arXiv subjects

Weipeng Zhang

Publications and source records attributed to Weipeng Zhang.

At least 19 recordsLinked to original sources

arXiv2026

Asymmetric On-Policy Distillation: Bridging Exploitation and Imitation at the Token Level

On-policy distillation (OPD) trains a student on its own trajectories with token-level teacher feedback and often outperforms off-policy distillation and standard reinforcement learning. However, we find that its standard advantage weighted policy gradient suffers from three structural weaknesses, including high variance updates, vanishing gradients in zero-advantage regions, and exploration bottlenecks when corrective signals are insufficient. We therefore propose Asymmetric On-Policy Distillation (AOPD), which replaces ineffective negative reinforcement with localized divergence minimization in non-positive advantage regions while preserving positive reinforcement learning. Experiments on mathematical reasoning benchmarks show that AOPD consistently outperforms standard OPD, with average gains of 4.09 / 8.34 under strong / weak initialization, respectively. AOPD also maintains higher policy entropy during training and better capability retention during sequential tool-use adaptation.

cs.LG↗

arXiv2026

Homodyne Photonic Tensor Processor exceeds 1,000-TOPS

High-performance computing underpins modern artificial intelligence (AI), enabling foundation models, real-time inference and perception in autonomous systems, and data-intensive scientific simulations. Recent advances in quantization techniques utilizing low-precision computation without degrading model accuracy, create new opportunities for analog photonic computing characterized by ultra-high clock rates and low energy consumption. Here we propose and demonstrate a coherent homodyne integrated circuit capable of general matrix multiplication (GEMM) with aggregate throughput that exceeds 1,000 TOPS (tera-operations per second), enabled by massive on-chip optical fanout and parallelism. By leveraging time multiplexing, the required modulator count is reduced from O($N^2$) to O(N), allowing dense integration of record-scale 256 $\times$ 256 homodyne units (each <0.0064 $mm^2$) within a single reticle. We employ wafer-scale fabricated 64 thin-film lithium niobate (TFLN) transmitters (each over 40-GHz bandwidth with propagation loss of 0.2 dB/cm) to encode data and chip-to-chip coupled to Si/SiN computing circuits (64 channels). Our system achieves up to 7-bit computational accuracy across 8 $\times$ 8 parallel channels at record computing clockrate 120 Gbaud/s, and 6-bit statistical accuracy across 256 $\times$ 100 channels at 20-128 Gbaud/s, representing a total throughput of 1,000-6,000 TOPS. Massive parallelism amortizes the optoelectronic (OE) conversion to allow 330-TOPS/W efficiency using foundry-available packaging technology. The system throughput is benchmarked with Qwen2.5-0.5 billion parameter models that generate accurate tokens. High throughput and energy efficiency establish a near-term pathway toward light-based accelerators for large-scale training and low-latency inference from datacenters to edges, accelerating new models toward artificial general intelligence.

cs.ET↗

arXiv2026

PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling

Reward models (RMs) are central to reinforcement learning from human feedback (RLHF), providing the critical supervision signals that align large language models (LLMs) with human preferences. Generative reward models (GRMs) provide greater interpretability than traditional scalar RMs, but they come with a critical trade-off: pairwise methods are hindered by a training-inference mismatch, while pointwise methods require expensive absolute annotations. To bridge this gap, we propose the Preference-aware Task-adaptive Reward Model (PaTaRM). Unlike prior approaches, PaTaRM enables robust pointwise training using readily available pairwise data via a novel Preference-Aware Reward (PAR) mechanism, eliminating the need for explicit rating labels. Furthermore, it incorporates a Task-Adaptive Rubric system that dynamically generates instance-specific criteria for precise evaluation. Extensive experiments demonstrate that PATRM achieves a 8.7% average improvement on RewardBench and RMBench across Qwen3-8B/14B models. Crucially, it boosts downstream RLHF performance by an average relative improvement of 13.6% across IFEval and InFoBench, validating its effectiveness for policy alignment. Our code is available at https://github.com/JaneEyre0530/PaTaRM.

cs.LG↗

arXiv2026

Learning Evolving Preferences: A Federated Continual Framework for User-Centric Recommendation

User-centric recommendation has become essential for delivering personalized services, as it enables systems to adapt to users' evolving behaviors while respecting their long-term preferences and privacy constraints. Although federated learning offers a promising alternative to centralized training, existing approaches largely overlook user behavior dynamics, leading to temporal forgetting and weakened collaborative personalization. In this work, we propose FCUCR, a federated continual recommendation framework designed to support long-term personalization in a privacy-preserving manner. To address temporal forgetting, we introduce a time-aware self-distillation strategy that implicitly retains historical preferences during local model updates. To tackle collaborative personalization under heterogeneous user data, we design an inter-user prototype transfer mechanism that enriches each client's representation using knowledge from similar users while preserving individual decision logic. Extensive experiments on four public benchmarks demonstrate the superior effectiveness of our approach, along with strong compatibility and practical applicability. Code is available.

cs.IR↗

arXiv2026

TRUST-SQL: Tool-Integrated Multi-Turn Reinforcement Learning for Text-to-SQL over Unknown Schemas

Text-to-SQL parsing has achieved remarkable progress under the Full Schema Assumption. However, this premise fails in real-world enterprise environments where databases contain hundreds of tables with massive noisy metadata. Rather than injecting the full schema upfront, an agent must actively identify and verify only the relevant subset, giving rise to the Unknown Schema scenario we study in this work. To address this, we propose TRUST-SQL (Truthful Reasoning with Unknown Schema via Tools). We formulate the task as a Partially Observable Markov Decision Process where our autonomous agent employs a structured four-phase protocol to ground reasoning in verified metadata. Crucially, this protocol provides a structural boundary for our novel Dual-Track GRPO strategy. By applying token-level masked advantages, this strategy isolates exploration rewards from execution outcomes to resolve credit assignment, yielding a 9.9% relative improvement over standard GRPO. Extensive experiments across five benchmarks demonstrate that TRUST-SQL achieves an average absolute improvement of 30.6% and 16.6% for the 4B and 8B variants respectively over their base models. Remarkably, despite operating entirely without pre-loaded metadata, our framework consistently matches or surpasses strong baselines that rely on schema prefilling. Data, code, and checkpoints are available at https://huggingface.co/collections/AIJian/trustsql.

cs.AI↗

arXiv2026

Harmonizing Dense and Sparse Signals in Multi-turn RL: Dual-Horizon Credit Assignment for Industrial Sales Agents

Optimizing large language models for industrial sales requires balancing long-term commercial objectives (e.g., conversion rate) with immediate linguistic constraints such as fluency and compliance. Conventional reinforcement learning often merges these heterogeneous goals into a single reward, causing high-magnitude session-level rewards to overwhelm subtler turn-level signals, which leads to unstable training or reward hacking. To address this issue, we propose Dual-Horizon Credit Assignment (DuCA), a framework that disentangles optimization across time scales. Its core, Horizon-Independent Advantage Normalization (HIAN), separately normalizes advantages from turn-level and session-level rewards before fusion, ensuring balanced gradient contributions from both immediate and long-term objectives to the policy update. Extensive experiments with a high-fidelity user simulator show DuCA outperforms the state-of-the-art GRPO baseline, achieving a 6.82% relative improvement in conversion rate, reducing inter-sentence repetition by 82.28%, and lowering identity detection rate by 27.35%, indicating a substantial improvement for an industrial sales scenario that effectively balances the dual demands of strategic performance and naturalistic language generation.

cs.AI↗

arXiv2026

Multimodal-enhanced Federated Recommendation: A Group-wise Fusion Approach

Federated Recommendation (FR) is a new learning paradigm to tackle the learn-to-rank problem in a privacy-preservation manner. How to integrate multi-modality features into federated recommendation is still an open challenge in terms of efficiency, distribution heterogeneity, and fine-grained alignment. To address these challenges, we propose a novel multimodal fusion mechanism in federated recommendation settings (GFMFR). Specifically, it offloads multimodal representation learning to the server, which stores item content and employs a high-capacity encoder to generate expressive representations, alleviating client-side overhead. Moreover, a group-aware item representation fusion approach enables fine-grained knowledge sharing among similar users while retaining individual preferences. The proposed fusion loss could be simply plugged into any existing federated recommender systems empowering their capability by adding multi-modality features. Extensive experiments on five public benchmark datasets demonstrate that GFMFR consistently outperforms state-of-the-art multimodal FR baselines.

cs.IR↗

arXiv2026

Compact, Large-Scale Photonic Neurons by Modulation-and-Weight Microring Resonators

Neuromorphic photonics promises sub-nanosecond latency, ultrawide bandwidth, and high parallelism, but practical scalability is constrained by fabrication tolerances, spectral alignment, and tuning energy. Here, we present a large-scale, compact, and reconfigurable photonic neuron in which each microring performs modulation and weighting simultaneously. By exploiting both carrier and thermal tuning within a single device, this architecture reduces footprint, relaxes spectral alignment requirements to just two optical components, and yields a steep transfer response that lowers tuning energy. The proposed neuron supports multiple operating configurations, allowing its dynamical behavior to be adapted to different computational tasks. In particular, a short electrical feedback path enables recurrent operation, providing tunable short- and long-term memory for temporal processing. Using a 10-microring resonator array, we demonstrate both spatial and temporal computing, including a 3$\times$3 convolution for image processing with an error of $<$5\% and high-frequency financial time-series prediction. Each modulation-weighting element occupies 80$\times$45 \SI{}{\micro\meter^2} and consumes an average of \SI{0.186}{\milli\watt}, corresponding to a compute density of \SI{4.67}{TOPS/s/\milli\meter^2}. Excluding electronic power, the on-chip tuning efficiency reaches approximately \SI{105}{TOPs/\watt}, which is comparable to state-of-the-art implementations. These results indicate that modulation-and-weighting microring resonator banks provide a scalable building block for large-scale neuromorphic photonic systems, offering a favorable combination of compact footprint, low power consumption, and functional flexibility.

physics.optics↗

arXiv2025

SkyMoE: A Vision-Language Foundation Model for Enhancing Geospatial Interpretation with Mixture of Experts

The emergence of large vision-language models (VLMs) has significantly enhanced the efficiency and flexibility of geospatial interpretation. However, general-purpose VLMs remain suboptimal for remote sensing (RS) tasks. Existing geospatial VLMs typically adopt a unified modeling strategy and struggle to differentiate between task types and interpretation granularities, limiting their ability to balance local detail perception and global contextual understanding. In this paper, we present SkyMoE, a Mixture-of-Experts (MoE) vision-language model tailored for multimodal, multi-task RS interpretation. SkyMoE employs an adaptive router that generates task- and granularity-aware routing instructions, enabling specialized large language model experts to handle diverse sub-tasks. To further promote expert decoupling and granularity sensitivity, we introduce a context-disentangled augmentation strategy that creates contrastive pairs between local and global features, guiding experts toward level-specific representation learning. We also construct MGRS-Bench, a comprehensive benchmark covering multiple RS interpretation tasks and granularity levels, to evaluate generalization in complex scenarios. Extensive experiments on 21 public datasets demonstrate that SkyMoE achieves state-of-the-art performance across tasks, validating its adaptability, scalability, and superior multi-granularity understanding in remote sensing.

cs.CV↗

arXiv2025

PC-UNet: An Enforcing Poisson Statistics U-Net for Positron Emission Tomography Denoising

Positron Emission Tomography (PET) is crucial in medicine, but its clinical use is limited due to high signal-to-noise ratio doses increasing radiation exposure. Lowering doses increases Poisson noise, which current denoising methods fail to handle, causing distortions and artifacts. We propose a Poisson Consistent U-Net (PC-UNet) model with a new Poisson Variance and Mean Consistency Loss (PVMC-Loss) that incorporates physical data to improve image fidelity. PVMC-Loss is statistically unbiased in variance and gradient adaptation, acting as a Generalized Method of Moments implementation, offering robustness to minor data mismatches. Tests on PET datasets show PC-UNet improves physical consistency and image fidelity, proving its ability to integrate physical information effectively.

cs.CV↗

arXiv2025

Online training and pruning of multi-wavelength photonic neural networks

CMOS-compatible photonic integrated circuits (PICs) are emerging as a promising platform in artificial intelligence (AI) computing. Owing to the compact footprint of microring resonators (MRRs) and the enhanced interconnect efficiency enabled by wavelength division multiplexing (WDM), MRR-based photonic neural networks (PNNs) are particularly promising for large-scale integration. However, the scalability and energy efficiency of such systems are fundamentally limited by the MRR resonance wavelength variations induced by fabrication process variations (FPVs) and environmental fluctuations. Existing solutions use post-fabrication approaches or thermo-optic tuning, incurring high control power and additional process complexity. In this work, we introduce an online training and pruning method that addresses this challenge, adapting to FPV-induced and thermally induced shifts in MRR resonance wavelength. By incorporating a power-aware pruning term into the conventional loss function, our approach simultaneously optimizes the PNN accuracy and the total power consumption for MRR tuning. In proof-of-concept on-chip experiments on the Iris dataset, our system PNNs can adaptively train to maintain a 96% classification accuracy, while achieving a 44.7% reduction in tuning power via pruning. Additionally, our approach reduces the power consumption by orders-of-magnitude on larger datasets. By addressing chip-to-chip variation and minimizing power requirements, our approach significantly improves the scalability and energy efficiency of MRR-based integrated analog photonic processors, paving the way for large-scale PICs to enable versatile applications including neural networks, photonic switching, LiDAR, and radio-frequency beamforming.

physics.optics↗

arXiv2025

LLM4Tag: Automatic Tagging System for Information Retrieval via Large Language Models

Tagging systems play an essential role in various information retrieval applications such as search engines and recommender systems. Recently, Large Language Models (LLMs) have been applied in tagging systems due to their extensive world knowledge, semantic understanding, and reasoning capabilities. Despite achieving remarkable performance, existing methods still have limitations, including difficulties in retrieving relevant candidate tags comprehensively, challenges in adapting to emerging domain-specific knowledge, and the lack of reliable tag confidence quantification. To address these three limitations above, we propose an automatic tagging system LLM4Tag. First, a graph-based tag recall module is designed to effectively and comprehensively construct a small-scale highly relevant candidate tag set. Subsequently, a knowledge-enhanced tag generation module is employed to generate accurate tags with long-term and short-term knowledge injection. Finally, a tag confidence calibration module is introduced to generate reliable tag confidence scores. Extensive experiments over three large-scale industrial datasets show that LLM4Tag significantly outperforms the state-of-the-art baselines and LLM4Tag has been deployed online for content tagging to serve hundreds of millions of users.

cs.IR↗

arXiv2024

Control-free and efficient integrated photonic neural networks via hardware-aware training and pruning

Integrated photonic neural networks (PNNs) are at the forefront of AI computing, leveraging on light's unique properties, such as large bandwidth, low latency, and potentially low power consumption. Nevertheless, the integrated optical components within PNNs are inherently sensitive to external disturbances and thermal interference, which can detrimentally affect computing accuracy and reliability. Current solutions often use complicated control methods, resulting in high hardware complexity impractical for large-scale PNNs. In response, we propose a novel hardware-aware training and pruning approach. The core idea is to train the parameters of a physical neural network towards its noise-robust and energy-efficient region. This innovation enables control-free and energy-efficient photonic computing. Our method is validated across diverse integrated PNN architectures. Through experimental validation, our approach significantly enhances the computing precision of MRR-based PNN, achieving a notable 4-bit improvement without the need for complex device control mechanisms or energy-intensive temperature stabilization circuits. Specifically, it improves the accuracy of experimental handwritten digit classification from 67.0% to 95.0%, nearing theoretical limits and achieved without a thermoelectric controller. Additionally, this approach reduces the energy by tenfold. We further extend the validation to various architectures, such as PCM-based PNN, demonstrating the broad applicability of our approach across different platforms. This advancement represents a significant step towards the practical, energy-efficient, and noise-resilient implementation of large-scale integrated PNNs.

physics.optics↗

arXiv2023

A system-on-chip microwave photonic processor solves dynamic RF interference in real time with picosecond latency

Radio-frequency interference is a growing concern as wireless technology advances, with potentially life-threatening consequences like interference between radar altimeters and 5G cellular networks. Mobile transceivers mix signals with varying ratios over time, posing challenges for conventional digital signal processing (DSP) due to its high latency. These challenges will worsen as future wireless technologies adopt higher carrier frequencies and data rates. However, conventional DSPs, already on the brink of their clock frequency limit, are expected to offer only marginal speed advancements. This paper introduces a photonic processor to address dynamic interference through blind source separation (BSS). Our system-on-chip processor employs a fully integrated photonic signal pathway in the analogue domain, enabling rapid demixing of received mixtures and recovering the signal-of-interest in under 15 picoseconds. This reduction in latency surpasses electronic counterparts by more than three orders of magnitude. To complement the photonic processor, electronic peripherals based on field-programmable gate array (FPGA) assess the effectiveness of demixing and continuously update demixing weights at a rate of up to 305 Hz. This compact setup features precise dithering weight control, impedance-controlled circuit board and optical fibre packaging, suitable for handheld and mobile scenarios. We experimentally demonstrate the processor's ability to suppress transmission errors and maintain signal-to-noise ratios in two scenarios, radar altimeters and mobile communications. This work pioneers the real-time adaptability of integrated silicon photonics, enabling online learning and weight adjustments, and showcasing practical operational applications for photonic processing.

physics.optics↗

arXiv2023

Real-Time Blind Photonic Interference Cancellation for mmWave MIMO

Multiple-input multiple-output (MIMO) mmWave devices broadcast multiple spatially-separated data streams simultaneously in order to increase data transfer rates. Data transfer can, however, be compromised by interference. Conventional techniques for mitigating interference require additional space and power not generally available in handheld mobile devices. Here, we propose a photonic mmWave MIMO receiver architecture capable of interference cancellation with greatly reduced space and power needs. We demonstrate real-time photonic interference cancellation with an integrated FPGA-photonic system that executes a novel zero-calibration micro-ring resonator control algorithm. The system achieves sub-second cancellation weight determination latency with sub-Nyquist sampling. We evaluate the impact of canceller design parameters on performance, establishing that effective photonic cancellation is possible in handheld devices with less than 30 ms weight determination latency.

eess.SP↗

arXiv2022

Broadband physical layer cognitive radio with an integrated photonic processor for blind source separation

The expansion of telecommunications incurs increasingly severe crosstalk and interference, and a physical layer cognitive method, called blind source separation (BSS), can effectively address these issues. BSS requires minimal prior knowledge to recover signals from their mixtures, agnostic to carrier frequency, signal format, and channel conditions. However, previous electronic implementations of BSS did not fulfill this versatility requirement due to the inherently narrow bandwidth of radio-frequency (RF) components, the high energy consumption of digital signal processors (DSP), and their shared weaknesses of low scalability. Here, we report a photonic BSS approach that inherits the advantages of optical devices and can fully fulfill its "blindness" aspect. Using a microring weight bank integrated on a photonic chip, we demonstrate energy-efficient, WDM-scalable BSS across 19.2 GHz of bandwidth, covering many standard frequency bands. Our system also has a high (9-bit) resolution for signal demixing thanks to a recently developed dithering control method, resulting in higher signal-to-interference ratios (SIR) even for ill-conditioned mixtures.

eess.SP↗

arXiv2022

Silicon microring synapses enable photonic deep learning beyond 9-bit precision

Deep neural networks (DNN) consist of layers of neurons interconnected by synaptic weights. A high bit-precision in weights is generally required to guarantee high accuracy in many applications. Minimizing error accumulation between layers is also essential when building large-scale networks. Recent demonstrations of photonic neural networks are limited in bit-precision due to crosstalk and the high sensitivity of optical components (e.g., resonators). Here, we experimentally demonstrate a record-high precision of 9 bits with a dithering control scheme for photonic synapses. We then numerically simulated the impact with increased synaptic precision on a wireless signal classification application. This work could help realize the potential of photonic neural networks for many practical, real-world tasks.

physics.app-ph↗

arXiv2022

Confidence Calibration for Intent Detection via Hyperspherical Space and Rebalanced Accuracy-Uncertainty Loss

Data-driven methods have achieved notable performance on intent detection, which is a task to comprehend user queries. Nonetheless, they are controversial for over-confident predictions. In some scenarios, users do not only care about the accuracy but also the confidence of model. Unfortunately, mainstream neural networks are poorly calibrated, with a large gap between accuracy and confidence. To handle this problem defined as confidence calibration, we propose a model using the hyperspherical space and rebalanced accuracy-uncertainty loss. Specifically, we project the label vector onto hyperspherical space uniformly to generate a dense label representation matrix, which mitigates over-confident predictions due to overfitting sparce one-hot label matrix. Besides, we rebalance samples of different accuracy and uncertainty to better guide model training. Experiments on the open datasets verify that our model outperforms the existing calibration methods and achieves a significant improvement on the calibration metric.

cs.CL↗