arXiv subjects

Yunhao Zhang

Publications and source records attributed to Yunhao Zhang.

At least 19 recordsLinked to original sources

arXiv2026

VikingRAG: Accurate and Token-efficient Retrieval-augmented Generation over Structured Documents

State-of-the-art retrieval-augmented generation (RAG) methods exploit document structures to acquire sufficient evidence, but often incur substantial token costs. To reduce structural-context tokens without compromising high RAG accuracy, we present {\sf VikingRAG}, a directory-aware semantic data management system that tightly integrates semantic and structural access to support structural-context-efficient, evidence-gap-driven multi-round retrieval. To further reduce token overhead of multi-round interaction, we materialize agentic multi-round retrieval traces as experience edges, and reuse these edges for similar queries, avoiding repeated multi-round exploration. To additionally reduce token costs when agentic multi-round retrieval is unnecessary, we introduce an adaptive escalation strategy that answers from one-round experience-augmented retrieval when the evidence is sufficient, and invokes agentic multi-round retrieval only otherwise. Experiments on real datasets show that the base system {\sf VikingRAG} matches high accuracy of state-of-the-art methods while consuming only 11.6\%--51.9\% of their tokens. With retrieval-trace reuse and adaptive escalation, token costs drop to 5.1\%--32.5\% while maintaining competitive accuracy and practical document-storage performance, showing the utility of this work for emerging AI knowledge bases.

cs.IR↗

arXiv2026

Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction

We introduce Tencent WorkBuddy Bench, a multi-domain evaluation suite for coding agents; this report documents its construction methodology, scoring protocol, and a cross-model leaderboard. At its core is a unified evaluation framework for constructing and running distribution-informed coding-agent tasks across four work domains - Code, Web, Office, and Security. Rather than adapting public issue text, every task is reverse-engineered from a real commit, pull request, or business scenario and rewritten as a short, colloquial, role-played request, so that a task's prompt is not recoverable by web-searching the underlying issue, pull request, or commit thread. Because the dataset is released openly - task directories, environment images, evaluation harness, tests, and reference solutions - contamination resistance rests on this construction together with dataset versioning rather than on secrecy. The four subsets - repository-level engineering, front-end development, office and business workflows, and red-/blue-team security - probe complementary facets of real work, each with its own verification style. All are packaged in a uniform task-directory format and run, under a uniform and reproducible protocol, on two agent harnesses (CodeBuddy Code and Claude Code); the full open release makes the benchmark reproducible end to end and directly auditable, since any third party can re-run each task and inspect its content. Because each subset uses a different scoring instrument, scores are not comparable across subsets and the suite reports no suite-wide average. We report a cross-model leaderboard across several model families.

cs.CL↗

arXiv2026

The First ChineseBabyLM Challenge: training data-efficient and cognitively plausible language models for Chinese

This paper presents the first ChineseBabyLM Challenge, organized as part of NLPCC 2026. The challenge asked participants to train language models from scratch using no more than 102M Chinese words. The models were evaluated on three tracks: natural language understanding, cognitive alignment, and Hanzi knowledge. There were no restrictions on tokenizers, model architectures, or the number of training epochs. Eighteen teams submitted 28 distinct models, generating 74 result files. The overall-winning team used a DeBERTa-v2 architecture and introduced an auxiliary pinyin-prediction objective during pretraining. Several submissions also explored curriculum-learning strategies and architectural innovations. Overall, the challenge provides a benchmark for advancing data-efficient and cognitively plausible approaches to Chinese language modeling.

cs.CL↗

arXiv2026

Time Series as Language: A Universal Tokenizer for General-Purpose Time Series Foundation Models

While Next-Token Prediction (NTP) has unified LLM pretraining, its adaptation to unbounded, continuous time series (TS) remains open. To bridge the gap, we introduce UniTok, a universal tokenizer that transforms TS into discrete tokens, and UniTok-FM, a foundation model pretrained via NTP on these tokens. UniTok-FM is a general-purpose foundation model that supports zero-shot and prompt-boosted forecasting, as well as few-shot generation and classification via training-free in-context inference--a capability not achieved by prior works. Technically, UniTok is a vector-quantized autoencoder incorporating prefix normalization for scale stabilization, a progressive-resolution causal architecture for encoding and decoding, and a structure-preserving reconstruction loss for training. UniTok-FM adopts an off-the-shelf LLM architecture without TS-specific modifications. Instead of pretraining on isolated TS, it performs NTP on context windows formed by multiple series with similar patterns, aiming to capture their shared dynamics. Experiments on forecasting, generation, and classification show that a single unified UniTok-FM consistently outperforms statistical and supervised baselines, achieves competitive performance with task-specific foundation models, and uniquely enables training-free in-context inference across tasks.

cs.LG↗

arXiv2026

Computational Lesions in Multilingual Language Models Separate Shared and Language-specific Brain Alignment

How the brain supports language across different languages is a basic question in neuroscience and a useful test for multilingual artificial intelligence. Neuroimaging has identified language-responsive brain regions across languages, but it cannot by itself show whether the underlying processing is shared or language-specific. Here we use six multilingual large language models (LLMs) as controllable systems and create targeted ``computational lesions'' by zeroing small parameter sets that are important across languages or especially important for one language. We then compare intact and lesioned models in predicting functional magnetic resonance imaging (fMRI) responses during 100 minutes of naturalistic story listening in native English, Chinese and French (112 participants). Lesioning a compact shared core reduces whole-brain encoding correlation by 60.32% relative to intact models, whereas language-specific lesions preserve cross-language separation in embedding space but selectively weaken brain predictivity for the matched native language. These results support a shared backbone with embedded specializations and provide a causal framework for studying multilingual brain-model alignment.

cs.CL↗

arXiv2026

Component-Level Lesioning of Language Models Reveals Clinically Aligned Aphasia Phenotypes

Large language models (LLMs) increasingly exhibit human-like linguistic behaviors and internal representations that they could serve as computational simulators of language cognition. We ask whether LLMs can be systematically manipulated to reproduce language-production impairments characteristic of aphasia following focal brain lesions. Such models could provide scalable proxies for testing rehabilitation hypotheses, and offer a controlled framework for probing the functional organization of language. We introduce a clinically grounded, component-level framework that simulates aphasia by selectively perturbing functional components in LLMs, and apply it to both modular Mixture-of-Experts models and dense Transformers using a unified intervention interface. Our pipeline (i) identifies subtype-linked components for Broca's and Wernicke's aphasia, (ii) interprets these components with linguistic probing tasks, and (iii) induces graded impairments by progressively perturbing the top-k subtype-linked components, evaluating outcomes with Western Aphasia Battery (WAB) subtests summarized by Aphasia Quotient (AQ). Across architectures and lesioning strategies, subtype-targeted perturbations yield more systematic, aphasia-like regressions than size-matched random perturbations, and MoE modularity supports more localized and interpretable phenotype-to-component mappings. These findings suggest that modular LLMs, combined with clinically informed component perturbations, provide a promising platform for simulating aphasic language production and studying how distinct language functions degrade under targeted disruptions.

cs.CL↗

arXiv2025

BubbleOKAN: A Physics-Informed Interpretable Neural Operator for High-Frequency Bubble Dynamics

In this work, we employ physics-informed neural operators to map pressure profiles from an input function space to the corresponding bubble radius responses. Our approach employs a two-step DeepONet architecture. To address the intrinsic spectral bias of deep learning models, our model incorporates the Rowdy adaptive activation function, enhancing the representation of high-frequency features. Moreover, we introduce the Kolmogorov-Arnold network (KAN) based two-step DeepOKAN model, which enhances interpretability (often lacking in conventional multilayer perceptron architectures) while efficiently capturing high-frequency bubble dynamics without explicit utilization of activation functions in any form. We particularly investigate the use of spline basis functions in combination with radial basis functions (RBF) within our architecture, as they demonstrate superior performance in constructing a universal basis for approximating high-frequency bubble dynamics compared to alternative formulations. Furthermore, we emphasize on the performance bottleneck of RBF while learning the high frequency bubble dynamics and showcase the advantage of using spline basis function for the trunk network in overcoming this inherent spectral bias. The model is systematically evaluated across three representative scenarios: (1) bubble dynamics governed by the Rayleigh-Plesset equation with a single initial radius, (2) bubble dynamics governed by the Keller-Miksis equation with a single initial radius, and (3) Keller-Miksis dynamics with multiple initial radii. We also compare our results with state-of-the-art neural operators, including Fourier Neural Operators, Wavelet Neural Operators, OFormer, and Convolutional Neural Operators. Our findings demonstrate that the two-step DeepOKAN accurately captures both low- and high-frequency behaviors, and offers a promising alternative to conventional numerical solvers.

cs.LG↗

arXiv2025

Discovering Semantic Subdimensions through Disentangled Conceptual Representations

Understanding the core dimensions of conceptual semantics is fundamental to uncovering how meaning is organized in language and the brain. Existing approaches often rely on predefined semantic dimensions that offer only broad representations, overlooking finer conceptual distinctions. This paper proposes a novel framework to investigate the subdimensions underlying coarse-grained semantic dimensions. Specifically, we introduce a Disentangled Continuous Semantic Representation Model (DCSRM) that decomposes word embeddings from large language models into multiple sub-embeddings, each encoding specific semantic information. Using these sub-embeddings, we identify a set of interpretable semantic subdimensions. To assess their neural plausibility, we apply voxel-wise encoding models to map these subdimensions to brain activation. Our work offers more fine-grained interpretable semantic subdimensions of conceptual meaning. Further analyses reveal that semantic dimensions are structured according to distinct principles, with polarity emerging as a key factor driving their decomposition into subdimensions. The neural correlates of the identified subdimensions support their cognitive and neuroscientific plausibility.

cs.CL↗

arXiv2025

Ordered Consensus with Equal Opportunity

The specification of state machine replication (SMR) has no requirement on the final total order of commands. In blockchains based on SMR, however, order matters, since different orders could provide their clients with different financial rewards. Ordered consensus augments the specification of SMR to include specific guarantees on such order, with a focus on limiting the influence of Byzantine nodes. Real-world ordering manipulations, however, can and do happen even without Byzantine replicas, typically because of factors, such as faster networks or closer proximity to the blockchain infrastructure, that give some clients an unfair advantage. To address this challenge, this paper proceeds to extend ordered consensus by requiring it to also support equal opportunity, a concrete notion of fairness, widely adopted in social sciences. Informally, equal opportunity requires that two candidates who, according to a set of criteria deemed to be relevant, are equally qualified for a position (in our case, a specific slot in the SMR total order), should have an equal chance of landing it. We show how randomness can be leveraged to keep bias in check, and, to this end, introduce the secret random oracle (SRO), a system component that generates randomness in a fault-tolerant manner. We describe two SRO designs based, respectively, on trusted hardware and threshold verifiable random functions, and instantiate them in Bercow, a new ordered consensus protocol that, by approximating equal opportunity up to within a configurable factor, can effectively mitigate well-known ordering attacks in SMR-based blockchains.

cs.DC↗

arXiv2025

Bridging Brains and Models: MoE-Based Functional Lesions for Simulating and Rehabilitating Aphasia

The striking alignment between large language models (LLMs) and human brain activity positions them as powerful models of healthy cognition. This parallel raises a fundamental question: if LLMs can model the intact brain, can we lesion them to simulate the linguistic deficits of the injured brain? In this work, we introduce a methodology to model aphasia - a complex language disorder caused by neural injury - by selectively disabling components in a modular Mixture-of-Experts (MoE) language model. We simulate distinct aphasia subtypes, validate their linguistic outputs against real patient speech, and then investigate functional recovery by retraining the model's remaining healthy experts. Our results demonstrate that lesioning functionally-specialized experts for syntax or semantics induces distinct impairments that closely resemble Broca's and Wernicke's aphasia, respectively. Crucially, we show that freezing the damaged experts and retraining the intact ones on conversational data restores significant linguistic function, demonstrating a computational analogue for rehabilitation. These findings establish modular LLMs as a powerful and clinically-relevant potential framework for modeling the mechanisms of language disorders and for computationally exploring novel pathways for therapy.

q-bio.NC↗

arXiv2025

How Syntax Specialization Emerges in Language Models

Large language models (LLMs) have been found to develop surprising internal specializations: Individual neurons, attention heads, and circuits become selectively sensitive to syntactic structure, reflecting patterns observed in the human brain. While this specialization is well-documented, how it emerges during training and what influences its development remains largely unknown. In this work, we tap into the black box of specialization by tracking its formation over time. By quantifying internal syntactic consistency across minimal pairs from various syntactic phenomena, we identify a clear developmental trajectory: Syntactic sensitivity emerges gradually, concentrates in specific layers, and exhibits a 'critical period' of rapid internal specialization. This process is consistent across architectures and initialization parameters (e.g., random seeds), and is influenced by model scale and training data. We therefore reveal not only where syntax arises in LLMs but also how some models internalize it during training. To support future research, we will release the code, models, and training checkpoints upon acceptance.

cs.CL↗

arXiv2025

Decoding the Multimodal Mind: Generalizable Brain-to-Text Translation via Multimodal Alignment and Adaptive Routing

Decoding language from the human brain remains a grand challenge for Brain-Computer Interfaces (BCIs). Current approaches typically rely on unimodal brain representations, neglecting the brain's inherently multimodal processing. Inspired by the brain's associative mechanisms, where viewing an image can evoke related sounds and linguistic representations, we propose a unified framework that leverages Multimodal Large Language Models (MLLMs) to align brain signals with a shared semantic space encompassing text, images, and audio. A router module dynamically selects and fuses modality-specific brain features according to the characteristics of each stimulus. Experiments on various fMRI datasets with textual, visual, and auditory stimuli demonstrate state-of-the-art performance, achieving an 8.48% improvement on the most commonly used benchmark. We further extend our framework to EEG and MEG data, demonstrating flexibility and robustness across varying temporal and spatial resolutions. To our knowledge, this is the first unified BCI architecture capable of robustly decoding multimodal brain activity across diverse brain signals and stimulus types, offering a flexible solution for real-world applications.

cs.CL↗

arXiv2025

SparseVoxFormer: Sparse Voxel-based Transformer for Multi-modal 3D Object Detection

Most previous 3D object detection methods that leverage the multi-modality of LiDAR and cameras utilize the Bird's Eye View (BEV) space for intermediate feature representation. However, this space uses a low x, y-resolution and sacrifices z-axis information to reduce the overall feature resolution, which may result in declined accuracy. To tackle the problem of using low-resolution features, this paper focuses on the sparse nature of LiDAR point cloud data. From our observation, the number of occupied cells in the 3D voxels constructed from a LiDAR data can be even fewer than the number of total cells in the BEV map, despite the voxels' significantly higher resolution. Based on this, we introduce a novel sparse voxel-based transformer network for 3D object detection, dubbed as SparseVoxFormer. Instead of performing BEV feature extraction, we directly leverage sparse voxel features as the input for a transformer-based detector. Moreover, with regard to the camera modality, we introduce an explicit modality fusion approach that involves projecting 3D voxel coordinates onto 2D images and collecting the corresponding image features. Thanks to these components, our approach can leverage geometrically richer multi-modal features while even reducing the computational cost. Beyond the proof-of-concept level, we further focus on facilitating better multi-modal fusion and flexible control over the number of sparse features. Finally, thorough experimental results demonstrate that utilizing a significantly smaller number of sparse features drastically reduces computational costs in a 3D object detector while enhancing both overall and long-range performance.

cs.CV↗

arXiv2025

Exclusive Generation of Single-Atom Sulfur for Ultrahigh Quality Monolayer MoS$_2$ Growth

Preparation of high-quality two-dimensional (2D) transition metal dichalcogenides (TMDCs) is the precondition for realizing their applications. However, the synthesized 2D TMDCs (e.g., MoS$_2$) crystals suffer from low quality due to the massive defects formed during the growth. Here, we report the single-atom sulfur (S1) as a highly reactive sulfur species to grow ultrahigh-quality monolayer MoS$_2$. Derived from battery waste, the sulfurized polyacrylonitrile (SPAN) is found to be exclusive and efficient in releasing S1. The monolayer MoS$_2$ prepared by SPAN exhibits an ultralow defect density of $~7\times 10^{12}$ cm$^{-2}$ and the narrowest photoluminescence (PL) emission peak with full-width at half-maximum of ~47.11 meV at room temperature. Moreover, the statistical resonance Raman and low-temperature PL results further verify the significantly lower defect density and higher optical quality of SPAN-grown MoS$_2$ than the conventional S-powder-grown samples. This work provides an effective approach for preparing ultrahigh-quality 2D single crystals, facilitating their industrial applications.

cond-mat.mtrl-sci↗

arXiv2025

Cu Intercalation-stabilized 1T'-MoS2 with Electrical Insulating Behavior

The intercalated two-dimensional (2D) transition metal dichalcogenides (TMDCs) have attracted much attention for their designable structure and novel properties. Among this family, host materials with low symmetry such as 1T' phase TMDCs are particularly interesting because of their potentials in inducing unconventional phenomena. However, such systems typically have low quality and poor stability, hindering further study in the structure-property relationship and applications. In this work, we intercalated Cu into 1T' MoS2 with high crystallinity and high thermal stability up to ~300 oC. We identified the distribution and arrangement of Cu intercalators for the first time, and the results show that Cu occupy partial of the tetrahedral interstices aligned with Mo sites. The obtained Cu-1T' MoS2 exhibits an insulating hopping transport behavior with a large temperature coefficient of resistance reaching -4 ~ -2 % K-1. This work broadens the artificial intercalated structure library and promotes structure design and property modulation of layered materials.

cond-mat.mtrl-sci↗

arXiv2024

LinSATNet: The Positive Linear Satisfiability Neural Networks

Encoding constraints into neural networks is attractive. This paper studies how to introduce the popular positive linear satisfiability to neural networks. We propose the first differentiable satisfiability layer based on an extension of the classic Sinkhorn algorithm for jointly encoding multiple sets of marginal distributions. We further theoretically characterize the convergence property of the Sinkhorn algorithm for multiple marginals. In contrast to the sequential decision e.g.\ reinforcement learning-based solvers, we showcase our technique in solving constrained (specifically satisfiability) problems by one-shot neural networks, including i) a neural routing solver learned without supervision of optimal solutions; ii) a partial graph matching network handling graphs with unmatchable outliers on both sides; iii) a predictive network for financial portfolios with continuous constraints. To our knowledge, there exists no one-shot neural solver for these scenarios when they are formulated as satisfiability problems. Source code is available at https://github.com/Thinklab-SJTU/LinSATNet

cs.AI↗

arXiv2024

Technical Report for CVPR 2024 WeatherProof Dataset Challenge: Semantic Segmentation on Paired Real Data

This technical report presents the implementation details of 2nd winning for CVPR'24 UG2 WeatherProof Dataset Challenge. This challenge aims at semantic segmentation of images degraded by various degrees of weather from all around the world. We addressed this problem by introducing a pre-trained large-scale vision foundation model: InternImage, and trained it using images with different levels of noise. Besides, we did not use additional datasets in the training procedure and utilized dense-CRF as post-processing in the final testing procedure. As a result, we achieved 2nd place in the challenge with 45.1 mIOU and fewer submissions than the other winners.

cs.CV↗

arXiv2024

Navigating Brain Language Representations: A Comparative Analysis of Neural Language Models and Psychologically Plausible Models

Neural language models, particularly large-scale ones, have been consistently proven to be most effective in predicting brain neural activity across a range of studies. However, previous research overlooked the comparison of these models with psychologically plausible ones. Moreover, evaluations were reliant on limited, single-modality, and English cognitive datasets. To address these questions, we conducted an analysis comparing encoding performance of various neural language models and psychologically plausible models. Our study utilized extensive multi-modal cognitive datasets, examining bilingual word and discourse levels. Surprisingly, our findings revealed that psychologically plausible models outperformed neural language models across diverse contexts, encompassing different modalities such as fMRI and eye-tracking, and spanning languages from English to Chinese. Among psychologically plausible models, the one incorporating embodied information emerged as particularly exceptional. This model demonstrated superior performance at both word and discourse levels, exhibiting robust prediction of brain activation across numerous regions in both English and Chinese.

cs.CL↗