arXiv subjects

Tianyu Dong

Publications and source records attributed to Tianyu Dong.

At least 19 recordsLinked to original sources

arXiv2026

$M^2PO$: Multi-Perspective Multi-Pair Preference Optimization for Machine Translation

Aligning Large Language Models (LLMs) with human preferences is pivotal for Machine Translation (MT), yet current approaches are often hindered by misleading reward signals. Our analysis reveals that prevailing Quality Estimation (QE) models exhibit a systematic blind spot toward partial errors, specifically partial hallucinations and omissions, often favoring superficially fluent but unfaithful translations. To address this issue, we propose $M^2PO$ (Multi-Perspective Multi-Pair Preference Optimization), a data-centric framework for preference optimization in machine translation. First, to correct the bias toward fluency, $M^2PO$ uses a dual-perspective mechanism that decouples semantic fidelity from fluency and prioritizes faithfulness through a curriculum strategy. Second, after correcting this bias, partial errors fall between perfect and severely incorrect translations, making them difficult to learn through standard best-versus-worst comparisons. We therefore introduce a multi-pair objective that leverages the full candidate list to capture these fine-grained error signals. Experiments on WMT23, WMT24, and FLORES-200 show that $M^2PO$ enables a 9B model to outperform leading open-source baselines and achieve parity with proprietary models such as GPT-4o and Gemini-2.0-Flash, demonstrating strong potential for efficient and high-fidelity LLM-based translation. Our code and dataset will be released.

cs.CL↗

arXiv2026

System optimality versus self-organization in metro networks: An optimal transport analysis

Urban transportation network design is typically approached through top-down planning grounded in engineering and economics. Yet cities are complex systems characterized by feedback loops between infrastructure and mobility demand: network structure shapes origin-destination (OD) flows, while OD flows adapt and in turn affect network development. Despite this interdependence, computational tools to benchmark real-world networks against alternative generative principles remain limited. Here, we introduce a data-driven framework to compare three types of metro networks: a self-organized (desire-path) network derived from local rules, a system-optimal network derived from global optimization, and the empirical real-world network. We quantify similarities and discrepancies using geometric comparisons and optimal transport theory. Applying the framework to Singapore's Mass Rapid Transit system, we find that the empirical metro network is substantially closer to the self-organized benchmark than to the system-optimal benchmark, with remaining discrepancies largely explained by geographic constraints. These findings highlight that system-wide optimality alone may be inadequate for guiding practical interventions, motivating planning approaches that explicitly incorporate local service needs. The framework is transferable across regions and can diagnose design-use misalignments to support adaptive infrastructure planning.

physics.soc-ph↗

arXiv2026

SARA: Unlocking Multilingual Knowledge in Mixture-of-Experts via Semantically Anchored Routing Alignment

Sparse Mixture-of-Experts (MoE) architectures have emerged as an increasingly influential paradigm as they offer a strategic balance between parameter scalability and computational efficiency. However, low-resource languages, which suffer from a scarcity of high-quality training data, often have their tokens routed to different experts than those predominantly activated by high-resource inputs, which limits cross-lingual expert sharing. This cross-lingual routing divergence consequently hinders their efficacy in multilingual contexts. To address this issue, we propose SARA (Semantically Anchored Routing Alignment), a framework designed to transfer specialized capabilities from high-resource languages as anchors to low-resource languages. SARA explicitly aligns the routing distribution of multilingual inputs with high-resource semantic anchors using a symmetric Jensen-Shannon (JS) divergence constraint. Unlike traditional distillation methods that operate on output logits, SARA directly aligns the internal routing distributions of MoE layers, encouraging mechanistic consistency in expert selection across languages. We conduct experiments on 2 LLMs across 5 low-resource languages and 3 benchmarks. Experiment results demonstrate that SARA outperforms standard instruction tuning, e.g., +0.8% on Qwen3-30B-A3B and +1.2% on Phi-3.5-MoE-instruct on Global-MMLU. Further analyses show that SARA effectively addresses performance bottlenecks in low-resource languages, providing a scalable pathway to enhance multilingual capabilities in sparse architectures.

cs.CL↗

arXiv2026

Mix-MoE: Improving Multilingual Machine Translation of Large Language Models through Mixed MoEs

Large Language Models (LLMs) have shown great promise in multilingual machine translation (MT), even with limited bilingual supervision. However, fine-tuning LLMs with parallel corpora presents major challenges, namely parameter interference. To address these issues, we propose Mix-MoE, a mixed Mixture-of-Experts framework designed to train LLMs for multilingual MT. Our framework operates in two distinct stages: (1) post-pretraining with MoE on monolingual corpora, and (2) post-pretraining with MoE on parallel corpora. Crucially, we divide the MoE layers into two specialized groups: Language Model Experts (LM Experts) and Machine Translation Experts (MT Experts). LM Experts are designed to capture and retain the monolingual knowledge learned by the pre-trained LLM. MT Experts, on the other hand, are specifically trained to acquire and store bilingual translation knowledge. Furthermore, to facilitate effective interaction between these specialized experts and leverage potential underlying structural patterns in text, we introduce a routing mechanism enhanced by Fourier Transform features derived from model representations. The experimental results demonstrate that Mix-MoE excels in multilingual MT, significantly outperforming existing baselines and showing notable progress in mitigating parameter interference.

cs.CL↗

arXiv2026

From Curated Data to Scalable Models: Continual Pre-training of Dense and MoE Large Language Models for Tibetan

Large language models (LLMs) have achieved remarkable success across a wide range of natural language processing tasks, yet their performance remains heavily biased toward high-resource languages. Tibetan, despite its cultural significance and large speaker population, is still substantially underrepresented. In this work, we present a comprehensive pipeline for advancing Tibetan language modeling through large-scale data curation and continual pre-training. We construct a 72 GB high-quality Tibetan corpus, the largest to date, and adapt Qwen2.5-7B through balanced multilingual continual pre-training with Tibetan, Chinese, and English, followed by multilingual instruction tuning. To further scale capacity efficiently, we extend the dense model to a 50B-A10B Mixture-of-Experts architecture. Due to the absence of standardized Tibetan benchmarks, we build multiple evaluation datasets via high-quality translation and human verification. Experimental results show that both dense and MoE models consistently outperform existing open-source and Tibetan-focused models of similar scale across diverse tasks. Our work advances Tibetan-centric LLM research and provides transferable insights for extending LLMs to other low-resource languages. We will release the model weights, evaluation benchmarks, and detailed data processing documentation in the follow-up.

cs.CL↗

arXiv2026

Incentivizing Parametric Knowledge via Reinforcement Learning with Verifiable Rewards for Cross-Cultural Entity Translation

Cross-cultural entity translation remains challenging for large language models (LLMs) as literal or phonetic renderings are usually yielded instead of culturally appropriate translations in context. However, relevant knowledge may already be encoded in model parameters during large-scale pre-training. To incentivize the effective use of parametric knowledge, we propose EA-RLVR (Entity-Anchored Reinforcement Learning with Verifiable Rewards), a training framework that optimizes cross-cultural entity translation without relying on external knowledge bases. EA-RLVR anchors supervision on a verifiable, entity-level reward signal and incorporates lightweight structural gates to stabilize optimization. This design steers the model toward learning a robust reasoning process rather than merely imitating reference translations. We evaluate EA-RLVR on XC-Translate and observe consistent improvements in both entity translation accuracy and out-of-domain generalization. Specifically, training on merely 7k samples boosts Qwen3-14B's entity translation accuracy from 23.66\% to 31.87\% on a 50k test set comprising entirely unseen entities. The learned entity translation ability also transfers to general translation, yielding +1.35 XCOMET on WMT24++, which scales to +1.59 with extended optimization. Extensive analyses of $pass@k$ dynamics and reward formulations attribute these gains to superior sampling efficiency and a stable optimization landscape.

cs.CL↗

arXiv2026

MNAFT: modality neuron-aware fine-tuning of multimodal large language models for image translation

Multimodal large language models (MLLMs) have shown impressive capabilities, yet they often struggle to effectively capture the fine-grained textual information within images crucial for accurate image translation. This often leads to a modality gap between visual text inputs and textual inputs/outputs for image translation. Existing methods, primarily relying on instruction fine-tuning, risk parameter redundancy of pre-trained knowledge, hindering generalization performance. To address this, we introduce modality neuron-aware fine-tuning (MNAFT), a novel approach that takes advantage of the specialized roles of individual neurons within MLLMs for enhanced image translation. MNAFT identifies language-agnostic and language-specific neurons in both vision and language modules through an instruction-driven activation analysis, evaluating their importance in various translation tasks. We then perform selective fine-tuning, updating only the parameters of language-specific and language-agnostic neurons within the selected layers relevant to the target task, while preserving the knowledge encoded in other neurons and layers. Our extensive experiments on multiple benchmarks demonstrate that MNAFT significantly outperforms state-of-the-art image translation methods, including cascaded models, standard full fine-tuning, and parameter-efficient tuning techniques. Furthermore, we provide comprehensive analysis, including visualizations of neuron activations and clustering patterns, to offer insights into the roles of different neuron groups in mediating cross-modal understanding and facilitating accurate language-specific translation.

cs.CL↗

arXiv2026

DEP: A Decentralized Large Language Model Evaluation Protocol

With the rapid development of Large Language Models (LLMs), a large number of benchmarks have been proposed. However, most benchmarks lack unified evaluation standard and require the manual implementation of custom scripts, making results hard to ensure consistency and reproducibility. Furthermore, mainstream evaluation frameworks are centralized, with datasets and answers, which increases the risk of benchmark leakage. To address these issues, we propose a Decentralized Evaluation Protocol (DEP), a decentralized yet unified and standardized evaluation framework through a matching server without constraining benchmarks. The server can be mounted locally or deployed remotely, and once adapted, it can be reused over the long term. By decoupling users, LLMs, and benchmarks, DEP enables modular, plug-and-play evaluation: benchmark files and evaluation logic stay exclusively on the server side. In remote setting, users cannot access the ground truth, thereby achieving data isolation and leak-proof evaluation. To facilitate practical adoption, we develop DEP Toolkit, a protocol-compatible toolkit that supports features such as breakpoint resume, concurrent requests, and congestion control. We also provide detailed documentation for adapting new benchmarks to DEP. Using DEP toolkit, we evaluate multiple LLMs across benchmarks. Experimental results verify the effectiveness of DEP and show that it reduces the cost of deploying benchmark evaluations. As of February 2026, we have adapted over 60 benchmarks and continue to promote community co-construction to support unified evaluation across various tasks and domains.

cs.CL↗

arXiv2025

Clock Pulling Enables Maximum-Efficiency Wireless Power Transfer

Nonlinear parity-time (PT) symmetry in non-Hermitian wireless power transfer (WPT) systems, while attracting significant attention from both physics and engineering communities, have posed formidable theoretical and practical challenges due to their complex dynamical mechanisms. Here, we revisit multistability in nonlinear non-Hermitian systems and find that the PT-symmetry state is not always stable even in PT-symmetry phase. We report a discovery on a nonlinear clock-pulling mechanism, which can forcibly break the PT symmetry. Proper implementation of this mechanism can switch the system stability, particularly in stabilizing the conventional unstable state which has the maximum transfer efficiency for WPT. Our work offers new tools for non-Hermitian physics and is expected to drive technological progress.

physics.app-ph↗

arXiv2025

FuxiMT: Sparsifying Large Language Models for Chinese-Centric Multilingual Machine Translation

In this paper, we present FuxiMT, a novel Chinese-centric multilingual machine translation model powered by a sparsified large language model (LLM). We adopt a two-stage strategy to train FuxiMT. We first pre-train the model on a massive Chinese corpus and then conduct multilingual fine-tuning on a large parallel dataset encompassing 65 languages. FuxiMT incorporates Mixture-of-Experts (MoEs) and employs a curriculum learning strategy for robust performance across various resource levels. Experimental results demonstrate that FuxiMT significantly outperforms strong baselines, including state-of-the-art LLMs and machine translation models, particularly under low-resource scenarios. Furthermore, FuxiMT exhibits remarkable zero-shot translation capabilities for unseen language pairs, indicating its potential to bridge communication gaps where parallel data are scarce or unavailable.

cs.CL↗

arXiv2025

Focal control and light tracing on curved surfaces with isotropic transformation medium

Optics related to non-Euclidean geometry has been attracting growing interest for emerged novel phenomena and the analog for general relativity, while most studies are limited to the free space on rotationally-symmetric surfaces. In this paper, we focus on the light control and ray tracing on complex surfaces filled with inhomogeneous transformation medium. Within the conformal transformation optics, focal control devices and absolute optical instruments have been extended to curved surfaces. According to the equivalence between geometry and material, the metric tensor of the curved surface and the refractive index tensor are unified as the optical metric for the Hamilton's equations of light propagation on a curved surface. By solving for ray trajectories in the local coordinate system of mesh element and illuminating the refraction between non-planar elements with discontinuous media, a mesh-based ray-tracing algorithm on curved surface with medium has been proposed to validate the light control. Our research establishes a theoretical framework for light ray control in non-Euclidean space and offers an efficient tool for ray tracing in inhomogeneous medium on curved surface.

physics.optics↗

arXiv2024

Topological invariant of non-Hermitian space-time modulated photonic crystals

We propose a medium transformation approach to formulate the adjoint system of space-time modulated photonic crystals (STMPCs), essential for the bi-orthogonal Berry connection when calculating the topological invariant. We show that the non-Abelian Zak phase of STMPCs comprising stacked photonic time crystals and dielectrics is quantized to 0 or 1 for both the entangled and isolated bands. We find that the eigenmodes at the center and edge of the Brillouin zone differ in symmetry for the band with non-trivial Zak phases, while they share the same symmetry for the trivial Zak phases. In addition, topological phase transitions owing to band inversion are observed. Moreover, a generalized Brillouin zone of the non-Hermitian STMPCs is established, which is identical to the Hermitian counterpart, implicating that the non-Bloch band theory is not required in this regard. The proposed medium transformation method may serve as an alternative approach to exploring more intricate topological phenomena in non-Hermitian systems when incorporating non-Bloch band theory.

physics.optics↗

arXiv2024

Three-dimensional Coupled PT-symmetric Electronic Resonators

In this article, the non-Hermitian characteristics of three-dimensional PT-symmetric coupled electronic resonators are theoretically analyzed. First, the concept of non-Hermitian PT symmetry is illustrated in the context of electronics using a pair of coupled electronic resonators. Two typical configurations of parallel-coupled PT-symmetric electronic trimers are then analyzed. The results indicate that, for the planar configuration, the system can exhibit two phase transitions as the coupling coefficient or gain-loss parameter changes, different from the linear configuration. By comparing system equations based on coupled-mode theory and circuit theory, it is shown that high dimensionality alone is not a sufficient condition for the existence of a higher-order exceptional point; an approximation condition is also required. A modified exceptional point is proposed, and the approximation conditions for the mean deviation $D$ for the real part of the three eigenfrequencies, satisfying $D \leq 1\%$ and $D \leq 0.1\%$, are discussed, respectively. The theoretical results presented in this paper not only reveal the unique non-Hermitian characteristics of high-dimensional PT-symmetric electronic systems but also offer theoretical support for wireless power transmission and wireless sensing technologies.

physics.app-ph↗

arXiv2024

Communication-Free Robust Wireless Power Transfer with Constant Output Power and Stable Frequency

A primary challenge in wireless power transfer (WPT) systems is to achieve efficient and stable power transmission without complex control strategies when load conditions change dynamically. Addressing this issue, we propose a third-order pseudo-Hermitian WPT system whose output characteristics exhibit a stable frequency and constant power. The frequency selection mechanism and energy efficiency of the nonlinear WPT system based on pseudo-Hermitian under the coupling mode theory approximation are analyzed. Theoretical analysis indicates that under certain coupling coefficients and load conditions, the proposed system can achieve frequency adaptation in a stable frequency mode without the need to change the circuit frequency. When the load changes dynamically, the stability of the power output is maintained using a proportional integral (PI) control strategy that only collects the voltage and current at the transmitting end, eliminating the need for wireless communication circuits with feedback from the receiving side. Experimental results demonstrate that the proposed design scheme can achieve constant power transmission when load conditions change, maintaining stable and relatively high transmission efficiency. The proposed scheme exhibits benefits in practical applications since no communication is required.

physics.app-ph↗

arXiv2024

Dispersive gains enhance wireless power transfer with asymmetric resonance

Parity-time symmetry is a fundamental concept in non-Hermitian physics that has recently gained attention for its potential in engineering advanced electronic systems and achieving robust wireless power transfer even in the presence of disturbances, through the incorporation of nonlinearity. However, the current parity-time-symmetric scheme falls short of achieving the theoretical maximum efficiency of wireless power transfer and faces challenges when applied to non-resistive loads. In this study, we propose a theoretical framework and provide experimental evidence demonstrating that asymmetric resonance, based on dispersive gain, can greatly enhance the efficiency of wireless power transfer beyond the limits of symmetric approaches. By leveraging the gain spectrum interleaving resulting from dispersion, we observe a mode switching phenomenon in asymmetric systems similar to the symmetry-breaking effect. This phenomenon reshapes the distribution of resonance energy and enables more efficient wireless power transfer compared to conventional methods. Our findings open up new possibilities for harnessing dispersion effects in various domains such as electronics, microwaves, and optics. This work represents a significant step towards exploiting dispersion as a means to optimize wireless power transfer and lays the foundation for future advancements in these fields.

physics.app-ph↗

arXiv2023

Free electron emission in vacuum assisted by photonic time crystals

The Cerenkov radiation and the Smith-Purcell effect state that free electron emission occurs exclusively in dielectrics when the velocity of the particles exceeds the speed of light in the medium or in the vicinity of periodic gratings close to each other within a vacuum. We demonstrate that free electrons in a vacuum can also emit highly directional monochromatic waves when they are in close proximity to a medium that is periodically modulated temporally, suggesting the existence of temporal Smith-Purcell effect. The momentum band gaps of time-varying media, such as photonic time crystals (PTCs), create new pathways for the injection of external energy, allowing the frequency, intensity, and spatial distribution of the electromagnetic fields to be controlled. Moreover, the PTC substrate enables the conversion of localized evanescent fields into amplified, highly directional propagating plane waves that are only sensitive to the velocity of particles and the modulation frequency, which allows us to observe and utilize Cerenkov-like radiation in free space. Our work exhibits significant opportunities for the utilization of time-varying structures in various fields, including particle identification, ultraweak signal detection, and improved radiation source design.

physics.optics↗

arXiv2023

M3KE: A Massive Multi-Level Multi-Subject Knowledge Evaluation Benchmark for Chinese Large Language Models

Large language models have recently made tremendous progress in a variety of aspects, e.g., cross-task generalization, instruction following. Comprehensively evaluating the capability of large language models in multiple tasks is of great importance. In this paper, we propose M3KE, a Massive Multi-Level Multi-Subject Knowledge Evaluation benchmark, which is developed to measure knowledge acquired by Chinese large language models by testing their multitask accuracy in zero- and few-shot settings. We have collected 20,477 questions from 71 tasks. Our selection covers all major levels of Chinese education system, ranging from the primary school to college, as well as a wide variety of subjects, including humanities, history, politics, law, education, psychology, science, technology, art and religion. All questions are multiple-choice questions with four options, hence guaranteeing a standardized and unified assessment process. We've assessed a number of state-of-the-art open-source Chinese large language models on the proposed benchmark. The size of these models varies from 335M to 130B parameters. Experiment results demonstrate that they perform significantly worse than GPT-3.5 that reaches an accuracy of ~ 48% on M3KE. The dataset is available at https://github.com/tjunlp-lab/M3KE.

cs.CL↗

arXiv2023

Observation of higher-order exceptional points in pseudo-Hermitian radio-frequency circuits

Exceptional points (EP) in non-Hermitian systems have been widely investigated due to their enhanced sensitivity in comparison to standard systems. In this letter, we report the observation of higher-order pseudo-Hermitian degeneracies in an electronic platform comprised of three inductively coupled gain-loss-loss LC resonators. Theoretical analysis demonstrates that the proposed system can realize third-order EP with asymmetric coupling between adjacent inductors and an arbitrary scaling factor between two loss resonators. When capacitive perturbation is introduced on the middle resonator, the perturbed eigenfrequencies follow a cube-root dependence on the perturbation parameter; in this case, the sensitivity is significantly greater than conventional wireless readout methods. Our work enriches the explorations of higher-order EP on electronic platforms and provides a new degree of design freedom for the non-Hermitian-EP-enhanced wireless sensing system.

physics.app-ph↗