arXiv subjects

Zheng Zhang

Publications and source records attributed to Zheng Zhang.

At least 109 records · Page 6Linked to original sources

arXiv2026

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↗

arXiv2026

CachePrune: Privacy-Aware and Fine-Grained KV Cache Sharing for Efficient LLM Inference

Large Language Models (LLMs) rely on Key-Value (KV) caching to accelerate inference, and many serving systems further share the KV cache across users' requests to reduce redundant computation. While widely adopted, unrestricted cross-user sharing introduces side-channel vulnerabilities, allowing an adversary to infer user inputs by probing for cache reuse. Existing defenses disable sharing entirely to prevent leakage; yet such a coarse-grained strategy sacrifices substantial reuse potential, since prompts often include large portions of privacy-irrelevant segments, such as system instructions or publicly accessible materials. Building on this, we present CachePrune, a privacy-aware KV cache sharing mechanism that enables fine-grained reuse of KV entries across requests. Realizing such fine granularity requires token-level cache management, as reusable segments vary in length and position due to sensitivity masking, making reuse more complex than the fixed-size or sentence-level chunking used in existing coarse-grained schemes. Specifically, CachePrune makes fine-grained reuse practical by addressing two key challenges: accurately and efficiently deriving reusable KV segments and efficiently retrieving them over variable-length spans. We implement CachePrune on top of vLLM and evaluate it on three datasets, showing that it eliminates direct leakage through KV cache reuse side channels while reducing TTFT by 4.5x and increasing cache hit rates by 44% compared with state-of-the-art approaches.

cs.CR↗

arXiv2026

FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning

Both full fine-tuning (Full FT) and parameter-efficient fine-tuning methods such as LoRA introduce weight updates without accounting for the spectral structure established during pretraining. As a result, noisy gradients from limited fine-tuning data can perturb robust pretrained features. We identify spectral preconditioning as the missing ingredient: reparameterizing each weight matrix through its full-rank singular value decomposition (SVD) and freezing one singular basis constrains updates to the pretrained column space, yielding a preconditioned optimization scheme that outperforms unconstrained Full FT at the same trainable parameter count. Building on this insight, we propose FuRA (Full-Rank Adaptation), an efficient full-rank adaptation framework based on a block tensor-train factorization W = LSR, where the large core L is fixed to the pretrained block-wise SVD basis, while only the compact core R and the block-wise singular values S are optimized. This design simultaneously provides full-rank spectral preconditioning, preserves full-rank update expressivity, and achieves parameter, memory, and step-time efficiency comparable to LoRA. FuRA consistently outperforms Full FT across multiple settings, including LLM fine-tuning (+1.37 on LLaMA-3-8B commonsense reasoning), LLM reinforcement learning for mathematical reasoning, and visual instruction tuning for VLMs. Furthermore, the 4-bit quantized variant, QFuRA, also surpasses QLoRA. Code is available at https://github.com/olokevin/FuRA-NIPS

cs.LG↗

arXiv2026

Moduli of Persson surfaces: The compactification via KSBA stable pairs and a generic global Torelli type theorem

We study a family of canonically polarized surfaces introduced by Persson, which arise as Galois $G=(\mathbb{Z}/2\mathbb{Z})^4$-covers of $\mathbf{P}^2$ branched along eight general lines. For this family, we construct the compactified moduli space and explicitly describe the stable degenerations in the sense of Kollár, Shepherd-Barron, and Alexeev (KSBA) via stable pairs of weighted hyperplane arrangements. By computing the $\mathbb{Q}$-Gorenstein obstructions and using the KSBA wall crossings, we show that the resulting compactified moduli stack is smooth. Furthermore, we establish a generic global Torelli type result: up to two possibilities, a generic smooth Persson surface can be recovered from the Hodge structure on the anti-invariant part of the second cohomology of its étale double cover, together with the associated $\widetilde{G}=(\mathbb{Z}/2\mathbb{Z})^5$-action.

math.AG↗

arXiv2026

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↗

arXiv2026

Value-Decomposed Reinforcement Learning Framework for Taxiway Routing with Hierarchical Conflict-Aware Observations

Taxiway routing and on-surface conflict avoidance are coupled safety-critical decision problems in airport surface operations. Existing planning and optimization methods are often limited by online computational cost, while reinforcement learning methods may struggle to represent downstream traffic conflicts and balance multiple objectives. This paper presents Conflict-aware Taxiway Routing (CaTR), a reinforcement learning framework for real-time multi-aircraft taxiway routing. CaTR constructs a grid-based airport surface environment with action masking, introduces a hierarchical foresight traffic representation to encode current and downstream conflict-related traffic conditions, and adopts a value-decomposed reinforcement learning strategy to prioritize sparse but safety-critical objectives. Experiments are conducted on a realistic environment based on Changsha Huanghua International Airport under multiple traffic density levels. Results show that CaTR achieves better safety--efficiency trade-offs than representative planning, optimization, and reinforcement learning baselines while maintaining practical runtime.

cs.AI↗

arXiv2026

Design Your Ad: Personalized Advertising Image and Text Generation with Unified Autoregressive Models

Generating realistic and user-preferred advertisements is a key challenge in e-commerce. Existing approaches utilize multiple independent models driven by click-through-rate (CTR) to controllably create attractive image or text advertisements. However, their pipelines lack cross-modal perception and rely on CTR that only reflects average preferences. Therefore, we explore jointly generating personalized image-text advertisements from historical click behaviors. We first design a Unified Advertisement Generative model (Uni-AdGen) that employs a single autoregressive framework to produce both advertising images and texts. By incorporating a foreground perception module and instruction tuning, Uni-AdGen enhances the realism of the generated content. To further personalize advertisements, we equip Uni-AdGen with a coarse-to-fine preference understanding module that effectively captures user interests from noisy multimodal historical behaviors to drive personalized generation. Additionally, we construct the first large-scale Personalized Advertising image-text dataset (PAd1M) and introduce a Product Background Similarity (PBS) metric to facilitate training and evaluation. Extensive experiments show that our method outperforms baselines in general and personalized advertisement generation. Our project is available at https://github.com/JD-GenX/Uni-AdGen.

cs.CV↗

arXiv2026

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↗

arXiv2026

Multi-agent decision making: A Blackwell's informativeness approach

The rapid development of large language models (LLMs) has motivated research on decision-making in multi-agent systems, where multiple agents collaborate to achieve shared objectives. Existing aggregation approaches, such as voting and debate, are largely ad-hoc and lack formal guarantees regarding the informativeness of the resulting decisions. In this paper, we provide a principled approach to analyse decisions made in the multi-LLM setting using Blackwell's informativeness framework. Within the Blackwell information-structure abstraction, we show that voting and debate induce information structures that are no more informative than the pooled private information of all agents. This result identifies Bayesian pooled posterior maximisation as an information-theoretic upper-bound decision rule under the Blackwell ordering. Motivated by this theoretical analysis, we introduce a practical method for LLM-based question-answering (QA) tasks that estimates each agent's posterior and approximates the pooled posterior using a product-of-posteriors estimator. Extensive experiments on six QA benchmarks demonstrate that our approach outperforms state-of-the-art multi-LLM debate and voting methods.

cs.LG↗

arXiv2026

Beamforming Design for Pinching Antenna Systems Enabled Cognitive Radio Systems

A pinching antenna system (PASS) assisted cognitive radio (CR) system is proposed. A secondary system sum rate maximization problem is formulated by jointly considering the base station (BS) power budget, the pinching antenna (PA) deployment constraints, and the interference tolerance requirements of primary users. To address the resulting non-convex problem, a tractable reformulation based on the weighted minimum mean-square error (WMMSE) approach is adopted, followed by the development of an alternating optimization (AO) algorithm. Within this framework, the auxiliary variables are updated in closed form, enabling an efficient transformation of the digital beamforming subproblem to a convex form, while the PA deployment is refined through a tailored element-wise optimization strategy. Numerical results validate the effectiveness of the proposed design and show consistent performance gains compared with conventional benchmark schemes.

eess.SP↗

arXiv2026

NTIRE 2026 3D Restoration and Reconstruction in Real-world Adverse Conditions: RealX3D Challenge Results

This paper presents a comprehensive review of the NTIRE 2026 3D Restoration and Reconstruction (3DRR) Challenge, detailing the proposed methods and results. The challenge seeks to identify robust reconstruction pipelines that are robust under real-world adverse conditions, specifically extreme low-light and smoke-degraded environments, as captured by our RealX3D benchmark. A total of 279 participants registered for the competition, of whom 33 teams submitted valid results. We thoroughly evaluate the submitted approaches against state-of-the-art baselines, revealing significant progress in 3D reconstruction under adverse conditions. Our analysis highlights shared design principles among top-performing methods and provides insights into effective strategies for handling 3D scene degradation.

cs.CV↗

arXiv2026

Efficient primal-dual algorithm for imaging applications with matrix stacking, applied to DBT image reconstruction

The primal-dual hybrid gradient (PDHG) algorithm for solving convex optimization problems that arise in tomographic imaging is revisited. In particular, simplification of the selection of step-size parameters is developed for optimization problems with multiple terms, each containing a linear transform subject to splitting. This simplification maintains algorithm efficiency while avoiding massive grid searches for the optimal parameter settings. The PDHG framework is demonstrated on an image reconstruction problem for wide-angle digital breast tomosythesis (DBT); use of the proposed optimization problem is enabled by the framework and it is demonstrated to have some advantage in quantitative accuracy of the reconstructed volume and in improving DBT depth resolution.

math.OC↗

arXiv2026

KD-CVG: A Knowledge-Driven Approach for Creative Video Generation

Creative Generation (CG) leverages generative models to automatically produce advertising content that highlights product features, and it has been a significant focus of recent research. However, while CG has advanced considerably, most efforts have concentrated on generating advertising text and images, leaving Creative Video Generation (CVG) relatively underexplored. This gap is largely due to two major challenges faced by Text-to-Video (T2V) models: (a) \textbf{ambiguous semantic alignment}, where models struggle to accurately correlate product selling points with creative video content, and (b) \textbf{inadequate motion adaptability}, resulting in unrealistic movements and distortions. To address these challenges, we develop a comprehensive Advertising Creative Knowledge Base (ACKB) as a foundational resource and propose a knowledge-driven approach (KD-CVG) to overcome the knowledge limitations of existing models. KD-CVG consists of two primary modules: Semantic-Aware Retrieval (SAR) and Multimodal Knowledge Reference (MKR). SAR utilizes the semantic awareness of graph attention networks and reinforcement learning feedback to enhance the model's comprehension of the connections between selling points and creative videos. Building on this, MKR incorporates semantic and motion priors into the T2V model to address existing knowledge gaps. Extensive experiments have demonstrated KD-CVG's superior performance in achieving semantic alignment and motion adaptability, validating its effectiveness over other state-of-the-art methods. The code and dataset will be open source at https://kdcvg.github.io/KDCVG/.

cs.CV↗

arXiv2026

Amorphous Nanoconfinement Enables Self-sustaining Sabatier Reaction at Ambient Conditions

The Sabatier reaction, the catalytic hydrogenation of CO2 into CH4, offers a cornerstone for carbon capture and utilization, and in-situ resource utilization during space exploration; however, it faces a fundamental thermodynamic-kinetic paradox: although highly exothermic, conventional catalysts still require continuous external heating to activate CO2 and maintain stable operation. Here we report an amorphous silica-embedded ruthenium catalyst that enables a long-term self-sustaining autothermal Sabatier reaction dispensing with external energy supply. Operating under ambient conditions, this system achieves a record-high CH4 yield of 0.50 mol gcat-1 h-1 with 100% selectivity, stable operation for over 2,000 hours, and a record-low catalyst bed temperature down to 100 oC. This exceptional self-sustaining behavior stems from the synergistic effect of the catalyst's ultralow effective thermal conductivity (0.27 W m-1 K-1), induced by amorphous nanoconfinement, and its superior intrinsic activity. This synergy generates localized hot spots at Ru sites while suppressing macroscopic heat loss. In situ measurements further reveal CH4 formation even at 54 oC and identify a *CO-mediated pathway for CO2 methanation. The reaction ignites readily with a lighter or focused sunlight and persists even under forced convection from an electric fan, demonstrating strong environmental tolerance. By removing the need for constant energy input, this "ignite-and-forget" system paves the way for decentralized Power-to-Gas systems and autonomous fuel production in resource-constrained environments like Mars.

cond-mat.mtrl-sci↗

arXiv2026

Competing Constraints on Superconductivity in Thick FeSe films

Superconducting films emerge from the complex interplay of multiple growth parameters, making their optimization challenging. In iron-based superconductors, compressive strain is known to enhance the transition temperature (Tc) of FeSe films, yet reported Tc values vary widely even on identical substrates, indicating factors beyond strain are critical. Here, we develop a high-throughput off-center pulsed laser deposition strategy that transforms plume inhomogeneity into combinatorial FeSe film libraries with continuous gradients in lattice parameter, composition, and disorder. We discover that the maximum Tc does not coincide with the plume center but can shift off-center, revealing a competition between favorable c-axis expansion, stoichiometry, and defect scattering. Systematic characterization of 80 thick films (>50 nm), combined with interpretable machine learning, shows that besides the strong correlate of c-axis lattice parameter to Tc, the stoichiometry and disorder scattering impose critical constraints on the achievable transition temperature, defining a narrow optimization window rather than a simple monotonic relationship. This framework yields Tconset=17.1 K in thick FeSe films and establishes a general framework combining combinatorial synthesis with machine learning to uncover constrained optimization landscapes in complex functional materials.

cond-mat.supr-con↗

arXiv2026

SkipKV: Selective Skipping of KV Generation and Storage for Efficient Inference with Large Reasoning Models

Large reasoning models (LRMs) often incur significant key-value (KV) cache overhead, due to their linear growth with the verbose chain-of-thought (CoT) reasoning. This incurs both memory overhead and throughput bottlenecks, limiting efficient deployment. To reduce KV cache size during inference, we first investigate the effectiveness of existing KV cache eviction methods for CoT reasoning. Interestingly, we find that due to unstable token-wise scoring and reduced effective KV budget caused by padding, state-of-the-art (SoTA) eviction methods fail to maintain accuracy in multi-batch settings. Additionally, these methods often generate longer sequences than the original model without eviction, as semantic-unaware token-wise eviction leads to repeated revalidation during reasoning. To address these issues, we present \textbf{SkipKV}, a \textbf{\textit{training-free}} KV compression method that performs selective \textit{eviction} and \textit{generation}, operating at a coarse-grained, sentence-level sequence removal for efficient CoT reasoning. In specific, it introduces a \textit{sentence-scoring metric} to identify and remove highly similar sentences while maintaining semantic coherence. To suppress redundant generation, SkipKV dynamically adjusts a steering vector to update the hidden activation states during inference, enforcing the LRM to generate concise responses. Extensive evaluations on multiple reasoning benchmarks demonstrate that SkipKV achieves up to $\mathbf{26.7}\%$ higher accuracy compared to baseline methods, at a similar compression budget. Additionally, compared to SoTA, SkipKV yields up to $\mathbf{1.6}\times$ shorter generation length while improving throughput by up to $\mathbf{1.7}\times$. Our code is released at: \href{https://github.com/TTTTTTris/SkipKV}{https://github.com/TTTTTTris/SkipKV}.

cs.AI↗

arXiv2026

Efficient Adversarial Training via Criticality-Aware Fine-Tuning

Vision Transformer (ViT) models have achieved remarkable performance across various vision tasks, with scalability being a key advantage when applied to large datasets. This scalability enables ViT models to exhibit strong generalization capabilities. However, as the number of parameters increases, the robustness of ViT models to adversarial examples does not scale proportionally. Adversarial training (AT), one of the most effective methods for enhancing robustness, typically requires fine-tuning the entire model, leading to prohibitively high computational costs, especially for large ViT architectures. In this paper, we aim to robustly fine-tune only a small subset of parameters to achieve robustness comparable to standard AT. To accomplish this, we introduce Criticality-Aware Adversarial Training (CAAT), a novel method that adaptively allocates resources to the most robustness-critical parameters, fine-tuning only selected modules. Specifically, CAAT efficiently identifies parameters that contribute most to adversarial robustness. It then leverages parameter-efficient fine-tuning (PEFT) to robustly adjust weight matrices where the number of critical parameters exceeds a predefined threshold. CAAT exhibits favorable generalization when scaled to larger vision transformer architectures, potentially paving the way for adversarial training at scale, e.g, compared with plain adversarial training, CAAT incurs only a 4.3% decrease in adversarial robustness while tuning approximately 6% of its parameters. Extensive experiments on three widely used adversarial learning datasets demonstrate that CAAT outperforms state-of-the-art lightweight AT methods with fewer trainable parameters.

cs.CV↗

arXiv2026

PoTable: Towards Systematic Thinking via Plan-then-Execute Stage Reasoning on Tables

In recent years, table reasoning has garnered substantial research interest, particularly regarding its integration with Large Language Models (LLMs), which have revolutionized natural language applications. Existing LLM-based studies typically achieve step-by-step thinking for table reasoning guided by task semantics. While these approaches emphasize autonomous exploration and enhance fine-grained table understanding, they often overlook systematic thinking in the reasoning process. This oversight can lead to omitted steps, disorganized logic and misleading results, especially in complex scenarios. In this paper, we propose PoTable, a novel stage-oriented plan-then-execute approach that incorporates systematic thinking into table reasoning. Specifically, PoTable involves several distinct analytical stages with clear objectives to provide adequate guidance. To accomplish stage-specific goals, PoTable employs a plan-then-execute mechanism: it first plans the operation chain based on the stage objective, and then executes operations sequentially through code generation, real-time running and feedback processing. Consequently, PoTable produces reliable table reasoning results with highly accurate, step-wise commented and completely executable programs. It mirrors the workflow of a professional data analyst, offering advantages in both accuracy and explainability. Finally, we conduct extensive experiments on four datasets from the WikiTQ and TabFact benchmarks, where the results demonstrate the effectiveness, efficiency and explainability of PoTable. Our code is available at: https://github.com/Double680/PoTable.

cs.IR↗