paradigm and implementation. Experiments on both IR and RS tasks show the effectiveness of our proposed method.

Large language models (LLMs) are increasingly applied to ranking tasks in retrieval and recommendation. Although reasoning prompting can enhance ranking utility, our preliminary exploration reveals that its benefits are inconsistent and come at a substantial computational cost, suggesting that when to reason is as crucial as how to reason. To address this issue, we propose a reasoning routing framework that employs a lightweight, plug-and-play router head to decide whether to use direct inference (Non-Think) or reasoning (Think) for each instance before generation. The router head relies solely on pre-generation signals: i) compact ranking-aware features (e.g., candidate dispersion) and ii) model-aware difficulty signals derived from a diagnostic checklist reflecting the model's estimated need for reasoning. By leveraging these features before generation, the router outputs a controllable token that determines whether to apply the Think mode. Furthermore, the router can adaptively select its operating policy along the validation Pareto frontier at deployment time, enabling dynamic allocation of computational resources toward instances most likely to benefit from Think under varying system constraints. Experiments on three public ranking datasets with different scales of open-source LLMs show consistent improvements in ranking utility with reduced token consumption (e.g., +6.3% NDCG@10 with –49.5% tokens on MovieLens with Qwen3-4B), demonstrating reasoning routing as a practical solution to the accuracy-efficiency trade-off.

Large Language Models (LLMs) have been widely adopted in ranking systems, specifically for reranking tasks. Despite the effectiveness, the auto-regressive decoding of LLMs leads to inference latency due to the memory-bandwidth-bound. To alleviate this bottleneck, prior studies have explored single token decoding as an approximation, but they suffer from performance degradation at the tail positions of ranking. In this paper, we propose a Single Forward Pass (SFP)-based method to pre-verify multiple rankings using tree attention, approximating auto-regressive decoding by relevant sub-rankings at each step. However, verifying all possible ranking permutations will lead to factorial-level token computation (N!), making it intractable within SFP. To this end, we first reduce item ranking permutations to combinations (2N), based on the empirical observation that the LLM's next-item generation is less sensitive to the exact ordering of preceding items. Furthermore, we divide the full ranking into K sub-rankings and aggregate their individual probabilities, which further decreases the verification space to 2⌈N/K⌉ • K << 2N. However, naïvely aggregating sub-ranking probabilities leads to inaccurate estimates in listwise ranking. To overcome this, we introduce a Möbius inversion model that explicitly decompose the individual contribution of subsets within a complete lattice, as verified by tree attention. Then, we learn their higher-order effects with a hierarchical self-attention model to reconstruct the full ranking probability. Experiments on both information retrieval and recommendation tasks show the effectiveness of our proposed method.

New-items play a crucial role in recommender systems (RSs) for delivering fresh and engaging user experiences. However, traditional methods struggle to effectively recommend new-items due to their short exposure time and limited interaction records, especially in dynamic recommender systems (DRSs) where new-items get continuously introduced and users' preferences evolve over time. This leads to significant unfairness towards new-items, which could accumulate over the successive model updates, ultimately compromising the stability of the entire system. Therefore, we propose FairAgent, a reinforcement learning (RL)-based new-item fairness enhancement framework specifically designed for DRSs. It leverages knowledge distillation to extract collaborative signals from traditional models, retaining strong recommendation capabilities for old-items. In addition, FairAgent introduces a novel reward mechanism for recommendation tailored to the characteristics of DRSs, which consists of three components: 1) a new-item exploration reward to promote the exposure of dynamically introduced new-items, 2) a fairness reward to adapt to users' personalized fairness requirements for new-items, and 3) an accuracy reward which leverages users' dynamic feedback to enhance recommendation accuracy. Extensive experiments on three public datasets and backbone models demonstrate the superior performance of FairAgent. The results present that FairAgent can effectively boost new-item exposure, achieve personalized new-item fairness, while maintaining high recommendation accuracy.

Collaborative filtering (CF) has been extensively studied in recommendation, spawning various solutions. While graph convolution networks (GCNs) are effective at representation learning, their efficiency is lacking. Comparatively, item similarity model efficiently establishes direct relationships between items. In spite of this, the modeling problem grows quadratically as the number of items increases. This poses critical scalability issues. In this paper, through an investigation of the latest GCN model, we reveal the feasibility of optimizing the process of similarity modeling using the underlying group structure in the item set. Based on these findings, we propose a novel model which introduces graph partitioning to reduce the scale of similarity modeling problem, dubbed FPSR+. Specifically, we divide similarity modeling of items into sub-problems within each partition, and incorporate global and local prior knowledge to alleviate information loss. Following an analysis of the properties of different items in partitioning, we propose a new hub set selection strategy that improves the robustness of FPSR+ in the small partition case. Extensive experiments on four real-world datasets demonstrate the superior performance of FPSR+ compared with state-of-the-art GCN models and item similarity models, as well as several-fold speedups and reductions in parameter storage.

In collaborative filtering, distance metric learning has been applied to matrix factorization techniques with promising results. However, matrix factorization lacks the ability of capturing collaborative information, which has been remarked by recent works and improved by interpreting user interactions as signals. This paper aims to find out how metric learning connect to these signal-based models. By adopting a generalized distance metric, we discovered that in signal-based models, it is easier to estimate the residual of distances, which refers to the difference between the distances from a user to a target item and another item, rather than estimating the distances themselves. Further analysis also uncovers a link between the normalization strength of interaction signals and the novelty of recommendation, which has been overlooked by existing studies. Based on the above findings, we propose a novel model to learn a generalized distance user-item distance metric to capture user preference in interaction signals by modeling the residuals of distance. The proposed CoRML model is then further improved in training efficiency by a newly introduced approximated ranking weight. Extensive experiments conducted on 4 public datasets demonstrate the superior performance of CoRML compared to the state-of-the-art baselines in collaborative filtering, along with high efficiency and the ability of providing novelty-promoted recommendations, shedding new light on the study of metric learning-based recommender systems.

Collaborative filtering (CF) is widely searched in recommendation with various types of solutions. Recent success of Graph Convolution Networks (GCN) in CF demonstrates the effectiveness of modeling high-order relationships through graphs, while repetitive graph convolution and iterative batch optimization limit their efficiency. Instead, item similarity models attempt to construct direct relationships through efficient interaction encoding. Despite their great performance, the growing item numbers result in quadratic growth in similarity modeling process, posing critical scalability problems. In this paper, we investigate the graph sampling strategy adopted in latest GCN model for efficiency improving, and identify the potential item group structure in the sampled graph. Based on this, we propose a novel item similarity model which introduces graph partitioning to restrict the item similarity modeling within each partition. Specifically, we show that the spectral information of the original graph is well in preserving global-level information. Then, it is added to fine-tune local item similarities with a new data augmentation strategy acted as partition-aware prior knowledge, jointly to cope with the information loss brought by partitioning. Experiments carried out on 4 datasets show that the proposed model outperforms state-of-the-art GCN models with 10x speed-up and item similarity models with 95% parameter storage savings.