Searcharxiv⌕ Search

arXiv subjects

Zhenrui Yue

Publications and source records attributed to Zhenrui Yue.

At least 19 recordsLinked to original sources
arXiv2026

Uncertainty-Aware Variational Reward Factorization via Probabilistic Preference Bases for LLM Personalization

Reward factorization personalizes large language models (LLMs) by decomposing rewards into shared basis functions and user-specific weights. Yet, existing methods estimate user weights from scarce data in isolation and as deterministic points, leading to inaccurate and unreliable inference. We introduce Variational Reward Factorization (VRF), an uncertainty-aware framework that represents each user's preferences as a variational distribution in a shared preference space. VRF infers user distributions via a variational encoder, derives weights through Wasserstein distance matching with shared probabilistic bases, and downweights uncertain estimates through a variance-attenuated loss. On three benchmarks, VRF outperforms all baselines across seen and unseen users, few-shot scenarios, and varying uncertainty levels, with gains extending to downstream alignment. Our code is available at https://github.com/Gyu-Seok-Lee/VRF_COLM26.

cs.CL↗
arXiv2026

Dr. Zero: Self-Evolving Search Agents without Training Data

As high-quality data becomes increasingly difficult to obtain, self-evolution without curated training data has emerged as a promising paradigm. This approach allows large language models (LLMs) to autonomously generate and solve complex problems, thereby improving their reasoning capabilities. However, multi-turn search agents struggle in this setting due to limited question diversity and the substantial compute required for multi-step reasoning and tool use. In this work, we introduce Dr. Zero, a framework that enables search agents to effectively self-evolve without human-annotated training data, relying solely on an external search engine as their knowledge environment. In particular, we design a self-evolution feedback loop where a proposer generates structurally diverse questions to train a solver initialized from the same base model. As the solver evolves, it incentivizes the proposer to produce increasingly difficult yet solvable tasks, thus establishing an automated curriculum to refine both agents. To enhance training efficiency, we also introduce hop-grouped relative policy optimization (HRPO). This method clusters structurally similar questions to construct group-level baselines, effectively minimizing the sampling overhead in evaluating each query's individual difficulty and solvability. Consequently, HRPO significantly reduces the compute requirements for proposer training and reward estimation without compromising performance or stability. Extensive experimental results demonstrate that Dr. Zero matches or surpasses fully supervised search agents on several question answering benchmarks, showing that strong agentic search and evidence-grounded reasoning can emerge solely through self-evolution.

cs.AI↗
arXiv2026

Topology-Aware Tokenization for Generative Recommendation

Generative recommendation reformulates sequential recommendation as an autoregressive generation task, yet a critical issue in this paradigm remains overlooked: topology distortion in item tokenization. In particular, we observe that the intrinsic adjacency relationships of items in the pretrained semantic embedding space are significantly disrupted after quantization. This topology distortion misleads the model's perception of item similarity, ultimately bottlenecking the accuracy of generative recommendations. To address this issue, we propose Topology-Aware Tokenization (TopoTok), an item tokenization framework that preserves item relational structure throughout the quantization hierarchy. Different from the prior monolithic supervision in tokenization, TopoTok introduces a multi-level distillation scheme to progressively recover the topology from coarse to fine granularity: 1) Inter-Group Distillation to capture global cluster-wise relations; 2) Intra-Group Distillation to refine local structures within semantic clusters; and 3) Inter-Item Distillation to enforce fine-grained alignment at the individual item level. Extensive experiments on three benchmark datasets demonstrate that TopoTok effectively alleviates topology distortion and consistently outperforms state-of-the-art tokenizers, achieving significant performance gains of up to 9.42% in Recall@5.

cs.IR↗
arXiv2026

The Cartesian Shortcut: Re-evaluate Vision Reasoning in Polar Coordinate Space

As current Multimodal Large Language Models rapidly saturate canonical visual reasoning benchmarks, a key question emerges: do these strong scores genuinely reflect robust visual understanding? We identify a pervasive vulnerability, the Cartesian Shortcut: visual reasoning benchmarks prevalently build on orthogonal grid-based layouts that can be readily discretized into explicit textual coordinates. Models systematically exploit this property, heavily leveraging text-based deductive reasoning to assist visual problem-solving. To systematically dismantle this shortcut, we introduce Polaris-Bench, which re-formulates 53 visual reasoning tasks in Polar coordinate space with paired Cartesian counterparts as reference, while preserving consistent logical constraints and task semantics -- thus fundamentally breaking the orthogonal prior that models exploit. Comprehensive evaluation across $14$ state-of-the-art MLLMs reveals that frontier models achieving $70$--$83\%$ on Cartesian layouts collapse to $31$--$39\%$ on Polar equivalents, with degradation persisting even under complete logical equivalence. Moreover, reasoning gains observed on Cartesian layouts are severely diminished on Polar equivalents. These findings expose a critical deficiency in current MLLMs: the lack of topology-invariant visual reasoning.

cs.CV↗
arXiv2026

Learning Decomposed Contextual Token Representations from Pretrained and Collaborative Signals for Generative Recommendation

Recent advances in generative recommenders adopt a two-stage paradigm: items are first tokenized into semantic IDs using a pretrained tokenizer, and then large language models (LLMs) are trained to generate the next item via sequence-to-sequence modeling. However, these two stages are optimized for different objectives: semantic reconstruction during tokenizer pretraining versus user interaction modeling during recommender training. This objective misalignment leads to two key limitations: (i) suboptimal static tokenization, where fixed token assignments fail to reflect diverse usage contexts; and (ii) discarded pretrained semantics, where pretrained knowledge - typically from language model embeddings - is overwritten during recommender training on user interactions. To address these limitations, we propose to learn $\underline{DE}$composed $\underline{CO}$ntextual Token $\underline{R}$epresentations (DECOR), a unified framework that preserves pretrained semantics while enhancing the adaptability of token embeddings. DECOR introduces contextualized token composition to refine token embeddings based on user interaction context, and decomposed embedding fusion that integrates pretrained codebook embeddings with newly learned collaborative embeddings. Experiments on three real-world datasets demonstrate that DECOR consistently outperforms state-of-the-art baselines in recommendation performance.

cs.IR↗
arXiv2026

SPRINT: Scalable and Predictive Intent Refinement for LLM-Enhanced Session-based Recommendation

Large language models (LLMs) have enhanced conventional recommendation models via user profiling, which generates representative textual profiles from users' historical interactions. However, their direct application to session-based recommendation (SBR) remains challenging due to severe session context scarcity and poor scalability. In this paper, we propose SPRINT, a scalable SBR framework that incorporates reliable and informative intents while ensuring high efficiency in both training and inference. SPRINT constrains LLM-based profiling with a global intent pool and validates inferred intents based on recommendation performance to mitigate noise and hallucinations under limited context. To ensure scalability, LLMs are selectively invoked only for uncertain sessions during training, while a lightweight intent predictor generalizes intent prediction to all sessions without LLM dependency at inference time. Experiments on real-world datasets show that SPRINT consistently outperforms state-of-the-art methods while providing more explainable recommendations.

cs.IR↗
arXiv2026

Retrieval Augmented Conversational Recommendation with Reinforcement Learning

Large language models (LLMs) exhibit enhanced capabilities in language understanding and generation. By utilizing their embedded knowledge, LLMs are increasingly used as conversational recommender systems (CRS), achieving improved performance across diverse scenarios. However, existing LLM-based methods rely on pretrained knowledge without external retrieval mechanisms for novel items. Additionally, the lack of a unified corpus poses challenges for integrating retrieval augmentation into CRS. Motivated by these challenges, we present RAR, a novel two-stage retrieval augmented conversational recommendation framework that aligns retrieval and generation to enhance both performance and factuality. To support this framework and provide a unified corpus, we construct a large-scale movie corpus, comprising over 300k movies with rich metadata, such as titles, casts and plot summaries. Leveraging this data, our primary contribution is RAR, the first framework to departs from standard two-stage CRS by dynamically bridging retrieval and generation. First, a retriever model generates candidate items based on user history; in the subsequent stage, an LLM refines the recommendations by incorporating conversational context with retrieved results. In addition, we introduce a novel reinforcement learning (RL) method that leverages LLM feedback to iteratively update the retriever. By creating a collaborative feedback loop that reinforces sampled candidate sets with higher ranking metrics, RAR effectively mitigates the misalignment between the retrieval and generation stages. Furthermore, grounding the LLM in factual metadata allows our RL-driven approach to capture subtle user intentions and generate context-aware recommendations with reduced hallucinations. We validate our approach through extensive experiments on multiple benchmarks, where RAR consistently outperforms state-of-the-art baseline methods.

cs.IR↗
arXiv2026

FASA: Frequency-aware Sparse Attention

The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of "dominant" FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100\% of full-KV performance when only keeping 256 tokens, and achieves 2.56$\times$ speedup using just 18.9\% of the cache on AIME24.

cs.CL↗
arXiv2025

Hybrid Latent Reasoning via Reinforcement Learning

Recent advances in large language models (LLMs) have introduced latent reasoning as a promising alternative to autoregressive reasoning. By performing internal computation with hidden states from previous steps, latent reasoning benefit from more informative features rather than sampling a discrete chain-of-thought (CoT) path. Yet latent reasoning approaches are often incompatible with LLMs, as their continuous paradigm conflicts with the discrete nature of autoregressive generation. Moreover, these methods rely on CoT traces for training and thus fail to exploit the inherent reasoning patterns of LLMs. In this work, we explore latent reasoning by leveraging the intrinsic capabilities of LLMs via reinforcement learning (RL). To this end, we introduce hybrid reasoning policy optimization (HRPO), an RL-based hybrid latent reasoning approach that (1) integrates prior hidden states into sampled tokens with a learnable gating mechanism, and (2) initializes training with predominantly token embeddings while progressively incorporating more hidden features. This design maintains LLMs' generative capabilities and incentivizes hybrid reasoning using both discrete and continuous representations. In addition, the hybrid HRPO introduces stochasticity into latent reasoning via token sampling, thereby enabling RL-based optimization without requiring CoT trajectories. Extensive evaluations across diverse benchmarks show that HRPO outperforms prior methods in both knowledge- and reasoning-intensive tasks. Furthermore, HRPO-trained LLMs remain interpretable and exhibit intriguing behaviors like cross-lingual patterns and shorter completion lengths, highlighting the potential of our RL-based approach and offer insights for future work in latent reasoning.

cs.CL↗
arXiv2025

Can Pre-training Indicators Reliably Predict Fine-tuning Outcomes of LLMs?

While metrics available during pre-training, such as perplexity, correlate well with model performance at scaling-laws studies, their predictive capacities at a fixed model size remain unclear, hindering effective model selection and development. To address this gap, we formulate the task of selecting pre-training checkpoints to maximize downstream fine-tuning performance as a pairwise classification problem: predicting which of two LLMs, differing in their pre-training, will perform better after supervised fine-tuning (SFT). We construct a dataset using 50 1B parameter LLM variants with systematically varied pre-training configurations, e.g., objectives or data, and evaluate them on diverse downstream tasks after SFT. We first conduct a study and demonstrate that the conventional perplexity is a misleading indicator. As such, we introduce novel unsupervised and supervised proxy metrics derived from pre-training that successfully reduce the relative performance prediction error rate by over 50%. Despite the inherent complexity of this task, we demonstrate the practical utility of our proposed proxies in specific scenarios, paving the way for more efficient design of pre-training schemes optimized for various downstream tasks.

cs.CL↗
arXiv2025

Boosting Data Utilization for Multilingual Dense Retrieval

Multilingual dense retrieval aims to retrieve relevant documents across different languages based on a unified retriever model. The challenge lies in aligning representations of different languages in a shared vector space. The common practice is to fine-tune the dense retriever via contrastive learning, whose effectiveness highly relies on the quality of the negative sample and the efficacy of mini-batch data. Different from the existing studies that focus on developing sophisticated model architecture, we propose a method to boost data utilization for multilingual dense retrieval by obtaining high-quality hard negative samples and effective mini-batch data. The extensive experimental results on a multilingual retrieval benchmark, MIRACL, with 16 languages demonstrate the effectiveness of our method by outperforming several existing strong baselines.

cs.IR↗
arXiv2025

Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning

Efficiently acquiring external knowledge and up-to-date information is essential for effective reasoning and text generation in large language models (LLMs). Prompting advanced LLMs with reasoning capabilities to use search engines during inference is often suboptimal, as the LLM might not fully possess the capability on how to interact optimally with the search engine. This paper introduces Search-R1, an extension of reinforcement learning (RL) for reasoning frameworks where the LLM learns to autonomously generate (multiple) search queries during step-by-step reasoning with real-time retrieval. Search-R1 optimizes LLM reasoning trajectories with multi-turn search interactions, leveraging retrieved token masking for stable RL training and a simple outcome-based reward function. Experiments on seven question-answering datasets show that Search-R1 improves performance by 41% (Qwen2.5-7B) and 20% (Qwen2.5-3B) over various RAG baselines under the same setting. This paper further provides empirical insights into RL optimization methods, LLM choices, and response length dynamics in retrieval-augmented reasoning. The code and model checkpoints are available at https://github.com/PeterGriffinJin/Search-R1.

cs.CL↗
arXiv2025

Inference Scaling for Long-Context Retrieval Augmented Generation

The scaling of inference computation has unlocked the potential of long-context large language models (LLMs) across diverse settings. For knowledge-intensive tasks, the increased compute is often allocated to incorporate more external knowledge. However, without effectively utilizing such knowledge, solely expanding context does not always enhance performance. In this work, we investigate inference scaling for retrieval augmented generation (RAG), exploring the combination of multiple strategies beyond simply increasing the quantity of knowledge, including in-context learning and iterative prompting. These strategies provide additional flexibility to scale test-time computation (e.g., by increasing retrieved documents or generation steps), thereby enhancing LLMs' ability to effectively acquire and utilize contextual information. We address two key questions: (1) How does RAG performance benefit from the scaling of inference computation when optimally configured? (2) Can we predict the optimal test-time compute allocation for a given budget by modeling the relationship between RAG performance and inference parameters? Our observations reveal that increasing inference computation leads to nearly linear gains in RAG performance when optimally allocated, a relationship we describe as the inference scaling laws for RAG. Building on this, we further develop the computation allocation model to estimate RAG performance across different inference configurations. The model predicts optimal inference parameters under various computation constraints, which align closely with the experimental results. By applying these optimal configurations, we demonstrate that scaling inference compute on long-context LLMs achieves up to 58.9% gains on benchmark datasets compared to standard RAG.

cs.CL↗
arXiv2025

Your Causal Self-Attentive Recommender Hosts a Lonely Neighborhood

In the context of sequential recommendation, a pivotal issue pertains to the comparative analysis between bi-directional/auto-encoding (AE) and uni-directional/auto-regressive (AR) attention mechanisms, where the conclusions regarding architectural and performance superiority remain inconclusive. Previous efforts in such comparisons primarily involve summarizing existing works to identify a consensus or conducting ablation studies on peripheral modeling techniques, such as choices of loss functions. However, far fewer efforts have been made in (1) theoretical and (2) extensive empirical analysis of the self-attention module, the very pivotal structure on which performance and designing insights should be anchored. In this work, we first provide a comprehensive theoretical analysis of AE/AR attention matrix in the aspect of (1) sparse local inductive bias, a.k.a neighborhood effects, and (2) low rank approximation. Analytical metrics reveal that the AR attention exhibits sparse neighborhood effects suitable for generally sparse recommendation scenarios. Secondly, to support our theoretical analysis, we conduct extensive empirical experiments on comparing AE/AR attention on five popular benchmarks with AR performing better overall. Empirical results reported are based on our experimental pipeline named Modularized Design Space for Self-Attentive Recommender (ModSAR), supporting adaptive hyperparameter tuning, modularized design space and HuggingFace plug-ins. We invite the recommendation community to utilize/contribute to ModSAR to (1) conduct more module/model-level examining beyond AE/AR comparison and (2) accelerate state-of-the-art model design. Lastly, we shed light on future design choices for performant self-attentive recommenders. We make our pipeline implementation and data available at https://github.com/yueqirex/SAR-Check.

cs.IR↗
arXiv2024

Transferable Sequential Recommendation via Vector Quantized Meta Learning

While sequential recommendation achieves significant progress on capturing user-item transition patterns, transferring such large-scale recommender systems remains challenging due to the disjoint user and item groups across domains. In this paper, we propose a vector quantized meta learning for transferable sequential recommenders (MetaRec). Without requiring additional modalities or shared information across domains, our approach leverages user-item interactions from multiple source domains to improve the target domain performance. To solve the input heterogeneity issue, we adopt vector quantization that maps item embeddings from heterogeneous input spaces to a shared feature space. Moreover, our meta transfer paradigm exploits limited target data to guide the transfer of source domain knowledge to the target domain (i.e., learn to transfer). In addition, MetaRec adaptively transfers from multiple source tasks by rescaling meta gradients based on the source-target domain similarity, enabling selective learning to improve recommendation performance. To validate the effectiveness of our approach, we perform extensive experiments on benchmark datasets, where MetaRec consistently outperforms baseline methods by a considerable margin.

cs.IR↗
arXiv2024

Train Once, Deploy Anywhere: Matryoshka Representation Learning for Multimodal Recommendation

Despite recent advancements in language and vision modeling, integrating rich multimodal knowledge into recommender systems continues to pose significant challenges. This is primarily due to the need for efficient recommendation, which requires adaptive and interactive responses. In this study, we focus on sequential recommendation and introduce a lightweight framework called full-scale Matryoshka representation learning for multimodal recommendation (fMRLRec). Our fMRLRec captures item features at different granularities, learning informative representations for efficient recommendation across multiple dimensions. To integrate item features from diverse modalities, fMRLRec employs a simple mapping to project multimodal item features into an aligned feature space. Additionally, we design an efficient linear transformation that embeds smaller features into larger ones, substantially reducing memory requirements for large-scale training on recommendation data. Combined with improved state space modeling techniques, fMRLRec scales to different dimensions and only requires one-time training to produce multiple models tailored to various granularities. We demonstrate the effectiveness and efficiency of fMRLRec on multiple benchmark datasets, which consistently achieves superior performance over state-of-the-art baseline methods. We make our code and data publicly available at https://github.com/yueqirex/fMRLRec.

cs.IR↗
arXiv2024

Retrieval Augmented Fact Verification by Synthesizing Contrastive Arguments

The rapid propagation of misinformation poses substantial risks to public interest. To combat misinformation, large language models (LLMs) are adapted to automatically verify claim credibility. Nevertheless, existing methods heavily rely on the embedded knowledge within LLMs and / or black-box APIs for evidence collection, leading to subpar performance with smaller LLMs or upon unreliable context. In this paper, we propose retrieval augmented fact verification through the synthesis of contrasting arguments (RAFTS). Upon input claims, RAFTS starts with evidence retrieval, where we design a retrieval pipeline to collect and re-rank relevant documents from verifiable sources. Then, RAFTS forms contrastive arguments (i.e., supporting or refuting) conditioned on the retrieved evidence. In addition, RAFTS leverages an embedding model to identify informative demonstrations, followed by in-context prompting to generate the prediction and explanation. Our method effectively retrieves relevant documents as evidence and evaluates arguments from varying perspectives, incorporating nuanced information for fine-grained decision-making. Combined with informative in-context examples as prior, RAFTS achieves significant improvements to supervised and LLM baselines without complex prompts. We demonstrate the effectiveness of our method through extensive experiments, where RAFTS can outperform GPT-based methods with a significantly smaller 7B LLM.

cs.CL↗
arXiv2024

Open-Vocabulary Federated Learning with Multimodal Prototyping

Existing federated learning (FL) studies usually assume the training label space and test label space are identical. However, in real-world applications, this assumption is too ideal to be true. A new user could come up with queries that involve data from unseen classes, and such open-vocabulary queries would directly defect such FL systems. Therefore, in this work, we explicitly focus on the under-explored open-vocabulary challenge in FL. That is, for a new user, the global server shall understand her/his query that involves arbitrary unknown classes. To address this problem, we leverage the pre-trained vision-language models (VLMs). In particular, we present a novel adaptation framework tailored for VLMs in the context of FL, named as Federated Multimodal Prototyping (Fed-MP). Fed-MP adaptively aggregates the local model weights based on light-weight client residuals, and makes predictions based on a novel multimodal prototyping mechanism. Fed-MP exploits the knowledge learned from the seen classes, and robustifies the adapted VLM to unseen categories. Our empirical evaluation on various datasets validates the effectiveness of Fed-MP.

cs.CL↗