arXiv subjects

Yu Zhao

Publications and source records attributed to Yu Zhao.

At least 145 records · Page 8Linked to original sources

arXiv2025

Structured Packing in LLM Training Improves Long Context Utilization

Recent advancements in long-context large language models have attracted significant attention, yet their practical applications often suffer from suboptimal context utilization. This study investigates structuring training data to enhance semantic interdependence, demonstrating that this approach effectively improves context utilization. To this end, we introduce the Structured Packing for Long Context (SPLiCe) method, which utilizes retrieval to collate mutually relevant documents into long and coherent training examples. We validate SPLiCe empirically across models of varying sizes -- 3B, 7B, and 13B -- achieving improved performance in long-context tasks, such as Qasper and HotpotQA. Remarkably, even brief fine-tuning with SPLiCe is sufficient to realize these benefits. Additionally, SPLiCe effectively mitigates the lost-in-middle phenomenon often observed in large models. Our comprehensive analysis of SPLiCe explores its design choices and reveals intriguing transfer effects; for instance, training on programming code enhances performance on natural language tasks.

cs.CL↗

arXiv2025

Multimodal Graph-Based Variational Mixture of Experts Network for Zero-Shot Multimodal Information Extraction

Multimodal information extraction on social media is a series of fundamental tasks to construct the multimodal knowledge graph. The tasks aim to extract the structural information in free texts with the incorporate images, including: multimodal named entity typing and multimodal relation extraction. However, the growing number of multimodal data implies a growing category set and the newly emerged entity types or relations should be recognized without additional training. To address the aforementioned challenges, we focus on the zero-shot multimodal information extraction tasks which require using textual and visual modalities for recognizing unseen categories. Compared with text-based zero-shot information extraction models, the existing multimodal ones make the textual and visual modalities aligned directly and exploit various fusion strategies to improve their performances. But the existing methods ignore the fine-grained semantic correlation of text-image pairs and samples. Therefore, we propose the multimodal graph-based variational mixture of experts network (MG-VMoE) which takes the MoE network as the backbone and exploits it for aligning multimodal representations in a fine-grained way. Considering to learn informative representations of multimodal data, we design each expert network as a variational information bottleneck to process two modalities in a uni-backbone. Moreover, we also propose the multimodal graph-based virtual adversarial training to learn the semantic correlation between the samples. The experimental results on the two benchmark datasets demonstrate the superiority of MG-VMoE over the baselines.

cs.MM↗

arXiv2025

Grammar Induction from Visual, Speech and Text

Grammar Induction could benefit from rich heterogeneous signals, such as text, vision, and acoustics. In the process, features from distinct modalities essentially serve complementary roles to each other. With such intuition, this work introduces a novel \emph{unsupervised visual-audio-text grammar induction} task (named \textbf{VAT-GI}), to induce the constituent grammar trees from parallel images, text, and speech inputs. Inspired by the fact that language grammar natively exists beyond the texts, we argue that the text has not to be the predominant modality in grammar induction. Thus we further introduce a \emph{textless} setting of VAT-GI, wherein the task solely relies on visual and auditory inputs. To approach the task, we propose a visual-audio-text inside-outside recursive autoencoder (\textbf{VaTiora}) framework, which leverages rich modal-specific and complementary features for effective grammar parsing. Besides, a more challenging benchmark data is constructed to assess the generalization ability of VAT-GI system. Experiments on two benchmark datasets demonstrate that our proposed VaTiora system is more effective in incorporating the various multimodal signals, and also presents new state-of-the-art performance of VAT-GI.

cs.CL↗

arXiv2025

On the peak height distribution of non-stationary Gaussian random fields: 1D general covariance and scale space

We study the peak height distribution of certain non-stationary Gaussian random fields. The explicit peak height distribution of smooth, non-stationary Gaussian processes in 1D with general covariance is derived. The formula is determined by two parameters, each of which has a clear statistical meaning. For multidimensional non-stationary Gaussian random fields, we generalize these results to the setting of scale space fields, which play an important role in peak detection by helping to handle peaks of different spatial extents. We demonstrate that these properties not only offer a better interpretation of the scale space field but also simplify the computation of the peak height distribution. Finally, two efficient numerical algorithms are proposed as a general solution for computing the peak height distribution of smooth multidimensional Gaussian random fields in applications.

stat.ME↗

arXiv2025

Utilizing Effective Dynamic Graph Learning to Shield Financial Stability from Risk Propagation

Financial risks can propagate across both tightly coupled temporal and spatial dimensions, posing significant threats to financial stability. Moreover, risks embedded in unlabeled data are often difficult to detect. To address these challenges, we introduce GraphShield, a novel approach with three key innovations: Enhanced Cross-Domain Infor mation Learning: We propose a dynamic graph learning module to improve information learning across temporal and spatial domains. Advanced Risk Recognition: By leveraging the clustering characteristics of risks, we construct a risk recognizing module to enhance the identification of hidden threats. Risk Propagation Visualization: We provide a visualization tool for quantifying and validating nodes that trigger widespread cascading risks. Extensive experiments on two real-world and two open-source datasets demonstrate the robust performance of our framework. Our approach represents a significant advancement in leveraging artificial intelligence to enhance financial stability, offering a powerful solution to mitigate the spread of risks within financial networks.

q-fin.RM↗

arXiv2025

Steering Knowledge Selection Behaviours in LLMs via SAE-Based Representation Engineering

Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context -- this phenomenon, known as \emph{context-memory knowledge conflicts}, can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. Analysing the internal activations of LLMs, we find that they can internally register the signals of knowledge conflict at mid-layers. Such signals allow us to detect whether a knowledge conflict occurs and use \emph{inference-time} intervention strategies to resolve it. In this work, we propose \textsc{SpARE}, a \emph{training-free} representation engineering method that uses pre-trained sparse auto-encoders (SAEs) to control the knowledge selection behaviour of LLMs. \textsc{SpARE} identifies the functional features that control the knowledge selection behaviours and applies them to edit the internal activations of LLMs at inference time. Our experimental results show that \textsc{SpARE} can effectively control the usage of either knowledge source to resolve knowledge conflict in open-domain question-answering tasks, surpassing existing representation engineering methods ($+10\%$) as well as contrastive decoding methods ($+15\%$).

cs.CL↗

arXiv2025

Analysing the Residual Stream of Language Models Under Knowledge Conflicts

Large language models (LLMs) can store a significant amount of factual knowledge in their parameters. However, their parametric knowledge may conflict with the information provided in the context. Such conflicts can lead to undesirable model behaviour, such as reliance on outdated or incorrect information. In this work, we investigate whether LLMs can identify knowledge conflicts and whether it is possible to know which source of knowledge the model will rely on by analysing the residual stream of the LLM. Through probing tasks, we find that LLMs can internally register the signal of knowledge conflict in the residual stream, which can be accurately detected by probing the intermediate model activations. This allows us to detect conflicts within the residual stream before generating the answers without modifying the input or model parameters. Moreover, we find that the residual stream shows significantly different patterns when the model relies on contextual knowledge versus parametric knowledge to resolve conflicts. This pattern can be employed to estimate the behaviour of LLMs when conflict happens and prevent unexpected answers before producing the answers. Our analysis offers insights into how LLMs internally manage knowledge conflicts and provides a foundation for developing methods to control the knowledge selection processes.

cs.CL↗

arXiv2025

A New Interpretation for the Hot Corona in Active Galactic Nuclei

This work attempts to provide a new interpretation for the hot corona in active galactic nuclei (AGNs). A thin parabolic magnetic reconnection layer, anchored at the innermost disk and extending along the boundary of the magnetic tower for a few tens of gravitational radii, serves as a hard-X-ray source above the disk. Within this reconnection layer, the tearing instability leads to the formation of a chain of plasmoids, which contain relativistic electrons that generate X-ray radiation through inverse-Compton (IC) scattering of soft photons emitted by the accretion disk. Based on previous theoretical works and numerical simulations, we develop a heuristic framework to parameterize the geometry and magnetization of the reconnection layer, as well as to compute both the power of the IC-scattering radiation and the height of the reconnection layer. Our model allows for a quantitative investigation of the relation between the height of the corona and the X-ray radiation luminosity, which can be directly compared against the observed relation from X-ray reverberation mapping of individual AGNs. The theoretical results are in good agreement with the observations of IRAS 13224-3809, indicating the validation of our model.

astro-ph.HE↗

arXiv2025

Are We Done with MMLU?

Maybe not. We identify and analyse errors in the popular Massive Multitask Language Understanding (MMLU) benchmark. Even though MMLU is widely adopted, our analysis demonstrates numerous ground truth errors that obscure the true capabilities of LLMs. For example, we find that 57% of the analysed questions in the Virology subset contain errors. To address this issue, we introduce a comprehensive framework for identifying dataset errors using a novel error annotation protocol. Then, we create MMLU-Redux, which is a subset of 5,700 manually re-annotated questions across all 57 MMLU subjects. We estimate that 6.49% of MMLU questions contain errors. Using MMLU-Redux, we demonstrate significant discrepancies with the model performance metrics that were originally reported. Our results strongly advocate for revising MMLU's error-ridden questions to enhance its future utility and reliability as a benchmark. https://huggingface.co/datasets/edinburgh-dawg/mmlu-redux-2.0.

cs.CL↗

arXiv2024

Melody-Guided Music Generation

We present the Melody-Guided Music Generation (MG2) model, a novel approach using melody to guide the text-to-music generation that, despite a simple method and limited resources, achieves excellent performance. Specifically, we first align the text with audio waveforms and their associated melodies using the newly proposed Contrastive Language-Music Pretraining, enabling the learned text representation fused with implicit melody information. Subsequently, we condition the retrieval-augmented diffusion module on both text prompt and retrieved melody. This allows MG2 to generate music that reflects the content of the given text description, meantime keeping the intrinsic harmony under the guidance of explicit melody information. We conducted extensive experiments on two public datasets: MusicCaps and MusicBench. Surprisingly, the experimental results demonstrate that the proposed MG2 model surpasses current open-source text-to-music generation models, achieving this with fewer than 1/3 of the parameters or less than 1/200 of the training data compared to state-of-the-art counterparts. Furthermore, we conducted comprehensive human evaluations involving three types of users and five perspectives, using newly designed questionnaires to explore the potential real-world applications of MG2.

cs.SD↗

arXiv2024

Bulge Oscillation Driven by Outflows of Active Galactic Nuclei. I. Fast Outflow Case

There is growing evidence for star formation inside outflows of active galactic nuclei (AGNs). The formed stars are injected into bulges and give rise to perturbation of bulges. In this paper, we investigate the issues of non-rotating, spherically symmetric bulges under the perturbation of fast, massive outflows with stars formed inside. We show that the potential perturbation of outflows together with injection and dynamical friction of these stars could drive bulge oscillations. Still, we find non-zero radial velocity of bulges will be driven by the episodic outflows of AGNs and after the AGN quenched, the radial velocity will tend to zero within a timescale $\simτ_{\rm AGN}$, which is the AGN's lifetime. For some typical values of bulges and AGNs, we find the expansion and contraction velocities are of a few $10\,\rm km\,s^{-1}$ for $10^{10}\,M_\odot$ bulges and mass outflowing rate $500\,M_\odot/\rm yr$, which would give observational signatures.

astro-ph.GA↗

arXiv2024

TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction

Wildfire monitoring and prediction are essential for understanding wildfire behaviour. With extensive Earth observation data, these tasks can be integrated and enhanced through multi-task deep learning models. We present a comprehensive multi-temporal remote sensing dataset for active fire detection, daily wildfire monitoring, and next-day wildfire prediction. Covering wildfire events in the contiguous U.S. from January 2017 to October 2021, the dataset includes 3552 surface reflectance images and auxiliary data such as weather, topography, land cover, and fuel information, totalling 71 GB. The lifecycle of each wildfire is documented, with labels for active fires (AF) and burned areas (BA), supported by manual quality assurance of AF and BA test labels. The dataset supports three tasks: a) active fire detection, b) daily burned area mapping, and c) wildfire progression prediction. Detection tasks use pixel-wise classification of multi-spectral, multi-temporal images, while prediction tasks integrate satellite and auxiliary data to model fire dynamics. This dataset and its benchmarks provide a foundation for advancing wildfire research using deep learning.

cs.CV↗

arXiv2024

RADARSAT Constellation Mission Compact Polarisation SAR Data for Burned Area Mapping with Deep Learning

Monitoring wildfires has become increasingly critical due to the sharp rise in wildfire incidents in recent years. Optical satellites like Sentinel-2 and Landsat are extensively utilized for mapping burned areas. However, the effectiveness of optical sensors is compromised by clouds and smoke, which obstruct the detection of burned areas. Thus, satellites equipped with Synthetic Aperture Radar (SAR), such as dual-polarization Sentinel-1 and quad-polarization RADARSAT-1/-2 C-band SAR, which can penetrate clouds and smoke, are investigated for mapping burned areas. However, there is limited research on using compact polarisation (compact-pol) C-band RADARSAT Constellation Mission (RCM) SAR data for this purpose. This study aims to investigate the capacity of compact polarisation RCM data for burned area mapping through deep learning. Compact-pol m-chi decomposition and Compact-pol Radar Vegetation Index (CpRVI) are derived from the RCM Multi-look Complex product. A deep-learning-based processing pipeline incorporating ConvNet-based and Transformer-based models is applied for burned area mapping, with three different input settings: using only log-ratio dual-polarization intensity images images, using only compact-pol decomposition plus CpRVI, and using all three data sources. The results demonstrate that compact-pol m-chi decomposition and CpRVI images significantly complement log-ratio images for burned area mapping. The best-performing Transformer-based model, UNETR, trained with log-ratio, m-chi decomposition, and CpRVI data, achieved an F1 Score of 0.718 and an IoU Score of 0.565, showing a notable improvement compared to the same model trained using only log-ratio images.

cs.CV↗

arXiv2024

Marco-LLM: Bridging Languages via Massive Multilingual Training for Cross-Lingual Enhancement

Large Language Models (LLMs) have achieved remarkable progress in recent years; however, their excellent performance is still largely limited to major world languages, primarily English. Many LLMs continue to face challenges with multilingual tasks, especially when it comes to low-resource languages. To address this issue, we introduced Marco-LLM: Massive multilingual training for cross-lingual enhancement LLM. We have collected a substantial amount of multilingual data for several low-resource languages and conducted extensive continual pre-training using the Qwen2 models. This effort has resulted in a multilingual LLM named Marco-LLM. Through comprehensive evaluations on various multilingual benchmarks, including MMMLU, AGIEval, Belebele, Flores-200, XCOPA and many others, Marco-LLM has demonstrated substantial improvements over state-of-the-art LLMs. Furthermore, Marco-LLM achieved substantial enhancements in any-to-any machine translation tasks, showing the effectiveness of our multilingual LLM. Marco-LLM is a pioneering multilingual LLM designed to not only perform exceptionally well in multilingual tasks, including low-resource languages, but also maintain strong performance in English and other major languages, closing the performance gap between high- and low-resource language capabilities. By bridging languages, this effort demonstrates our dedication to ensuring LLMs work accurately across various languages.

cs.CL↗

arXiv2024

Marco-o1: Towards Open Reasoning Models for Open-Ended Solutions

Currently OpenAI o1 sparks a surge of interest in the study of large reasoning models (LRM). Building on this momentum, Marco-o1 not only focuses on disciplines with standard answers, such as mathematics, physics, and coding -- which are well-suited for reinforcement learning (RL) -- but also places greater emphasis on open-ended resolutions. We aim to address the question: ''Can the o1 model effectively generalize to broader domains where clear standards are absent and rewards are challenging to quantify?'' Marco-o1 is powered by Chain-of-Thought (CoT) fine-tuning, Monte Carlo Tree Search (MCTS), reflection mechanisms, and innovative reasoning strategies -- optimized for complex real-world problem-solving tasks.

cs.CL↗

arXiv2024

A Simple and Effective $L_2$ Norm-Based Strategy for KV Cache Compression

The deployment of large language models (LLMs) is often hindered by the extensive memory requirements of the Key-Value (KV) cache, especially as context lengths increase. Existing approaches to reduce the KV cache size involve either fine-tuning the model to learn a compression strategy or leveraging attention scores to reduce the sequence length. We analyse the attention distributions in decoder-only Transformers-based models and observe that attention allocation patterns stay consistent across most layers. Surprisingly, we find a clear correlation between the $L_2$ and the attention scores over cached KV pairs, where a low $L_2$ of a key embedding usually leads to a high attention score during decoding. This finding indicates that the influence of a KV pair is potentially determined by the key embedding itself before being queried. Based on this observation, we compress the KV cache based on the $L_2$ of key embeddings. Our experimental results show that this simple strategy can reduce the KV cache size by 50% on language modelling and needle-in-a-haystack tasks and 90% on passkey retrieval tasks without losing accuracy. Moreover, without relying on the attention scores, this approach remains compatible with FlashAttention, enabling broader applicability.

cs.CL↗

arXiv2024

Harnessing with Twisting: Single-Arm Deformable Linear Object Manipulation for Industrial Harnessing Task

Wire-harnessing tasks pose great challenges to be automated by the robot due to the complex dynamics and unpredictable behavior of the deformable wire. Traditional methods, often reliant on dual-robot arms or tactile sensing, face limitations in adaptability, cost, and scalability. This paper introduces a novel single-robot wire-harnessing pipeline that leverages a robot's twisting motion to generate necessary wire tension for precise insertion into clamps, using only one robot arm with an integrated force/torque (F/T) sensor. Benefiting from this design, the single robot arm can efficiently apply tension for wire routing and insertion into clamps in a narrow space. Our approach is structured around four principal components: a Model Predictive Control (MPC) based on the Koopman operator for tension tracking and wire following, a motion planner for sequencing harnessing waypoints, a suite of insertion primitives for clamp engagement, and a fix-point switching mechanism for wire constraint updating. Evaluated on an industrial-level wire harnessing task, our method demonstrated superior performance and reliability over conventional approaches, efficiently handling both single and multiple wire configurations with high success rates.

cs.RO↗

arXiv2024

Training-free LLM-generated Text Detection by Mining Token Probability Sequences

Large language models (LLMs) have demonstrated remarkable capabilities in generating high-quality texts across diverse domains. However, the potential misuse of LLMs has raised significant concerns, underscoring the urgent need for reliable detection of LLM-generated texts. Conventional training-based detectors often struggle with generalization, particularly in cross-domain and cross-model scenarios. In contrast, training-free methods, which focus on inherent discrepancies through carefully designed statistical features, offer improved generalization and interpretability. Despite this, existing training-free detection methods typically rely on global text sequence statistics, neglecting the modeling of local discriminative features, thereby limiting their detection efficacy. In this work, we introduce a novel training-free detector, termed \textbf{Lastde} that synergizes local and global statistics for enhanced detection. For the first time, we introduce time series analysis to LLM-generated text detection, capturing the temporal dynamics of token probability sequences. By integrating these local statistics with global ones, our detector reveals significant disparities between human and LLM-generated texts. We also propose an efficient alternative, \textbf{Lastde++} to enable real-time detection. Extensive experiments on six datasets involving cross-domain, cross-model, and cross-lingual detection scenarios, under both white-box and black-box settings, demonstrated that our method consistently achieves state-of-the-art performance. Furthermore, our approach exhibits greater robustness against paraphrasing attacks compared to existing baseline methods.

cs.CL↗