Searcharxiv⌕ Search

arXiv subjects

Ziyue Liu

Publications and source records attributed to Ziyue Liu.

At least 19 recordsLinked to original sources

Training Specialist LLMs on Your Phone: Forward-Only Optimization via Block-Spectral Coordinates

Large language models (LLMs) on personal devices keep user data local, but limited memory restricts their general capability. Since most user-specific tasks do not require the breadth of frontier models, a practical solution is to adapt a shared on-device model into lightweight specialists. However, backpropagation is prohibitive in memory, and users rarely provide labeled data. We present MobileSpec, a system that trains specialists with forward passes alone and learns from automatically checked task outcomes, as in reinforcement learning with verifiable rewards (RLVR). Because forward-only updates are built from random perturbations, training is efficient only when these perturbations resemble the weight changes of RLVR, which largely preserve the singular values of pretrained weights and rotate their singular vectors. The core algorithm of MobileSpec, FuRA-S, produces such perturbations while training only the block-spectral coordinates of pretrained weights, namely the singular values of their column blocks. As we show theoretically and empirically, rescaling these blocks independently shifts their relative contributions to each weight matrix and thereby rotates its singular vectors. Since this update can be rewritten as a transformation of the layer input, FuRA-S also adapts INT4 models without modifying their weights. With specialists of only 1-5 MB that update 0.015-0.05% of the parameters, MobileSpec raises the success rate of a phone-use agent from 6-23% up to 89-100% and trains a personalized 8B model on an iPhone 16 Pro overnight. On RLVR benchmarks, FuRA-S is competitive with full-parameter evolution strategies (ES) and outperforms ES over LoRA adapters and quantized ES on INT4 models.

cs.LG↗

Controlling Action-Head Co-Adaptation to VLM Representation Anisotropy

Modern learning systems increasingly combine pretrained models with downstream heads. Despite the growing importance of such modular designs, how the downstream model co-adapts to the geometry of pretrained representations remains relatively underexplored. We study this problem through vision-language-action models (VLAs), where a pretrained vision-language model (VLM) conditions an action model for control. We find that the action model progressively concentrates its sensitivity on high-variance dimensions of the VLM outputs. Through controlled experiments across different backbone-head models, we further show that counteracting this sensitivity concentration can be beneficial under representation--task mismatch settings. Motivated by these observations, we introduce \emph{Adaptive Interface Equalizer} (AdaIE), a parameter-free mechanism that reweights pretrained representations according to the evolving downstream sensitivity during training. AdaIE improves VLA success rates on simulation benchmarks, with gains of up to $+5.28$ percentage points (pp) on RoboCasa-GR1 and $+1.68$ pp on RoboTwin 2.0, and achieves an average gain of $+19$ pp across five real-world manipulation tasks. It further generalizes to diffusion policies and world models, improving downstream performance in both settings. Our results highlight interface co-adaptation as an important optimization problem when coupling pretrained representations with downstream models.

cs.AI↗

HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training

HoloBoxing presents a simple recipe for learning humanoid boxing by building on a pretrained whole-body controller. The recipe separates reusable motor skills from task decisions: a compact planner requests future body-point targets, while the frozen controller handles joint-level execution. A learned decoder maps 32-dimensional actions to structured target windows, and measured-state imitation initializes the planner before reinforcement learning. Task learning then refines interaction behavior while reusing the decoder and controller, avoiding the need to relearn whole-body control. This combination keeps task optimization focused on a compact upper policy and supplies an initial behavior before reward-driven exploration. Starting from that available controller, the recipe accounts for about seven minutes of additional training, including reused representation fitting and a single-GPU task run. This excludes controller pretraining, data preparation, evaluation, and search. Simulated boxing experiments and component controls evaluate the resulting behavior and the roles of imitation initialization and the target representation.

cs.RO↗

DeepOHeat-v2: Self-Improving Operator Learning for Fast and Trustworthy Thermal Optimization in 3D-IC Design

Thermal-aware optimization of multi-die 3D integrated circuits evaluates many designs, each a costly heat-equation solve. Operator-learning surrogates replace this solve with a fast forward pass, ideally trained from physics alone, without labeled data. DeepOHeat-v1 made such surrogates fast and trustworthy, but only on low-contrast geometries. High-contrast multi-die stacks break it in two ways: discontinuous conductivities make the continuous physics loss ill-defined at material interfaces, and ill-conditioning ($κ_2(A_h) \approx 6 \times 10^4$) puts the discretized strong-form loss beyond first-order optimization. We propose DeepOHeat-v2 to overcome both. First, we train on a discretized physics loss that handles the discontinuities natively; its energy form reduces the prediction-space loss-Hessian conditioning from $κ^2$ to $κ$, and a matrix-preconditioned optimizer cuts the mean peak temperature error from over 30 K to 0.55 K. Second, because optimization leaves the training distribution, we propose a self-improving framework: a hotspot trust gate sends flagged placements to a reference solver, and the surrogate incrementally retrains on the refined solutions, keeping an update only when it improves held-out validation error. On a multi-die benchmark, the surrogate-true peak gap on the returned design falls from 1.12 K to 0.11 K, matching a solve-at-every-step optimizer while running $56\times$ faster.

cs.LG↗

Deep Learning Imputation of Missing Radius of Maximum Winds (Rmax) Values in Tropical Cyclone Best-Track Data

Probabilistic coastal hazard assessments require accurate characterization of tropical cyclone (TC) parameters, yet datasets often contain missing records for the radius of maximum winds (Rmax), a key variable in Joint Probability Method analyses. This study evaluates data-driven approaches for Rmax imputation, including one-dimensional Convolutional Neural Networks (1DCNNs), Long Short-Term Memory (LSTM) networks, and conventional machine learning models. We examine physics-informed input augmentation, temporal modeling, and transfer learning using synthetic RAFT and STORM datasets for pre-training and observational IBTrACS data for fine-tuning. Including the radius of 34-knot winds (R34) substantially improves performance across all model types. Temporal models achieve higher average correlations than non-temporal models despite using approximately an order of magnitude fewer samples, indicating better preservation of relative Rmax variability across storms. This advantage is more pronounced when R34 is unavailable, suggesting temporal information can partially compensate for missing storm-size predictors. Transfer learning does not improve performance, likely because synthetic datasets have lower and less variable Rmax distributions than IBTrACS. These findings demonstrate the potential of temporal deep learning for reconstructing incomplete TC records and highlight the importance of physics-informed inputs, observational data availability, and distributional consistency in coastal hazard assessment.

cs.LG↗

MuonQ: Enhancing Low-Bit Muon Quantization via Directional Fidelity Optimization

The Muon optimizer has emerged as a compelling alternative to Adam for training large language models, achieving remarkable computational savings through gradient orthogonalization. However, Muon's optimizer state is more sensitive to quantization errors: because the orthogonalization discards the magnitudes of singular values and retains only directional information, even small quantization errors in singular vector directions are amplified in the update. In this work, we propose MuonQ, a low-bit Muon training framework built on the principle of directional fidelity optimization. First, we apply a pre-quantization normalization so that each step introduces quantization errors of the same magnitude, preventing the accumulated error from developing a preferred direction. Second, we introduce a structural decomposition that separately quantizes the dominant singular components via power iteration, ensuring that quantization errors perturb only singular value magnitudes rather than rotating singular vector directions. Third, we adopt $μ$-law companding quantization to allocate higher resolution to densely packed momentum values, shifting the quantization objective from outlier preservation to dense-region distinguishability. Together, these techniques enable stable 4-bit quantization of Muon's optimizer states. Pre-training experiments on GPT-style and LLaMA-style models demonstrate that MuonQ at 4-bit precision recovers most of full-precision Muon's training loss and downstream accuracy while reducing optimizer-state memory by up to 7.3$\times$, offering a favorable point on the accuracy-memory trade-off. Our code is available at https://github.com/YupengSu/MuonQ.

cs.LG↗

Parameter-Efficient Continual Fine-Tuning: A Survey

The emergence of large pre-trained networks has revolutionized the AI field, unlocking new possibilities and achieving unprecedented performance. However, these models inherit a fundamental limitation from traditional Machine Learning approaches: their strong dependence on the \textit{i.i.d.} assumption hinders their adaptability to dynamic learning scenarios. We believe the next breakthrough in AI lies in enabling efficient adaptation to evolving environments -- such as the real world -- where new data and tasks arrive sequentially. This challenge defines the field of Continual Learning (CL), a Machine Learning paradigm focused on developing lifelong learning neural models. One alternative to efficiently adapt these large-scale models is known Parameter-Efficient Fine-Tuning (PEFT). These methods tackle the issue of adapting the model to a particular data or scenario by performing small and efficient modifications, achieving similar performance to full fine-tuning. However, these techniques still lack the ability to adjust the model to multiple tasks continually, as they suffer from the issue of Catastrophic Forgetting. In this survey, we first provide an overview of CL algorithms and PEFT methods before reviewing the state-of-the-art on Parameter-Efficient Continual Fine-Tuning (PECFT). We examine various approaches, discuss evaluation metrics, and explore potential future research directions. Our goal is to highlight the synergy between CL and Parameter-Efficient Fine-Tuning, guide researchers in this field, and pave the way for novel future research directions.

cs.LG↗

VAR-Q: Tuning-free KV Cache Quantization for Visual Autoregressive Image and Video Generation

Visual autoregressive generation spans two dominant paradigms: next-scale prediction for images and video, and next-frame prediction for video. Both decode block-by-block rather than token-by-token, yielding KV caches whose statistics diverge from those of language models. At scale, KV memory exceeds the weights and limits large-scale deployment caused by a large amount of visual tokens. Our analysis of these caches reveals three findings. First, a small set of massive-value channels carries the functional load: ablating these outlier channels collapses generation, whereas ablating a matched set of inlier channels leaves quality intact. Second, the outliers organize along two axes, feature and block, with magnitudes persistent within each cell and shifting across cells. Third, this structure arises from two independent mechanisms: projection bias and RoPE-induced scattering. Guided by these findings, we propose VAR-Q, a tuning-free KV-cache quantization framework that groups along the same two axes. The generation loop provides block boundaries for free, removing the need for outlier detection or calibration. Paired with bit-packing kernels for runtime memory reduction, VAR-Q delivers low-bit quantization across nine models spanning both paradigms, cutting KV memory by up to 78.5% with negligible quality degradation and enabling 6x larger inference batches.

cs.CV↗

Muon$^2$: Boosting Muon via Adaptive Second-Moment Preconditioning

Muon has emerged as a promising optimizer for large-scale foundation model pre-training by exploiting the matrix structure of neural network updates through iterative orthogonalization. However, the orthogonalization quality of Muon hinges on the number of Newton--Schulz (NS) iterations performed, which poses efficiency challenges due to its non-trivial computation and communication cost. We propose Muon$^2$, an extension of Muon, to improve both quality and efficiency by applying Adam-style adaptive second-moment preconditioning before orthogonalization. Our key insight is that the core challenge of polar approximation in Muon lies in the ill-conditioned momentum matrix, of which the spectrum is substantially improved by Muon$^2$, leading to faster convergence toward a practically sufficient orthogonalization. We further characterize the practical orthogonalization quality via directional alignment, under which Muon$^2$ demonstrates dramatic improvement over Muon at each polar step. Across GPT, LLaMA, and Mixture-of-Experts pre-training experiments up to 13B parameters, Muon$^2$ (and its memory-efficient variant Muon$^2$-F that preserves most of its benefits) consistently outperforms Muon and its variants while reducing NS iterations by 40%, and saves up to 1/4 training time over Muon when achieving the same loss.

cs.LG↗

ReCoVer: Resilient LLM Pre-Training System via Fault-Tolerant Collective and Versatile Workload

Pre-training large language models on massive GPU clusters has made hardware faults routine rather than rare, driving the need for resilient training systems. Yet existing frameworks either focus on specific parallelism schemes or risk drifting away from a failure-free training trajectory. We propose ReCoVer, a resilient LLM pre-training system that upholds a single invariant: each iteration keeps the number of microbatches constant, ensuring per-iteration gradients remain stochastically equivalent to a failure-free run. The framework is organized as three decoupled protocol layers: (1) Fault-tolerant collectives that isolate faults from propagating across replicas; (2) in-step fine-grained recovery that preserves intra-iteration progress and prevents gradient corruption; (3) versatile-workload policy that dynamically redistributes microbatch quotas across the survivors. The design is parallelism-agnostic, integrating directly with both 3D parallelism and Hybrid Sharded Data Parallel (HSDP) as a drop-in substrate. We evaluate our implementation on end-to-end pre-training tasks for up to 512 GPUs, ReCoVer successfully preserves the training trajectory from a failure-free reference despite of 256 GPUs lost spread across the run. For comparison with checkpoint-and-restart baselines, ReCoVer demonstrates $2.23\times$ higher effective throughput after successive failures. This advantage results in ReCoVer processing 74.9% more tokens at 234 GPU-hours, with the gap widening as the training prolongs.

cs.DC↗

MUON+: Towards More Effective Muon via One Additional Normalization Step for LLM Pre-training

Muon has recently emerged as a strong optimizer for large language model pre-training, orthogonalizing the momentum matrix via Newton--Schulz polar iterations. A natural intuition is that polar iterations, by flattening the singular spectrum to all ones, should also eliminate column- and row-wise norm imbalance in the update. We show that this is not true in practice: practical polar steps can substantially amplify the imbalance. We term this the post-polar imbalanced update problem, and prove that such imbalance tightens the second-order term in a blockwise descent analysis, weakening Muon's per-step descent guarantee. Motivated by this analysis, we propose Muon+, a one-line fix that inserts a single normalization step after polar orthogonalization. Muon+ adds no optimizer state. Across pre-training experiments on GPT and LLaMA models from 60M to 7B parameters, spanning both compute-optimal budgets and extended token-to-parameter ratios up to approximately 200, Muon+ consistently outperforms Muon in terms of training and validation perplexity, leading to significant overall pre-training speedup.

cs.LG↗

BOOST: BOttleneck-Optimized Scalable Training Framework for Low-Rank Large Language Models

The scale of transformer model pre-training is constrained by the increasing computation and communication cost. Low-rank bottleneck architectures offer a promising solution to significantly reduce the training time and memory footprint with minimum impact on accuracy. Despite algorithmic efficiency, bottleneck architectures scale poorly under standard tensor parallelism. Simply applying 3D parallelism designed for full-rank methods leads to excessive communication and poor GPU utilization. To address this limitation, we propose BOOST, an efficient training framework tailored for large-scale low-rank bottleneck architectures. BOOST introduces a novel Bottleneck-aware Tensor Parallelism, and combines optimizations such as online-RMSNorm, linear layer grouping, and low-rank activation checkpointing to achieve end-to-end training speedup. Evaluations on different low-rank bottleneck architectures demonstrate that BOOST achieves 1.46-1.91$\times$ speedup over full-rank model baselines and 1.87-2.27$\times$ speedup over low-rank model with naively integrated 3D parallelism, with improved GPU utilization and reduced communication overhead.

cs.LG↗

Assessing the Risks of Typhoon-Induced Multi-Hazards in South Korea

Tropical cyclone-induced coastal hazards can significantly damage coastal infrastructure, and these risks may intensify under future climate change. As a result, there is increasing interest in conducting comprehensive assessments of coastal hazards-including storm surge, storm wind, storm rainfall, and their combined impacts-associated with tropical cyclone events. Risk assessments that overlook the compounding nature of these hazards may lead to ineffective or insufficient mitigation strategies. This study seeks to identify and evaluate the available data, models, and methodologies for assessing both individual and compound typhoon-induced hazards in South Korea. Particular effort is devoted to exploring how established approaches from the North Atlantic region can be adapted, integrated, and extended for application in the South Korean context. Multiple sites across South Korea are analyzed to illustrate the strengths and limitations of these methods.

physics.ao-ph↗

Multi-Level Conditioning by Pairing Localized Text and Sketch for Fashion Image Generation

Sketches offer designers a concise yet expressive medium for early-stage fashion ideation by specifying structure, silhouette, and spatial relationships, while textual descriptions complement sketches to convey material, color, and stylistic details. Effectively combining textual and visual modalities requires adherence to the sketch visual structure when leveraging the guidance of localized attributes from text. We present LOcalized Text and Sketch with multi-level guidance (LOTS), a framework that enhances fashion image generation by combining global sketch guidance with multiple localized sketch-text pairs. LOTS employs a Multi-level Conditioning Stage to independently encode local features within a shared latent space while maintaining global structural coordination. Then, the Diffusion Pair Guidance stage integrates both local and global conditioning via attention-based guidance within the diffusion model's multi-step denoising process. To validate our method, we develop Sketchy, the first fashion dataset where multiple text-sketch pairs are provided per image. Sketchy provides high-quality, clean sketches with a professional look and consistent structure. To assess robustness beyond this setting, we also include an "in the wild" split with non-expert sketches, featuring higher variability and imperfections. Experiments demonstrate that our method strengthens global structural adherence while leveraging richer localized semantic guidance, achieving improvement over state-of-the-art. The dataset, platform, and code are publicly available.

cs.CV↗

SWE-AGI: Benchmarking Specification-Driven Software Construction with MoonBit in the Era of Autonomous Agents

Although large language models (LLMs) have demonstrated impressive coding capabilities, their ability to autonomously build production-scale software from explicit specifications remains an open question. We introduce SWE-AGI, an open-source benchmark for evaluating end-to-end, specification-driven construction of software systems written in MoonBit. SWE-AGI tasks require LLM-based agents to implement parsers, interpreters, binary decoders, and SAT solvers strictly from authoritative standards and RFCs under a fixed API scaffold. Each task involves implementing 1,000-10,000 lines of core logic, corresponding to weeks or months of engineering effort for an experienced human developer. By leveraging the nascent MoonBit ecosystem, SWE-AGI minimizes data leakage, forcing agents to rely on long-horizon architectural reasoning rather than code retrieval. Across frontier models, gpt-5.3-codex achieves the best overall performance (solving 19/22 tasks, 86.4%), outperforming claude-opus-4.6 (15/22, 68.2%), and kimi-2.5 exhibits the strongest performance among open-source models. Performance degrades sharply with increasing task difficulty, particularly on hard, specification-intensive systems. Behavioral analysis further reveals that as codebases scale, code reading, rather than writing, becomes the dominant bottleneck in AI-assisted development. Overall, while specification-driven autonomous software engineering is increasingly viable, substantial challenges remain before it can reliably support production-scale development.

cs.SE↗

Scalable Back-Propagation-Free Training of Optical Physics-Informed Neural Networks

Physics intelligence and digital twins often require rapid and repeated performance evaluation of various engineering systems (e.g. robots, autonomous vehicles, semiconductor chips) to enable (almost) real-time actions or decision making. This has motivated the development of accelerated partial differential equation (PDE) solvers, in resource-constrained scenarios if the PDE solvers are to be deployed on the edge. Physics-informed neural networks (PINNs) have shown promise in solving high-dimensional PDEs, but the training time on state-of-the-art digital hardware (e.g., GPUs) is still orders-of-magnitude longer than the latency required for enabling real-time decision making. Photonic computing offers a potential solution to address this huge latency gap because of its ultra-high operation speed. However, the lack of photonic memory and the large device sizes prevent training real-size PINNs on photonic chips. This paper proposes a completely back-propagation-free (BP-free) and highly salable framework for training real-size PINNs on silicon photonic platforms. Our approach involves three key innovations: (1) a sparse-grid Stein derivative estimator to avoid the BP in the loss evaluation of a PINN, (2) a dimension-reduced zeroth-order optimization via tensor-train decomposition to achieve better scalability and convergence in BP-free training, and (3) a scalable on-chip photonic PINN training accelerator design using photonic tensor cores. We validate our numerical methods on both low- and high-dimensional PDE benchmarks. Through pre-silicon simulation based on real device parameters, we further demonstrate the significant performance benefit (e.g., real-time training, huge chip area reduction) of our photonic accelerator.

cs.LG↗

TEON: Tensorized Orthonormalization Beyond Layer-Wise Muon for Large Language Model Pre-Training

The Muon optimizer has demonstrated strong empirical performance in pre-training large language models by performing matrix-level gradient (or momentum) orthogonalization in each layer independently. In this work, we propose TEON, a principled generalization of Muon that extends orthogonalization beyond individual layers by modeling the gradients of a neural network as a structured higher-order tensor. We present TEON's improved convergence guarantee over layer-wise Muon, and further develop a practical instantiation of TEON based on the theoretical analysis with corresponding ablation. We evaluate our approach on two widely adopted architectures: GPT-style models, ranging from 130M to 774M parameters, and LLaMA-style models, ranging from 60M to 1B parameters. Experimental results show that TEON consistently improves training and validation perplexity across model scales and exhibits strong robustness under various approximate SVD schemes.

cs.LG↗

An Efficient Regional Storm Surge Surrogate Model Training Strategy Under Evolving Landscape and Climate Scenarios

Coastal communities face significant risk from storm-induced coastal flooding, which causes substantial societal and economic losses worldwide. Machine learning techniques have increasingly been integrated into coastal hazard modeling, particularly for storm surge prediction, due to advances in computational capacity. However, incorporating multiple projected future climate and landscape scenarios requires extensive numerical simulations of synthetic storm suites over large geospatial domains, resulting in rapidly escalating computational costs. This study proposes a cost-effective training data reduction strategy for machine learning based storm surge surrogate models that enables efficient incorporation of new future scenarios while minimizing computational burden. The proposed strategy reduces training data across three dimensions: grid points, input features, and storm suite size. Reducing the storm suite size for future scenario simulations is highly effective in guiding numerical simulations, yielding substantial reductions in simulation cost. The performance of surrogate models trained on reduced datasets was evaluated using different machine learning algorithms. Results demonstrate that the proposed reduction strategy is robust across different model types. When trained using 5,000 out of 80,000 grid points, 10 out of 12 input features, and 60 out of 90 storms, the total training dataset is reduced to approximately 5% of its original size. Despite this reduction, the trained model achieves a correlation coefficient of 0.94, comparable to models trained on the full dataset. In addition, storm selection methodologies are introduced to support efficient storm set expansion for future scenario analyses.

physics.ao-ph↗