arXiv subjects

Xiang Shi

Publications and source records attributed to Xiang Shi.

At least 19 recordsLinked to original sources

arXiv2026

Intern-S2-Preview: Scientific Agentic Foundation Model

Scientific discovery increasingly requires AI systems that can reason over scientific evidence of heterogeneous modalities, interact with scientific tools and environments, and sustain progress across long task horizons. We present Intern-S2-Preview, a series of scientific agentic foundation models designed to support multimodal scientific understanding, reasoning, generation, and long-horizon tasks. The training pipeline begins with scientific multimodal pre-training over rendered scientific documents, interleaved image-text data, and diverse scientific corpora. Starting from the pretrained checkpoint, we apply a unified post-training pipeline consisting of supervised fine-tuning, scalable multi-task reinforcement learning (RL), black- and white-box agentic RL, and on-policy distillation. This pipeline is supported by practical techniques that improve rollout and training stability and efficiency, including partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly for agentic tasks. At the architecture level, Intern-S2-Preview-397B extends time series modelling from efficient long-sequence understanding to numerical forecasting, while Memory Decoder is studied as a separate memory-augmented path for rapid scientific specialization without modifying the frozen 397B backbone. Evaluations across scientific, multimodal, agentic, and general-purpose benchmarks show that Intern-S2-Preview-397B achieves competitive or leading results in multiple settings. The time series modules improve scientific signal understanding and forecasting on SciTS, while the separate Intern-MemDec-4B extension improves the Biology-Instructions average score from 56.92 to 60.32 without modifying the frozen 397B backbone.

cs.LG↗

arXiv2026

MemSFT: Mitigating Alignment Tax with an External Parametric Memory

Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose MemSFT, which mitigates the alignment tax by decoupling domain specialization from backbone parameter updates through a plug-and-play parametric memory. The memory is trained to imitate the behavior of a non-parametric retriever operating over domain data, thereby memorizing knowledge and patterns that would otherwise be accessed through retrieval. Once trained on a specific domain, the memory can be reused across LLMs of different sizes. During generation, a learned router dynamically fuses the output distributions of the memory and backbone at each decoding step, allowing domain expertise to be invoked selectively. Across biology, geoscience, and law, evaluations with models ranging from Qwen3-8B to Qwen3-235B-A22B show that MemSFT consistently improves domain performance with negligible degradation in general performance, whereas full SFT suffers severe forgetting on general tasks. Overall, our results demonstrate a practical path to decoupling general model capabilities from domain-specific knowledge at the parameter level, thereby equipping LLMs with new specialized capabilities without compromising their general capabilities.

cs.LG↗

arXiv2026

Deciphering Region-Level Signatures from Latency Measurements in LEO Satellite Internet

Low-Earth orbit (LEO) satellite Internet has become an indispensable infrastructure that provide growing coverage for global users. Despite extensive measurement efforts, the principles underlying region-level performance characteristics remain insufficiently understood, limiting the ability to identify region-specific latency signatures under dynamic network conditions. In this paper, we formulate the problem of region-level latency characterization using Starlink round-trip time (RTT) measurements from the public LENS dataset. We then propose a hierarchical analytical framework that transforms raw RTT sequences into multi-scale statistical features for cross-region comparison. Using data from five geographically representative regions, we demonstrate that latency differences are strongly associated with deployment factors, particularly infrastructure availability and Starlink dish-to-Point-of-Presence distance. Mutual information analysis identifies minimum RTT as the most discriminative feature, which is further supported by XGBoost-based feature importance. The proposed model well achieves 83% accuracy on short-term data. However, its performance degrades over longer periods, indicating limited temporal generalization and motivating the need for adaptive models and feature representations for long-term performance in the future.

cs.LG↗

arXiv2026

A GAN-Based Framework for Robust Data Synthesis in Satellite Internet Observations

Low-Earth orbit (LEO) satellite Internet has become an important infrastructure for enabling ubiquitous connectivity to align with the International Telecommunications Union vision for 6G telecommunications networks. However, current LEO satellite Internet observations often suffer from missing data, which complicates data augmentation task and limits the expansion of representative datasets. Given the complex characteristics of these datasets, generative AI (GenAI) presents a promising approach, yet its application in this domain has received little attention to date. In this paper, we propose a GenAI-based framework to synthesize high-fidelity data directly from incomplete LEO network observations. We propose the representative data missing scenarios, and evaluate the performance with the latest GAN- and VAE-based GenAI models on the recent WetLinks dataset. We design block-wise and point-wise missing scenarios to closely simulate the data loss that happens on real-world LEO satellite networks. Our results show the effectiveness of our proposed GAN-based framework and GT-GAN model exhibits the best performance among all models in both missing scenarios. Even under extreme conditions (e.g., 40% of the input data is missing), GT-GAN shows the highest robustness, consistently capturing the underlying input data distribution and being the least affected in terms of generalization. Our results shed light on future directions for GenAI-based data augmentation methods and data-driven research on satellite network measurement.

cs.AI↗

arXiv2026

Beyond Numerical Features: CNN-Driven Algorithm Selection via Contour Plots for Continuous Black-Box Optimization

The present paper introduces a new representation-driven approach to per-instance algorithm selection, applied to black-box optimization, for automatically choosing the most promising solver from a fixed portfolio. Prior work in continuous optimization largely relies on numerical descriptors, including Exploratory Landscape Analysis features and learned embeddings such as Deep-ELA. This work studies a complementary representation: contour-map visualizations of probed landscapes. A CNN regressor takes multiple instance-specific contour views (stacked or encoded per view and aggregated) and predicts per-solver performance, enabling selection by the predicted best value. On the standard BBOB 2009 single-objective protocol, the resulting selectors significantly outperform the single best solver (SBS) and are competitive with feature-based baselines. A subsequent bi-objective evaluation under the DeepELA setting further indicates that the same image-based principle can be competitive when using windowed contour views. Overall, the results suggest that simple vision models can exploit spatial structure in probed landscapes for algorithm selection without handcrafted ELA features.

cs.LG↗

arXiv2026

The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation

Hypernetwork-based methods such as Doc-to-LoRA internalize a document into an LLM's weights in a single forward pass, but they fail systematically on conflicts: when the document contradicts pretraining knowledge, accuracy collapses to 46.4% on the deepest facts. We show the failure is a magnitude problem rather than a representational one. The hypernetwork already targets the right layers, but its adapter margin is approximately constant across documents while the pretrained margin grows with training frequency, so deep conflicts lose by construction. The account predicts that failure should track prior strength: sorting 194 conflicts by the base model's log-probability on the contradicted fact, baseline accuracy falls from 68% on weak-prior questions to 16% on strong-prior ones, a 52 percentage-point gap. The cure is amplitude. Selective Layer Boosting scales the adapter at its top-norm layers, and Conflict-Aware Internalization triggers boosting only when the base model is confident. Both are training-free; together they raise deep-conflict accuracy from 46.4% to 71.0% on Gemma-2B and from 53.6% to 72.5% on Mistral-7B while preserving novel-knowledge recall, and beat vanilla retrieval-augmented generation on medium conflicts by 18 percentage points despite operating entirely in parameter space. We release KID-Bench, a 489-question benchmark that separates novel recall, cross-knowledge combination, and prior-graded conflicts.

cs.LG↗

arXiv2026

Formalising the Logit Shift Induced by LoRA: A Technical Note

This technical note provides a first-order formalisation of the logit shift and fact-margin change induced by Low-Rank Adaptation (LoRA). Using a first-order Fr\'echet approximation around the base model trajectory, we show that the multi-layer LoRA effect can be decomposed into a linear summation of layerwise contributions and a higher-order remainder term representing inter-layer coupling.

cs.LG↗

arXiv2026

Spatiotemporal Analysis of VIIRS Satellite Observations and Network Traffic During the 2025 Manitoba Wildfires

Climate change has intensified extreme weather and wildfire conditions globally. Canada experienced record-breaking wildfires in 2023 and 2025, burning millions of hectares and severely impacting the Prairie provinces, with Manitoba facing its worst season in 30 years. These events highlight the urgent need to understand and mitigate escalating fire risks. While existing research largely focuses on wildfire management approaches, few studies have explored the relationship between user network traffic and wildfire activity, despite the potential of such correlations to provide valuable spatiotemporal insights into wildfire dynamics. This paper investigates the relationship between wildfire intensity and network performance during the 2025 Manitoba wildfire season, using Visible Infrared Imaging Radiometer Suite (VIIRS) satellite-derived Fire Radiative Power data and large-scale Speedtest measurements. We found statistically significant correlations between wildfire intensity and several network performance metrics in both the province-wide and region-wide case studies, as measured by Spearman's correlation coefficients ($\rho$) and corresponding p-values. Throughput-related metrics showed inverse correlations with wildfire intensity (e.g., download speed: $\rho = -0.214$, $p\_value = 0.004$), whereas latency-related metrics showed positive correlations (e.g., round-trip time latency: $\rho = 0.162$, $p\_value = 0.0308$). The findings suggest satellite fire indicators and network performance metrics together can reveal vulnerabilities during extreme environmental events and support diaster response and recovery efforts.

cs.CY↗

arXiv2026

GeoPAS: Geometric Probing for Algorithm Selection in Continuous Black-Box Optimization

Automated algorithm selection for continuous black-box optimization depends on representing problem information under limited probing and selecting solvers under heavy-tailed performance distributions. This paper proposes a geometric probing framework that represents each problem instance by randomly sampled multi-scale two-dimensional slices of the objective landscape. The slices are encoded with validity-mask-aware visual pooling and aggregated into an instance representation. Solver selection is then performed by a logarithmic composite score combining a learned instance-conditioned estimate with an algorithm-side empirical prior. The framework is evaluated on a standard single-objective black-box optimization benchmark suite with a portfolio of twelve solvers under instance-level, grouped random, and problem-level transfer protocols. Under the two within-suite protocols, it reduces aggregate mean relative expected running time from 30.37 for the single best solver to 3.14 and 3.61, while also improving median and upper-tail performance. Under problem-level transfer, the canonical adaptive setting improves typical and moderate-tail performance but leaves the mean dominated by rare extreme failures; a prior-heavy scoring variant mitigates this failure mode, although its robustness may be benchmark-dependent. The results suggest that coarse geometric probes provide useful solver-relevant information, while robust cross-problem selection also depends on metric-aligned decision scoring.

cs.LG↗

arXiv2026

ROI-Driven Foveated Attention for Unified Egocentric Representations in Vision-Language-Action Systems

The development of embodied AI systems is increasingly constrained by the availability and structure of physical interaction data. Despite recent advances in vision-language-action (VLA) models, current pipelines suffer from high data collection cost, limited cross-embodiment alignment, and poor transfer from internet-scale visual data to robot control. We propose a region-of-interest (ROI) driven engineering workflow that introduces an egocentric, geometry-grounded data representation. By projecting end-effector poses via forward kinematics (FK) into a single external camera, we derive movement-aligned hand-centric ROIs without requiring wrist-mounted cameras or multi-view systems. Unlike directly downsampling the full frame, ROI is cropped from the original image before resizing, preserving high local information density for contact-critical regions while retaining global context. We present a reproducible pipeline covering calibration, synchronization, ROI generation, deterministic boundary handling, and metadata governance. The resulting representation is embodiment-aligned and viewpoint-normalized, enabling data reuse across heterogeneous robots. We argue that egocentric ROI serves as a practical data abstraction for scalable collection and cross-embodiment learning, bridging internet-scale perception and robot-specific control.

cs.RO↗

arXiv2026

SaiVLA-0: Cerebrum--Pons--Cerebellum Tripartite Architecture for Compute-Aware Vision-Language-Action

We revisit Vision-Language-Action through a neuroscience-inspired triad. Biologically, the Cerebrum provides stable high-level multimodal priors and remains frozen; the Pons Adapter integrates these cortical features with real-time proprioceptive inputs and compiles intent into execution-ready tokens; and the Cerebellum (ParaCAT) performs fast, parallel categorical decoding for online control, with hysteresis/EMA/temperature/entropy for stability. A fixed-ratio schedule and two-stage feature caching make the system compute-aware and reproducible. Inspired by active, foveated vision, our wrist ROIs are geometrically tied to the end-effector via calibrated projection, providing a movement-stabilized, high-resolution view that is sensitive to fine-grained pose changes and complements the global context of the main view. The design is modular: upgrading the Cerebrum only retrains the Pons; changing robots only trains the Cerebellum; cerebellum-only RL can further refine control without touching high-level semantics. As a concept-and-protocol paper with preliminary evidence, we outline a timing protocol under matched conditions (GPU, resolution, batch) to verify anticipated efficiency gains. We also report preliminary LIBERO evidence showing that split feature caching reduces training time (7.5h to 4.5h) and improves average success (86.5% to 92.5%) under official N1.5 head-only training, and that SaiVLA0 reaches 99.0% mean success.

cs.RO↗

arXiv2025

Interplay Between Belief Propagation and Transformer: Differential-Attention Message Passing Transformer

Transformer-based neural decoders have emerged as a promising approach to error correction coding, combining data-driven adaptability with efficient modeling of long-range dependencies. This paper presents a novel decoder architecture that integrates classical belief propagation principles with transformer designs. We introduce a differentiable syndrome loss function leveraging global codebook structure and a differential-attention mechanism optimizing bit and syndrome embedding interactions. Experimental results demonstrate consistent performance improvements over existing transformer-based decoders, with our approach surpassing traditional belief propagation decoders for short-to-medium length LDPC codes.

cs.IT↗

arXiv2025

Modality Equilibrium Matters: Minor-Modality-Aware Adaptive Alternating for Cross-Modal Memory Enhancement

Multimodal fusion is susceptible to modality imbalance, where dominant modalities overshadow weak ones, easily leading to biased learning and suboptimal fusion, especially for incomplete modality conditions. To address this problem, we propose a Shapley-guided alternating training framework that adaptively prioritizes minor modalities to balance and thus enhance the fusion. Our method leverages Shapley Value-based scheduling to improve the training sequence adaptively, ensuring that under-optimized modalities receive sufficient learning. Additionally, we introduce the memory module to refine and inherit modality-specific representations with a cross-modal mapping mechanism to align features at both the feature and sample levels. To further validate the adaptability of the proposed approach, the encoder module empirically adopts both conventional and LLM-based backbones. With building up a novel multimodal equilibrium metric, namely, equilibrium deviation metric (EDM), we evaluate the performance in both balance and accuracy across four multimodal benchmark datasets, where our method achieves state-of-the-art (SOTA) results. Meanwhile, robustness analysis under missing modalities highlights its strong generalization capabilities. Accordingly, our findings reveal the untapped potential of alternating training, demonstrating that strategic modality prioritization fundamentally balances and promotes multimodal learning, offering a new paradigm for optimizing multimodal training dynamics.

cs.LG↗

arXiv2025

Understanding Stragglers in Large Model Training Using What-if Analysis

Large language model (LLM) training is one of the most demanding distributed computations today, often requiring thousands of GPUs with frequent synchronization across machines. Such a workload pattern makes it susceptible to stragglers, where the training can be stalled by few slow workers. At ByteDance we find stragglers are not trivially always caused by hardware failures, but can arise from multiple complex factors. This work aims to present a comprehensive study on the straggler issues in LLM training, using a five-month trace collected from our ByteDance LLM training cluster. The core methodology is what-if analysis that simulates the scenario without any stragglers and contrasts with the actual case. We use this method to study the following questions: (1) how often do stragglers affect training jobs, and what effect do they have on job performance; (2) do stragglers exhibit temporal or spatial patterns; and (3) what are the potential root causes for stragglers?

cs.DC↗

arXiv2024

Scalable Hierarchical Reinforcement Learning for Hyper Scale Multi-Robot Task Planning

To improve the efficiency of warehousing system and meet huge customer orders, we aim to solve the challenges of dimension disaster and dynamic properties in hyper scale multi-robot task planning (MRTP) for robotic mobile fulfillment system (RMFS). Existing research indicates that hierarchical reinforcement learning (HRL) is an effective method to reduce these challenges. Based on that, we construct an efficient multi-stage HRL-based multi-robot task planner for hyper scale MRTP in RMFS, and the planning process is represented with a special temporal graph topology. To ensure optimality, the planner is designed with a centralized architecture, but it also brings the challenges of scaling up and generalization that require policies to maintain performance for various unlearned scales and maps. To tackle these difficulties, we first construct a hierarchical temporal attention network (HTAN) to ensure basic ability of handling inputs with unfixed lengths, and then design multi-stage curricula for hierarchical policy learning to further improve the scaling up and generalization ability while avoiding catastrophic forgetting. Additionally, we notice that policies with hierarchical structure suffer from unfair credit assignment that is similar to that in multi-agent reinforcement learning, inspired of which, we propose a hierarchical reinforcement learning algorithm with counterfactual rollout baseline to improve learning performance. Experimental results demonstrate that our planner outperform other state-of-the-art methods on various MRTP instances in both simulated and real-world RMFS. Also, our planner can successfully scale up to hyper scale MRTP instances in RMFS with up to 200 robots and 1000 retrieval racks on unlearned maps while keeping superior performance over other methods.

cs.RO↗

arXiv2024

Minder: Faulty Machine Detection for Large-scale Distributed Model Training

Large-scale distributed model training requires simultaneous training on up to thousands of machines. Faulty machine detection is critical when an unexpected fault occurs in a machine. From our experience, a training task can encounter two faults per day on average, possibly leading to a halt for hours. To address the drawbacks of the time-consuming and labor-intensive manual scrutiny, we propose Minder, an automatic faulty machine detector for distributed training tasks. The key idea of Minder is to automatically and efficiently detect faulty distinctive monitoring metric patterns, which could last for a period before the entire training task comes to a halt. Minder has been deployed in our production environment for over one year, monitoring daily distributed training tasks where each involves up to thousands of machines. In our real-world fault detection scenarios, Minder can accurately and efficiently react to faults within 3.6 seconds on average, with a precision of 0.904 and F1-score of 0.893.

cs.DC↗

arXiv2024

Every Part Matters: Integrity Verification of Scientific Figures Based on Multimodal Large Language Models

This paper tackles a key issue in the interpretation of scientific figures: the fine-grained alignment of text and figures. It advances beyond prior research that primarily dealt with straightforward, data-driven visualizations such as bar and pie charts and only offered a basic understanding of diagrams through captioning and classification. We introduce a novel task, Figure Integrity Verification, designed to evaluate the precision of technologies in aligning textual knowledge with visual elements in scientific figures. To support this, we develop a semi-automated method for constructing a large-scale dataset, Figure-seg, specifically designed for this task. Additionally, we propose an innovative framework, Every Part Matters (EPM), which leverages Multimodal Large Language Models (MLLMs) to not only incrementally improve the alignment and verification of text-figure integrity but also enhance integrity through analogical reasoning. Our comprehensive experiments show that these innovations substantially improve upon existing methods, allowing for more precise and thorough analysis of complex scientific figures. This progress not only enhances our understanding of multimodal technologies but also stimulates further research and practical applications across fields requiring the accurate interpretation of complex visual data.

cs.CL↗

arXiv2024

Let's Learn Step by Step: Enhancing In-Context Learning Ability with Curriculum Learning

Demonstration ordering, which is an important strategy for in-context learning (ICL), can significantly affects the performance of large language models (LLMs). However, most of the current approaches of ordering require high computational costs to introduce the priori knowledge. In this paper, inspired by the human learning process, we propose a simple but effective demonstration ordering method for ICL, named the few-shot In-Context Curriculum Learning (ICCL). The ICCL implies gradually increasing the complexity of prompt demonstrations during the inference process. The difficulty can be assessed by human experts or LLMs-driven metrics, such as perplexity. Then we design extensive experiments to discuss the effectiveness of the ICCL at both corpus-level and instance-level. Moreover, we also investigate the formation mechanism of LLM's ICCL capability. Experimental results demonstrate that ICCL, developed during the instruction-tuning stage, is effective for representative open-source LLMs. To facilitate further research and applications by other scholars, we make the code publicly available.

cs.CL↗