arXiv subjects

Wei Wei

Publications and source records attributed to Wei Wei.

At least 91 records · Page 5Linked to original sources

arXiv2026

Suppressing Acoustomigration and Temperature Rise for High-power Robust Acoustics

High-frequency acoustic wave transducers, vibrating at gigahertz (GHz), favored for their compact size, are not only dominating the front-end of mobile handsets but are also expanding into various interdisciplinary fields, including quantum acoustics, acoustic-optics, acoustic-fluids, acoustoelectric, and sustainable power conversion systems. However, like strong vibration can "shake off" substances and produce heat, a long-standing bottleneck has been the ability to harness acoustics under high-power vibration loads, while simultaneously suppressing temperature rise, especially for IDT-based surface acoustic wave (SAW) systems. Here, we proposed a layered acoustic wave (LAW) platform, utilizing a quasi-infinite multifunctional top layer, that redefines mechanical and thermal boundary conditions to overcome three fundamental challenges in high-power acoustic wave vibration: self-heating, thermal instability, and acoustomigration. By simply leveraging a simplified, thick single-material overlayer to achieve electro-thermo-mechanical co-design, this acoustic platform moves beyond prior substrate-focused thermal management in SAW technology. It demonstrates, for the first time from the top boundary, simultaneous redistribution of the von Mises stress field and the creation of an efficient vertical thermal dissipation path. The LAW transducer, vibrating at over 2 GHz, achieves a 70% reduction in temperature rise under identical power loads, a first-order temperature coefficient of frequency (TCF) of -13 ppm/C with minimal dispersion, and an unprecedented threshold power density of 45.61 dBm/mm2 - over one order-of-magnitude higher than that of state-of-the-art thin-film surface acoustic wave (TF-SAW) counterparts at the same wavelength.

eess.SP↗

arXiv2026

Ares: Adaptive Reasoning Effort Selection for Efficient LLM Agents

Modern agents powered by thinking LLMs achieve high accuracy through long chain-of-thought reasoning but incur substantial inference costs. While many LLMs now support configurable reasoning levels (e.g., high/medium/low), static strategies are often ineffective: using low-effort modes at every step leads to significant performance degradation, while random selection fails to preserve accuracy or provide meaningful cost reduction. However, agents should reserve high reasoning effort for difficult steps like navigating complex website structures, while using lower-effort modes for simpler steps like opening a target URL. In this paper, we propose Ares, a framework for per-step dynamic reasoning effort selection tailored for multi-step agent tasks. Ares employs a lightweight router to predict the lowest appropriate reasoning level for each step based on the interaction history. To train this router, we develop a data generation pipeline that identifies the minimum reasoning effort required for successful step completion. We then fine-tune the router to predict these levels, enabling plug-and-play integration for any LLM agents. We evaluate Ares on a diverse set of agent tasks, including TAU-Bench for tool use agents, BrowseComp-Plus for deep-research agents, and WebArena for web agents. Experimental results show that Ares reduces reasoning token usage by up to 52.7% compared to fixed high-effort reasoning, while introducing minimal degradation in task success rates.

cs.AI↗

arXiv2026

Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory

Large language model (LLM) agents are fundamentally bottlenecked by finite context windows on long-horizon tasks. As trajectories grow, retaining tool outputs and intermediate reasoning in-context quickly becomes infeasible: the working context becomes prohibitively long, eventually exceeds the context budget, and makes distant evidence harder to use even when it is still present. Existing solutions typically shorten context through truncation or running summaries, but these methods are fundamentally lossy because they compress or discard past evidence itself. We introduce Memex, an indexed experience memory mechanism that instead compresses context without discarding evidence. Memex maintains a compact working context consisting of concise structured summaries and stable indices, while storing full-fidelity underlying interactions in an external experience database under those indices. The agent can then decide when to dereference an index and recover the exact past evidence needed for the current subgoal. We optimize both write and read behaviors with our reinforcement learning framework MemexRL, using reward shaping tailored to indexed memory usage under a context budget, so the agent learns what to summarize, what to archive, how to index it, and when to retrieve it. This yields a substantially less lossy form of long-horizon memory than summary-only approaches. We further provide a theoretical analysis showing the potential of the Memex loop to preserve decision quality with bounded dereferencing while keeping effective in-context computation bounded as history grows. Empirically, on challenging long-horizon tasks, Memex agent trained with MemexRL improves task success while using a significantly smaller working context.

cs.CL↗

arXiv2026

Make LoRA Great Again: Boosting LoRA with Adaptive Singular Values and Mixture-of-Experts Optimization Alignment

While Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning for Large Language Models (LLMs), its performance often falls short of Full Fine-Tuning (Full FT). Current methods optimize LoRA by initializing with static singular value decomposition (SVD) subsets, leading to suboptimal leveraging of pre-trained knowledge. Another path for improving LoRA is incorporating a Mixture-of-Experts (MoE) architecture. However, weight misalignment and complex gradient dynamics make it challenging to adopt SVD prior to the LoRA MoE architecture. To mitigate these issues, we propose \underline{G}reat L\underline{o}R\underline{A} Mixture-of-Exper\underline{t} (GOAT), a framework that (1) adaptively integrates relevant priors using an SVD-structured MoE, and (2) aligns optimization with full fine-tuned MoE by deriving a theoretical scaling factor. We demonstrate that proper scaling, without modifying the architecture or training algorithms, boosts LoRA MoE's efficiency and performance. Experiments across 25 datasets, including natural language understanding, commonsense reasoning, image classification, and natural language generation, demonstrate GOAT's state-of-the-art performance, closing the gap with Full FT.

cs.CL↗

arXiv2026

Small Object Detection Model with Spatial Laplacian Pyramid Attention and Multi-Scale Features Enhancement in Aerial Images

Detecting objects in aerial images confronts some significant challenges, including small size, dense and non-uniform distribution of objects over high-resolution images, which makes detection inefficient. Thus, in this paper, we proposed a small object detection algorithm based on a Spatial Laplacian Pyramid Attention and Multi-Scale Feature Enhancement in aerial images. Firstly, in order to improve the feature representation of ResNet-50 on small objects, we presented a novel Spatial Laplacian Pyramid Attention (SLPA) module, which is integrated after each stage of ResNet-50 to identify and emphasize important local regions. Secondly, to enhance the model's semantic understanding and features representation, we designed a Multi-Scale Feature Enhancement Module (MSFEM), which is incorporated into the lateral connections of C5 layer for building Feature Pyramid Network (FPN). Finally, the features representation quality of traditional feature pyramid network will be affected because the features are not aligned when the upper and lower layers are fused. In order to handle it, we utilized deformable convolutions to align the features in the fusion processing of the upper and lower levels of the Feature Pyramid Network, which can help enhance the model's ability to detect and recognize small objects. The extensive experimental results on two benchmark datasets: VisDrone and DOTA demonstrate that our improved model performs better for small object detection in aerial images compared to the original algorithm.

cs.CV↗

arXiv2026

Training-Free Agentic AI: Probabilistic Control and Coordination in Multi-Agent LLM Systems

Multi-agent large language model (LLM) systems enable complex, long-horizon reasoning by composing specialized agents, but practical deployment remains hindered by inefficient routing, noisy feedback, and high interaction cost. We introduce REDEREF, a lightweight and training-free controller for multi-agent LLM collaboration that improves routing efficiency during recursive delegation. REDEREF integrates (i) belief-guided delegation via Thompson sampling to prioritize agents with historically positive marginal contributions, (ii) reflection-driven re-routing using a calibrated LLM or programmatic judge, (iii) evidence-based selection rather than output averaging, and (iv) memory-aware priors to reduce cold-start inefficiency. Across multi-agent split-knowledge tasks, we show that while recursive retry alone saturates task success, belief-guided routing reduces token usage by 28%, agent calls by 17%, and time-to-success by 19% compared to random recursive delegation, and adapts gracefully under agent or judge degradation. These results demonstrate that simple, interpretable probabilistic control can meaningfully improve the efficiency and robustness of multi-agent LLM systems without training or fine-tuning.

cs.CL↗

arXiv2026

The Fuzzy Front Ends: Reflections on the Never-Ending Story of Visualization Co-Design

Co-design is an increasingly popular approach in HCI and visualization, yet there is little guidance on how to effectively apply this method in visualization contexts. In this paper, we visually present our experience of a two-and-a-half-year co-design project with the local arts community. Focusing on facilitating community exploration and sense-making around arts funding distribution, the project involved a series of co-design sessions between visualization researchers and members of the arts community. Through these iterative sessions, we built shared understanding and developed visualization prototypes tailored to community needs. However, the practice is far from complete, and we found ourselves continually returning to the "fuzzy front end" of the co-design process. We share this ongoing story through comic-style visuals and reflect on three fuzzy front ends that we encountered during the project. By sharing these experiences with the visualization community, we hope to offer insights that others can draw on in their own community-engaged co-design work.

cs.HC↗

arXiv2026

Performance and pulse shape discrimination of glass scintillator SG101 for neutron detection

We present a detailed characterization of the thermal neutron sensitive transparent glass scintillator SG101, benchmarked against the conventional LiF ZnS(Ag)based scintillator EJ426. The detection efficiency, energy resolution, and pulse shape discrimination (PSD) performance ofSG101 were evaluated under AmBe neutron irradiation. When coupled with organic scintillators(EJ200 or EJ276),the SG101 EJ200 system achieves a figure of merit (FOM) of 3.81 for thermal neutron/gamma separation, while the SG101 EJ276 configuration resolves three distinct particle populations gamma rays, fast neutrons, and thermal neutrons with FOM values of 3.46 and2.21, respectively. Correlation analysis reveals that the number of fast thermal neutron coincidence events significantly exceeds the accidental background, and the count of gamma fast thermal neutron triple-coincidence events is also far higher than the expected accidental rate, confirming significant physical correlations for both event types within a 100 us time window. These results demonstrate that SG101 is a promising candidate for applications requiring high-efficiency thermal neutron detection and precise event tagging coupling with a scintillator with PSD approach

physics.ins-det↗

arXiv2026

EpicCBR: Item-Relation-Enhanced Dual-Scenario Contrastive Learning for Cold-Start Bundle Recommendation

Bundle recommendation aims to recommend a set of items to users for overall consumption. Existing bundle recommendation models primarily depend on observed user-bundle interactions, limiting exploration of newly-emerged bundles that are constantly created. It pose a critical representation challenge for current bundle methods, as they usually treat each bundle as an independent instance, while neglecting to fully leverage the user-item (UI) and bundle-item (BI) relations over popular items. To alleviate it, in this paper we propose a multi-view contrastive learning framework for cold-start bundle recommendation, named EpicCBR. Specifically, it precisely mine and utilize the item relations to construct user profiles, identifying users likely to engage with bundles. Additionally, a popularity-based method that characterizes the features of new bundles through historical bundle information and user preferences is proposed. To build a framework that demonstrates robustness in both cold-start and warm-start scenarios, a multi-view graph contrastive learning framework capable of integrating these diverse scenarios is introduced to ensure the model's generalization capability. Extensive experiments conducted on three popular benchmarks showed that EpicCBR outperforms state-of-the-art by a large margin (up to 387%), sufficiently demonstrating the superiority of the proposed method in cold-start scenario. The code and dataset can be found in the GitHub repository: https://github.com/alexlovecoding/EpicCBR.

cs.IR↗

arXiv2026

Learning from Literature: Integrating LLMs and Bayesian Hierarchical Modeling for Oncology Trial Design

Designing modern oncology trials requires synthesizing evidence from prior studies to inform hypothesis generation and sample size determination. Trial designs based on incomplete or imprecise summaries can lead to misspecified hypotheses and underpowered studies, resulting in false positive or negative conclusions. To address this challenge, we developed LEAD-ONC (Literature to Evidence for Analytics and Design in Oncology), an AI-assisted framework that transforms published clinical trial reports into quantitative, design-relevant evidence. Given expert-curated trial publications that meet prespecified eligibility criteria, LEAD-ONC uses large language models to extract baseline characteristics and reconstruct individual patient data from Kaplan-Meier curves, followed by Bayesian hierarchical modeling to generate predictive survival distributions for a prespecified target trial population. We demonstrate the framework using five phase III trials in first-line non-small-cell lung cancer evaluating PD-1 or PD-L1 inhibitors with or without CTLA-4 blockade. Clustering based on baseline characteristics identified three clinically interpretable populations defined by histology. For a prospective randomized trial in the mixed-histology population comparing mono versus dual immune checkpoint inhibition, LEAD-ONC projected a modest median overall survival difference of 2.8 months (95 percent credible interval -2.0 to 7.6) and an estimated probability of at least a 3-month benefit of approximately 0.45. As LEAD-ONC remains under active development, these results are intended as preliminary demonstrations of the frameworks potential to support evidence-driven oncology trial design rather than definitive clinical conclusions.

stat.AP↗

arXiv2026

HumanLM: Simulating Users with State Alignment Beats Response Imitation

Large Language Models (LLMs) are increasingly used to simulate how specific users respond to a given context, enabling more user-centric applications that rely on user feedback. However, existing user simulators mostly imitate surface-level patterns and language styles, which fail to reflect the underlying states of real users (e.g., beliefs and emotions). To address these limitations, we propose a novel training framework, HumanLM, which builds user simulators that accurately reflect real users. Our key insight is that, in addition to generating responses, the model should generate natural-language latent states that align with ground-truth responses through reinforcement learning. These latent states correspond to a set of psychologically grounded state dimensions that drive how real users respond. HumanLM further synthesizes these aligned latent states into responses that accurately represent real users. For extensive evaluation, we develop Humanual, a comprehensive benchmark for simulating real users based on public data. Humanual consists of six large-scale datasets with 26k users and 216k responses in total, spanning diverse tasks such as generating user responses to daily life issues, political blogs, and chat sessions with LLM assistants. Across datasets, HumanLM significantly outperforms alternative approaches, achieving an average relative improvement of 16.3% in alignment scores from an LLM judge. In a real-time simulation study with 111 participants, HumanLM achieves the highest similarity to real user responses and competitive human-likeness scores.

cs.CL↗

arXiv2026

CompactRAG: Reducing LLM Calls and Token Overhead in Multi-Hop Question Answering

Retrieval-augmented generation (RAG) has become a key paradigm for knowledge-intensive question answering. However, existing multi-hop RAG systems remain inefficient, as they alternate between retrieval and reasoning at each step, resulting in repeated LLM calls, high token consumption, and unstable entity grounding across hops. We propose CompactRAG, a simple yet effective framework that decouples offline corpus restructuring from online reasoning. In the offline stage, an LLM reads the corpus once and converts it into an atomic QA knowledge base, which represents knowledge as minimal, fine-grained question-answer pairs. In the online stage, complex queries are decomposed and carefully rewritten to preserve entity consistency, and are resolved through dense retrieval followed by RoBERTa-based answer extraction. Notably, during inference, the LLM is invoked only twice in total - once for sub-question decomposition and once for final answer synthesis - regardless of the number of reasoning hops. Experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue demonstrate that CompactRAG achieves competitive accuracy while substantially reducing token consumption compared to iterative RAG baselines, highlighting a cost-efficient and practical approach to multi-hop reasoning over large knowledge corpora. The implementation is available at GitHub.

cs.CL↗

arXiv2026

Med3D-R1: Incentivizing Clinical Reasoning in 3D Medical Vision-Language Models for Abnormality Diagnosis

Developing 3D vision-language models with robust clinical reasoning remains a challenge due to the inherent complexity of volumetric medical imaging, the tendency of models to overfit superficial report patterns, and the lack of interpretability-aware reward designs. In this paper, we propose Med3D-R1, a reinforcement learning framework with a two-stage training process: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). During SFT stage, we introduce a residual alignment mechanism to bridge the gap between high-dimensional 3D features and textual embeddings, and an abnormality re-weighting strategy to emphasize clinically informative tokens and reduce structural bias in reports. In RL stage, we redesign the consistency reward to explicitly promote coherent, step-by-step diagnostic reasoning. We evaluate our method on medical multiple-choice visual question answering using two 3D diagnostic benchmarks, CT-RATE and RAD-ChestCT, where our model attains state-of-the-art accuracies of 41.92\% on CT-RATE and 44.99\% on RAD-ChestCT. These results indicate improved abnormality diagnosis and clinical reasoning and outperform prior methods on both benchmarks. Overall, our approach holds promise for enhancing real-world diagnostic workflows by enabling more reliable and transparent 3D medical vision-language systems.

cs.CV↗

arXiv2026

Locatability and Locatability Robustness of Visual Variables in Single Target Localization

Finding a particular object in a display is important for viewers in many visualizations, for example, when reacting to brushing or to a highlighted object. This can be enabled by making the target object different in one of the visual variables that determine the object's appearance; for example, by changing its color or size. Certain interpretations of the visual search literature have promoted the view that using visual variables such as hue-often labeled as preattentive-would make the target object automatically "popout," implying that an object can be located almost instantly, regardless of the number of objects in the display. In this paper we present a study that serves as a bridge between the extensive visual search literature and visualization, establishing empirical base measurements for the localization task. By testing displays with up to hundreds of objects, we are able to show that none of the common visual variables is immune to the increase in the number of objects. We also provide the first empirically informed comparisons between visual variables for this task in the context of visualization, and show how different visual variables have varying robustness with respect to two additional dimensions: the location of the target and the overall visual arrangement (layout). A free copy of this paper and all supplemental materials are available on our online repository: https://osf.io/z68ak/overview.

cs.HC↗

arXiv2026

No Other Representation Component Is Needed: Diffusion Transformers Can Provide Representation Guidance by Themselves

Recent studies have demonstrated that learning a meaningful internal representation can accelerate generative training. However, existing approaches necessitate to either introduce an off-the-shelf external representation task or rely on a large-scale, pre-trained external representation encoder to provide representation guidance during the training process. In this study, we posit that the unique discriminative process inherent to diffusion transformers enables them to offer such guidance without requiring external representation components. We propose SelfRepresentation Alignment (SRA), a simple yet effective method that obtains representation guidance using the internal representations of learned diffusion transformer. SRA aligns the latent representation of the diffusion transformer in the earlier layer conditioned on higher noise to that in the later layer conditioned on lower noise to progressively enhance the overall representation learning during only the training process. Experimental results indicate that applying SRA to DiTs and SiTs yields consistent performance improvements, and largely outperforms approaches relying on auxiliary representation task. Our approach achieves performance comparable to methods that are dependent on an external pre-trained representation encoder, which demonstrates the feasibility of acceleration with representation alignment in diffusion transformers themselves.

cs.CV↗

arXiv2026

Optimizing Network Topology Efficiency: A Resource-Centric Analysis of Non-Blocking Architectures

In modern network design, "efficiency" is often conflated with raw performance metrics like latency or aggregate throughput. This paper proposes a resource-centric definition of efficiency, isolating the hardware cost required to maintain a non-blocking throughput constraint. By modeling network cost as a function of the Traffic Multiplier (Hop Count) and Router Complexity (Radix), we demonstrate that the optimal topology is determined by the technological ratio between link interface costs ($α$), crossbar switching costs ($β$), and the network concentration ratio. We conclude that while high-radix direct networks optimize efficiency at small to medium scales, indirect networks (e.g., Fat Trees) are required to cap router complexity at massive scales. Furthermore, we posit that redundancy is most efficiently handled via parallel network instances (e.g., multi-plane Star networks) rather than intrinsic topological path diversity.

cs.NI↗

arXiv2026

Stable-DiffCoder: Pushing the Frontier of Code Diffusion Large Language Model

Diffusion-based language models (DLLMs) offer non-sequential, block-wise generation and richer data reuse compared to autoregressive (AR) models, but existing code DLLMs still lag behind strong AR baselines under comparable budgets. We revisit this setting in a controlled study and introduce Stable-DiffCoder, a block diffusion code model that reuses the Seed-Coder architecture, data, and training pipeline. To enable efficient knowledge learning and stable training, we incorporate a block diffusion continual pretraining (CPT) stage enhanced by a tailored warmup and block-wise clipped noise schedule. Under the same data and architecture, Stable-DiffCoder overall outperforms its AR counterpart on a broad suite of code benchmarks. Moreover, relying only on the CPT and supervised fine-tuning stages, Stable-DiffCoder achieves stronger performance than a wide range of \~8B ARs and DLLMs, demonstrating that diffusion-based training can improve code modeling quality beyond AR training alone. Moreover, diffusion-based any-order modeling improves structured code modeling for editing and reasoning, and through data augmentation, benefits low-resource coding languages.

cs.CL↗

arXiv2026

AERO: Adaptive and Efficient Runtime-Aware OTA Updates for Energy-Harvesting IoT

Energy-harvesting (EH) Internet of Things (IoT) devices operate under intermittent energy availability, which disrupts task execution and makes energy-intensive over-the-air (OTA) updates particularly challenging. Conventional OTA update mechanisms rely on reboots and incur significant overhead, rendering them unsuitable for intermittently powered systems. Recent live OTA update techniques reduce reboot overhead but still lack mechanisms to ensure consistency when updates interact with runtime execution. This paper presents AERO, an Adaptive and Efficient Runtime-Aware OTA update mechanism that integrates update tasks into the device's Directed Acyclic Graph (DAG) and schedules them alongside routine tasks under energy and timing constraints. By identifying update-affected execution regions and dynamically adjusting dependencies, AERO ensures consistent up date integration while adapting to intermittent energy availability. Experiments on representative workloads demonstrate improved update reliability and efficiency compared to existing live update approaches.

cs.AR↗