Searcharxiv⌕ Search

arXiv subjects

Zongwei Wang

Publications and source records attributed to Zongwei Wang.

At least 19 recordsLinked to original sources
arXiv2026

Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning

Test-time adaptation has emerged as a lightweight alternative to costly post-training for improving the reasoning capabilities of Large Language Models (LLMs) on downstream tasks. Predictive entropy provides a model-derived signal for such adaptation, guiding models toward higher-confidence reasoning states without external verifiers or reward models. However, higher confidence does not necessarily imply correctness, as LLMs may remain highly confident along incorrect reasoning trajectories. We observe that high-confidence reasoning is more likely to be correct when confidence remains stable under local perturbations. Based on this observation, we propose Test-Time Adaptation via Stability-Aware Confidence Optimization (TASCO), a framework that incorporates local stability into confidence-based test-time adaptation while keeping the LLM frozen. TASCO operationalizes local stability by optimizing a lightweight task-level prefix under two alternative perturbation strategies: Random Perturbation promotes distributional stability across trajectories induced by nearby perturbed prefixes, whereas Sharpness-Aware Perturbation targets worst-case local sensitivity. Experiments demonstrate that TASCO improves reasoning accuracy and token efficiency across diverse LLMs and reasoning benchmarks, while behavioral analyses show that it maintains stable confidence under local perturbations without prematurely concentrating the model's predictive distribution.

cs.AI↗
arXiv2026

SwiftExplorer: Training-free Diffusion Model Alignment with Swift Diversity Exploration

Diffusion models have general generative abilities but struggle to align with specific objectives. Fine-tuning can improve alignment, yet its training cost is often prohibitive. This led to training-free methods that apply objective-guided terms in sampling to bias the generation distribution toward designated regions, e.g., high-reward areas. However, these methods face two issues: (1) the strong directional bias narrows the pretrained distribution and generation diversity, and (2) indiscriminate constant guidance fails to prune redundant signals, hurting both quality and efficiency. To address the above challenges, we propose SwiftExplorer, a plugin that mitigates distribution collapse caused by excessive diversity loss and reduces compute costs. First, we adopt an Inheritance-Restart exploration mechanism to avoid early convergence, while exploration also increases the likelihood of high-reward trajectories. Additionally, it balances diversity and fidelity, adding diversity without causing a distribution over-shift. Second, our Quality-Efficiency arbitration mechanism improves guidance by removing incorrect signals, and it reduces computation by dynamically stopping generation when completeness and marginal reward gain are optimal. In an extensive number of experiments and different types of evaluation metrics, the proposed SwiftExplorer achieves excellent performance on all metrics, including preference, fidelity, diversity, and richness.

cs.LG↗
arXiv2026

Personalized Communication Skills for Agentic Recommender Systems

Agentic recommender systems increasingly employ large language model-based UserAgents to evaluate candidate items through simulated feedback before recommendations are delivered. However, existing UserAgents typically reason in isolation based on limited personal histories, which may lead to perspective narrowing: the agent evaluates candidates from a local and incomplete view, overlooks relevant preference facets, and consequently produces inaccurate judgments. A natural way to alleviate this problem is to introduce other users as advisor agents, whose diverse histories provide complementary evidence that helps the target user reconsider overlooked preference signals. Nevertheless, a generic user-advisor communication process is insufficient, as different user decision states require different forms of external advice. Based on this insight, we propose AgentCom, a personalized communication skill framework for agentic recommender systems. AgentCom organizes reusable communication skills into a shared why--what--how--who skill bank: why identifies the decision deficiency that necessitates communication, what specifies the information task, how determines the advisor interaction protocol, and who retrieves advisors capable of executing that protocol. To make the shared skill bank personalized at use time and adaptive over time, AgentCom introduces two complementary mechanisms: personalized skill routing and failure-driven skill evolution. Personalized skill routing constructs a communication path by sequentially selecting suitable skills for each user and recommendation context. Failure-driven skill evolution learns from unsuccessful communication cases and enriches the shared bank with reusable skills that address previously uncovered communication needs. Experiments show that AgentCom consistently improves recommendation performance across traditional, social, and agentic recommenders.

cs.IR↗
arXiv2026

Explore or Converge? Stage-Guided Per-Step Optimization for Diffusion Models

Diffusion models have strong generative capabilities. However, their maximum likelihood training objective only focuses on reconstructing the data distribution, making it difficult to align with specific preferences. Reinforcement learning (RL) for preference alignment in diffusion models is promising but limited by reward sparsity. Since a single reward cannot support optimization, existing RL methods usually backpropagate the final reward to all previous steps. However, denoising is stage-wise, with distinct semantics and controllability. Repeating the final reward across all steps creates a temporal objective mismatch, encouraging reward shortcuts that lead to reward hacking. At the same time, due to reward backfilling, each time step receives the same reward, making it impossible to distinguish between actions, thereby weakening the optimization process. To resolve this issue, we propose Stage-Guided Per-Step Optimization (SGPO) for diffusion models, which jointly leverages signal-to-noise ratio and semantic changes to identify generation stages and adaptively assign stage-specific objectives. Early denoising is chaotic and far from the final reward, resulting in weak reward-behavior correlation. This stage should prioritize exiting the chaotic state. In the mid stage, the latent transitions to a stable structure, where the final reward better corresponds to generative behavior. Therefore, this stage optimizes the final reward while exploring diversity to avoid early convergence to a single mode. In the late stage, the latent's core structure is largely fixed, and preference optimization mainly amplifies local details, risking overfitting. Therefore, stable convergence is preferred to avoid quality degradation. Results from 16 comparative experiments validate SGPO. Our method achieves 26.7% average gains in generative quality and 36.7% higher convergence speed.

cs.CV↗
arXiv2026

PAST: Prompt-Adaptive Sampling Termination for Efficient Diffusion Model

While diffusion models have made significant progress in text-to-image tasks, they still exhibit limitations when directly optimizing downstream objectives. Although Reinforcement Learning (RL) enables targeted optimization, existing methods are generally constrained by low-efficiency fine-tuning and sparse rewards. To address these challenges, we propose PAST, which provides differentiated rewards while adaptively regulating training episode length by jointly perceiving denoising progress and prompt difficulty. Specifically, we design an intrinsic reward paradigm to compensate for sparse extrinsic rewards and guide the model to explore paths that diverge more efficiently from noise patterns. We further provide theoretical justification for intrinsic rewards. Then, PAST dynamically monitors denoising completion and semantic alignment between image structures and prompt semantics. When both metrics satisfy generation requirements, the system adaptively terminates training. This enables appropriate allocation of episode lengths based on prompt difficulty and the current generation process. Finally, based on the predicted residual noise level, we establish a dual adaptive coordination mechanism. Specifically, it not only balances the extrinsic and intrinsic rewards but also balances the exploration and convergence. Experimental results demonstrate that PAST enhances computational efficiency of existing RL fine-tuning methods by up to 66.7%, while improving preference optimization quality by up to 29.5% through its dual adaptive regulation mechanism.

cs.CV↗
arXiv2026

HEMERA: A Heterogeneous Memory-Centric Accelerator with Recursive Dataflow for Edge-Constrained State-Space-Duality Models Inference

Structured State Space Models (SSMs), such as Mamba, enable efficient long-sequence modeling with linear time complexity. Recent implementations realize this capability through Structured State Space Duality (SSD), which transforms recursive state evolution into matrix-form computations. However, SSD introduces substantial system-level overheads, including quadratic intermediate materialization, irregular data movement, and prefix-dependent execution, leading to excessive memory traffic and bandwidth demand on conventional architectures. Although prior accelerators mitigate these overheads through optimized dataflows or compute-in-memory techniques, they largely retain matrix-oriented SSD execution and cannot simultaneously avoid quadratic intermediate storage and efficiently map dependency-bound state propagation. This paper presents HEMERA, a heterogeneous memory-centric accelerator for efficient Mamba-2 inference. Rather than directly executing the matrix-form SSD computation, HEMERA reformulates it into an algebraically equivalent streaming-recursive dataflow that avoids quadratic intermediate storage while preserving the original computation. The resulting heterogeneous execution paradigm maps dense linear operations onto in-memory computing units and recursive state updates onto a dedicated streaming engine. Across Mamba-2 models ranging from 130M to 2.8B, HEMERA achieves average latency speedups of 1.4x-3.6x and energy-efficiency improvements of 12.2x-27.0x over the official optimized fused Mamba-2 kernel on NVIDIA A100. It further reduces the average SSD-related execution-time ratio across model scales to 14.12% during long-sequence inference, demonstrating its potential for efficient deployment under edge constraints.

cs.AR↗
arXiv2026

Where Reasoning Matters: Rethinking Latent Reasoning in Semantic ID-based Generative Recommendation

Semantic ID-based generative recommendation predicts an item by generating a short sequence of semantic ID tokens, where each token is produced autoregressively. Latent reasoning has recently been introduced to improve this process through additional hidden-state computation before each token decision. This raises a practical question: when one item is represented by a sequence of semantic ID tokens, should each token receive the same fixed number of latent refinement steps, or should these steps be allocated more effectively across positions? We study this question through position-wise information-gain (IG), which measures how much each semantic ID position reduces the uncertainty of the target item. We observe that earlier semantic ID positions usually provide higher information-gain, while later positions contribute less additional information. We further analyze that applying more refinement to high-IG positions tends to bring larger expected benefits. Based on this observation, we propose IBA, an Information-Gain Budget Allocation framework for semantic ID-based generative recommendation. IBA treats latent refinement steps as a limited computational resource and learns how to allocate them across semantic ID positions, assigning more refinement to informative positions and less to positions with smaller contribution. Experiments on multiple public datasets show that IBA consistently improves strong generative recommendation baselines and achieves a better accuracy--computation trade-off than fixed or poorly matched step allocations.

cs.IR↗
arXiv2026

DynaGraph: Lightweight Multi-Model Interaction Framework via Dynamic Topological Reconfiguration

Tackling complex reasoning tasks typically relies on massive monolithic LLMs, which suffer from severe computational redundancy. While task decomposition through structured pipelines or multi-agent collaborations offers an alternative, these approaches inevitably fall into a critical dilemma: predefined static topologies are highly vulnerable to cascading errors, whereas unconstrained dynamic agents suffer from trajectory divergence and unpredictable memory bloat. To address this, we present DynaGraph, a lightweight multi-model framework driven by dynamic topological reconfiguration. At the execution level, DynaGraph multiplexes time-division PEFT adapters over a shared base model, enabling both full system training and inference deployment on a single consumer-grade GPU. At the routing level, the Evaluator continuously monitors execution confidence to trigger hierarchical self-healing: Fine-grained Patching for localized data gaps and Subgraph Reconstruction for severe logical ruptures. Experiments on StrategyQA, MATH, and FinQA demonstrate our 8B model closely approximates the reasoning capabilities of a 72B monolithic model (e.g., 87.6% on StrategyQA, 82.7% on MATH). Furthermore, it reduces latency by up to 68.1% and token consumption by 68.6% compared to unconstrained dynamic architectures.

cs.MA↗
arXiv2026

NASiC: 3D NAND-based CAM-Selected Multibit CIM Architecture for Efficient On-Device Mixture-of-Experts LLM Inference

The Mixture-of-Experts (MoE) models have emerged as the state-of-the-art paradigm for scaling up large language models (LLMs) without proportionally increased computational cost. However, its on-device deployment faces a critical challenge due to the large memory requirement for storing all expert parameters. 3D NAND-based computing-in-memory (CIM) architectures uniquely offer high storage capacity and reduced data movement, while they are ill-suited for MoE models with dynamically sparse expert activation, leading to a degradation of effective computational parallelism, along with underutilization of multibit storage capability of Flash cells. In this work, we proposed a 3D NAND-based content addressable-selected CIM architecture, dubbed as NASiC, which is tailored to MoE models. By leveraging the intrinsic string structure of 3D NAND technology, NASiC fuses the dynamical expert selection through CAM-based masking mechanism and activated expert computation through CIM into a single computation cycle, eradicating redundant computation and enhancing computational parallelism. Moreover, circuit-level optimizations and multibit CIM cell are co-designed with proposed NASiC architecture, featuring block-wise parallel computation with in-situ signed multibit input and weight expansion, substantially improving the throughput and energy-efficiency of NAND CIM array, as well as the utilization of high-density 3D NAND technology for MoE models. With extensive experimental results, we demonstrate NASiC achieves 4-114.8x improved performance and 3.9-70x improved energy efficiency over state-of-the-art designs, along with high accuracy, showing its great potential for efficient on-device MoE LLM inference.

cs.AR↗
arXiv2026

Do Less, Achieve More: Do We Need Every-Step Optimization for RL Fine-tuning of Diffusion Models?

Despite strong image-generation performance, diffusion models' reconstruction objectives limit alignment with human preferences. RL enables such alignment through explicit rewards. However, most studies apply RL to the full denoising trajectory, making it computationally costly and weakening preference alignment, i.e., doing more but achieving less. We observe that the impact of RL fine-tuning varies significantly across denoising stages. In the early stage, image structures are unstable and distant from the final reward signal. Applying RL at this stage leads to delayed rewards and action-reward mismatching, resulting in high variance and inefficient updates. Conversely, in the later stage, reward gains saturate, and continued training tends to overfit local details, intensifying reward hacking. To tackle these challenges, we propose AdaScope, an RL-enhanced plug-in that improves generation quality while reducing computational cost. Specifically, AdaScope adaptively identifies the optimal intervention timing for RL by perceiving the structural evolution and semantic consistency during denoising, and dynamically terminates training once the denoising converges and reward gains saturate. As a result, it achieves a rare 'dual benefit': a reduction in computational costs alongside a significant performance improvement. We offer theoretical grounds for the design of AdaScope. Compared with state-of-the-art methods, AdaScope improves performance by 66% while cutting computational cost by 59%.

cs.CV↗
arXiv2026

Generative Bid Shading in Real-Time Bidding Advertising

Bid shading plays a crucial role in Real-Time Bidding (RTB) by adaptively adjusting the bid to avoid advertisers overspending. Existing mainstream two-stage methods, which first model bid landscapes and then optimize surplus using operations research techniques, are constrained by unimodal assumptions that fail to adapt for non-convex surplus curves and are vulnerable to cascading errors in sequential workflows. Additionally, existing discretization models of continuous values ignore the dependence between discrete intervals, reducing the model's error correction ability, while sample selection bias in bidding scenarios presents further challenges for prediction. To address these issues, this paper introduces Generative Bid Shading (GBS), which comprises two primary components: 1) an end-to-end generative model that utilizes an autoregressive approach to generate shading ratios by stepwise residuals, capturing complex value dependencies without relying on predefined priors; and 2) a reward preference alignment system, which incorporates a channel-aware hierarchical dynamic network (CHNet) as the reward model to extract fine-grained features, along with modules for surplus optimization and exploration utility reward alignment, ultimately optimizing both short-term and long-term surplus using group relative policy optimization (GRPO). Extensive experiments on both offline and online A/B tests validate GBS's effectiveness. Moreover, GBS has been deployed on the Meituan DSP platform, serving billions of bid requests daily.

cs.GT↗
arXiv2026

Disagreement as Signals: Dual-view Calibration for Sequential Recommendation Denoising

Sequential recommendation seeks to model the evolution of user interests by capturing temporal user intent and item-level transition patterns. Transformer-based recommenders demonstrate a strong capacity for learning long-range and interpretable dependencies, yet remain vulnerable to behavioral noise that is misaligned with users' true preferences. Recent large language model (LLM)-based approaches attempt to denoise interaction histories through static semantic editing. Such methods neglect the learning dynamics of recommendation models and fail to account for the evolving nature of user interests. To address this limitation, we propose a Dual-view Calibration framework for Sequential Recommendation denoising (DC4SR). Specifically, we introduce a semantic prior, derived from an LLM fine-tuned via labeled historical interactions, to estimate the noise distribution from a semantic perspective. From the learning perspective, we further employ a model-side posterior that infers the noise distribution based on the model's learning dynamics. The disagreement between the two distributions is then leveraged to jointly refine semantic understanding and learning-aware model-side representations. Through iterative updates, dynamic dual-view calibration is achieved for both the global semantic prior and the model-side posterior, enabling consistent alignment with evolving user interests. Extensive experiments demonstrate that DC4SR consistently outperforms strong Transformer-based recommenders and LLM-based denoising methods, exhibiting enhanced robustness across training stages and noise conditions.

cs.IR↗
arXiv2026

Green-Red Watermarking for Recommender Systems

The widespread open-sourcing of advanced recommendation algorithms and the rising threat of model extraction attacks have made safeguarding the intellectual property of recommender systems an imperative task. While watermarking serves as a potent defense, existing methods primarily rely on forcing models to memorize pre-defined interaction patterns. Such memorization-based approaches often require excessive synthetic data injection and are vulnerable to removal attacks due to their detectable statistical deviations from natural user behavior. To address these limitations, we propose GREW, a novel Green-REd Watermarking framework for recommender systems. GREW leverages a secret key to partition the item space into "green" items for soft promotion and "red" items as anchors, thereby shifting the paradigm from fragile memorization to a stealthy, key-controlled output bias. By integrating watermark signals directly into the intrinsic ranking process, GREW employs three recommendation-tailored modules: (1) Semantic-Consistent Hashing, which utilizes the secret key to cluster green items for performance-aware stealthiness; (2) Decision-Aligned Masking, which confines signal injection to the competitive item subset to preserve ranking logic; and (3) Confidence-Aware Scaling, which dynamically modulates injection intensity based on model uncertainty. Ownership verification is performed via statistical hypothesis testing on aggregated black-box outputs, enabled by the keyed re-partitioning of the item space. Experiments on multiple base models demonstrate that GREW achieves strong ownership verification and robustness against extraction attacks compared to existing baselines while requiring no data injection. Our code is available at https://github.com/Loche2/GREW.

cs.IR↗
arXiv2026

Prompt-Unknown Promotion Attacks against LLM-based Sequential Recommender Systems

Large language model-powered sequential recommender systems (LLM-SRSs) have recently demonstrated remarkable performance, enabling recommendations through prompt-driven inference over user interaction sequences. However, this paradigm also introduces new security vulnerabilities, particularly text-level manipulations, rendering them appealing targets for promotion attacks that purposely boost the ranking of specific target items. Although such security risks have been receiving increasing attention, existing studies typically rely on an unrealistic assumption of access to either the victim model or prompt to unveil attack mechanisms. In this work, we investigate the item promotion attack in LLM-SRSs under a more realistic setting where both the system prompt and victim model are unknown to the attacker, and propose a Prompt-Unknown Dual-poisoning Attack (PUDA) framework. To simulate attacks under this full black-box setting, we introduce an LLM-based evolutionary refinement strategy that infers discrete system prompts, enabling the training of an effective surrogate model that mimics the behaviors of the victim model. Leveraging the distilled prompt and surrogate model, we devise a promotion attack that adversarially revises target item texts under semantic constraints, which is further complemented by the highly plausible, surrogate-generated poisoning sequences to enable cost-effective target item promotion. Extensive experiments on real-world datasets demonstrate that PUDA consistently outperforms state-of-the-art competitors in boosting the exposure of unpopular target items. Our findings reveal critical security risks in modern LLM-SRSs even when both prompts and models are protected, and highlight the need for more robust defensive means.

cs.IR↗
arXiv2026

AFMRL: Attribute-Enhanced Fine-Grained Multi-Modal Representation Learning in E-commerce

Multimodal representation is crucial for E-commerce tasks such as identical product retrieval. Large representation models (e.g., VLM2Vec) demonstrate strong multimodal understanding capabilities, yet they struggle with fine-grained semantic comprehension, which is essential for distinguishing highly similar items. To address this, we propose Attribute-Enhanced Fine-Grained Multi-Modal Representation Learning (AFMRL), which defines product fine-grained understanding as an attribute generation task. It leverages the generative power of Multimodal Large Language Models (MLLMs) to extract key attributes from product images and text, and enhances representation learning through a two-stage training framework: 1) Attribute-Guided Contrastive Learning (AGCL), where the key attributes generated by the MLLM are used in the image-text contrastive learning training process to identify hard samples and filter out noisy false negatives. 2) Retrieval-aware Attribute Reinforcement (RAR), where the improved retrieval performance of the representation model post-attribute integration serves as a reward signal to enhance MLLM's attribute generation during multimodal fine-tuning. Extensive experiments on large-scale E-commerce datasets demonstrate that our method achieves state-of-the-art performance on multiple downstream retrieval tasks, validating the effectiveness of harnessing generative models to advance fine-grained representation learning.

cs.CL↗
arXiv2026

Multi-LLM Token Filtering and Routing for Sequential Recommendation

Large language models (LLMs) have recently shown promise in recommendation by providing rich semantic knowledge. While most existing approaches rely on external textual corpora to align LLMs with recommender systems, we revisit a more fundamental yet underexplored question: Can recommendation benefit from LLM token embeddings alone without textual input? Through a systematic empirical study, we show that directly injecting token embeddings from a single LLM into sequential recommenders leads to unstable or limited gains, due to semantic misalignment, insufficient task adaptation, and the restricted coverage of individual LLMs. To address these challenges, we propose MLTFR, a Multi-LLM Token Filtering and Routing framework for corpus-free sequential recommendation. MLTFR follows an interaction-guided LLM knowledge integration paradigm, where task-relevant token embeddings are selected via user-guided token filtering to suppress noisy and irrelevant vocabulary signals. To overcome the limitations of single-LLM representations, MLTFR integrates multiple LLM token spaces through a Mixture-of-Experts architecture, with a Fisher-weighted semantic consensus expert to balance heterogeneous experts and prevent domination during training. By jointly filtering informative tokens and aggregating complementary semantic knowledge across multiple LLMs, MLTFR enables stable and effective utilization of LLM token embeddings without textual inputs or backbone modification. Extensive experiments demonstrate that MLTFR consistently outperforms state-of-the-art sequential recommendation baselines and existing alignment methods. Our code is available at: https://github.com/ccwwhhh/MLTFR.

cs.IR↗
arXiv2026

Self-Distilled Reinforcement Learning for Co-Evolving Agentic Recommender Systems

Large language model-empowered agentic recommender systems (ARS) reformulate recommendation as a multi-turn interaction between a recommender agent and a user agent, enabling iterative preference elicitation and refinement beyond conventional one-shot prediction. However, existing ARS are mainly optimized in a Reflexion-style paradigm, where past interaction trajectories are stored as textual memory and retrieved as prompt context for later reasoning. Although this design allows agents to recall prior feedback and observations, the accumulated experience remains external to model parameters, leaving agents reliant on generic reasoning rather than progressively acquiring recommendation-specific decision-making ability through learning. Reinforcement learning (RL) therefore provides a natural way to internalize such interaction experience into parameters. Yet existing RL methods for ARS still suffer from two key limitations. First, they fail to capture the interactive nature of ARS, in which the recommender agent and the user agent continuously influence each other and can naturally generate endogenous supervision through interaction feedback. Second, they reduce a rich multi-turn interaction process to final outcomes, overlooking the dense supervision embedded throughout the trajectory. To this end, we propose CoARS, a self-distilled reinforcement learning framework for co-evolving agentic recommender systems. CoARS introduces two complementary learning schemes: interaction reward, which derives coupled task-level supervision for the recommender agent and the user agent from the same interaction trajectory, and self-distilled credit assignment, which converts historical trajectories into token-level credit signals under teacher-student conditioning. Experiments on multiple datasets show that CoARS outperforms representative ARS baselines in recommendation performance and user alignment.

cs.IR↗
arXiv2026

Prompt-Induced Linguistic Fingerprints for LLM-Generated Fake News Detection

With the rapid development of large language models, the generation of fake news has become increasingly effortless, posing a growing societal threat and underscoring the urgent need for reliable detection methods. Early efforts to identify LLM-generated fake news have predominantly focused on the textual content itself; however, because much of that content may appear coherent and factually consistent, the subtle traces of falsification are often difficult to uncover. Through distributional divergence analysis, we uncover prompt-induced linguistic fingerprints: statistically distinct probability shifts between LLM-generated real and fake news when maliciously prompted. Based on this insight, we propose a novel method named Linguistic Fingerprints Extraction (LIFE). By reconstructing word-level probability distributions, LIFE can find discriminative patterns that facilitate the detection of LLM-generated fake news. To further amplify these fingerprint patterns, we also leverage key-fragment techniques that accentuate subtle linguistic differences, thereby improving detection reliability. Our experiments show that LIFE achieves state-of-the-art performance in LLM-generated fake news and maintains high performance in human-written fake news. The code and data are available at https://anonymous.4open.science/r/LIFE-E86A.

cs.CL↗