arXiv subjects

Chen Liang

Publications and source records attributed to Chen Liang.

At least 37 records · Page 2Linked to original sources

arXiv2026

Turning Stale Gradients into Stable Gradients: Coherent Coordinate Descent with Implicit Landscape Smoothing for Lightweight Zeroth-Order Optimization

Zeroth-Order (ZO) optimization is pivotal for scenarios where backpropagation is unavailable, such as memory-constrained on-device learning and black-box optimization. However, existing methods face a stark trade-off: they are either sample-inefficient (e.g., standard finite differences) or suffer from high variance due to randomized estimation (e.g., random subspace methods). In this work, we propose Coherent Coordinate Descent (CoCD), a deterministic, sample-efficient, and budget-aware ZO optimizer. Theoretically, we formalize the notion of gradient coherence and demonstrate that CoCD is equivalent to Block Cyclic Coordinate Descent (BCCD) with ``warm starts,'' effectively converting historical (stale) gradients from a liability into a computational asset. This mechanism enables $O(1)$ query complexity per step while maintaining global descent directions. Furthermore, we derive error bounds revealing a counter-intuitive insight: larger finite-difference step sizes can induce an implicit smoothing effect on the optimization landscape by reducing the effective smoothness constant, thereby improving convergence stability. Experiments on MLP, CNN, and ResNet architectures (up to 270k parameters) demonstrate that CoCD significantly outperforms BCCD in terms of sample efficiency and convergence loss/accuracy, and exhibits superior stability over randomized ZO methods. Our results suggest that deterministic, structure-aware updates offer a superior alternative to randomization for lightweight ZO optimization.

cs.LG↗

arXiv2026

Training-Free Looped Transformers

We introduce training-free looped transformers, in which a lightweight inference-time wrapper loops a contiguous mid-stack block of layers of a frozen checkpoint without additional fine-tuning, continued training, or architectural changes. Unlike prior looped transformer methods that train with the looped structure end-to-end, we retrofit recurrence onto pretrained models at test time. We show that naive block reapplication usually degrades performance, highlighting the importance of the loop application strategy. Motivated by viewing a pre-norm transformer block as a forward Euler step on an ODE, we instead treat looping as a refinement of the same approximation, replacing one large update with smaller damped sub-steps. Across seven dense, sparse MoE, and MLA+MoE model families, our method improves Qwen3-4B-Instruct by +2.64 pp on MMLU-Pro, Qwen3-30B-A3B-Instruct by +1.14 pp on CommonsenseQA, and Moonlight-16B-A3B-Instruct by +1.20 pp on OpenBookQA.

cs.LG↗

arXiv2026

How Well Do Models Follow Visual Instructions? VIBE: A Systematic Benchmark for Visual Instruction-Driven Image Editing

Recent generative models have achieved remarkable progress in image editing. However, existing systems and benchmarks remain largely text-guided. In contrast, human communication is inherently multimodal, where visual instructions such as sketches efficiently convey spatial and structural intent. To address this gap, we introduce VIBE, the Visual Instruction Benchmark for Image Editing with a three-level interaction hierarchy that captures deictic grounding, morphological manipulation, and causal reasoning. Across these levels, we curate high-quality and diverse test cases that reflect progressively increasing complexity in visual instruction following. We further propose a robust LMM-as-a-judge evaluation framework with task-specific metrics to enable scalable and fine-grained assessment. Through a comprehensive evaluation of 17 representative open-source and proprietary image editing models, we find that proprietary models exhibit early-stage visual instruction-following capabilities and consistently outperform open-source models. However, performance degrades markedly with increasing task difficulty even for the strongest systems, highlighting promising directions for future research.

cs.CV↗

arXiv2026

Rethinking Generative Image Pretraining: How Far Are We From Scaling Up Next-Pixel Prediction?

This paper investigates the scaling properties of autoregressive next-pixel prediction, a simple, end-to-end yet under-explored framework for unified vision models. Starting with images at resolutions of 32x32, we train a family of Transformers using IsoFlops profiles across compute budgets up to 7e19 FLOPs and evaluate three distinct target metrics: next-pixel prediction objective, ImageNet classification accuracy, and generation-based completion measured by Fr'echet Distance. First, optimal scaling strategy is critically task-dependent. At a fixed resolution of 32x32 alone, the optimal scaling properties for image classification and image generation diverge, where generation optimal setup requires the data size grow three to five times faster than for the classification optimal setup. Second, as image resolution increases, the optimal scaling strategy indicates that the model size must grow much faster than data size. Surprisingly, by projecting our findings, we discover that the primary bottleneck is compute rather than the amount of training data. As compute continues to grow four to five times annually, we forecast the feasibility of pixel-by-pixel modeling of images within the next five years.

cs.CV↗

arXiv2026

$ϕ$-Balancing for Mixture-of-Experts Training

Mixture-of-Experts (MoE) models rely on balanced expert utilization to fully realize their scalability. However, existing load-balancing methods are largely heuristic and operate on noisy mini-batch assignment statistics, introducing bias relative to population-level objectives. We propose $ϕ$-balancing, a principled framework that directly targets population-level expert balance by minimizing a strictly convex, symmetric, and differentiable potential of the expected routing distribution. Using convex duality, we derive an equivalent min-max formulation and obtain a simple online algorithm via mirror descent, yielding an efficient EMA-based routing adjustment with negligible overhead. Across large-scale pretraining and downstream fine-tuning, $ϕ$-balancing consistently outperforms prior Switch-style and loss-free baselines, demonstrating more stable and effective expert utilization.

cs.LG↗

arXiv2026

The Efficiency Gap in Byte Modeling

Modern language models have historically relied on two dominant design choices: subword tokenization and autoregressive (AR) ordering. These design decisions bake in priors that dictate a model's learning. Recently, two alternative paradigms have challenged this: byte-level modeling, which bypasses static statistically-derived token vocabularies, and masked diffusion modeling (MDM), which conducts parallel, non-sequential generation. Their intersection represents a fully end-to-end modality-agnostic generative prototype; however, removing these structural priors incurs a significant computational cost. In this work, we investigate this cost through a compute-matched scaling study. Our results reveal that the performance penalty of byte modeling is not uniform; across scale, the scaling overhead of byte modeling is worse for MDM than for AR. We hypothesize that this disparity stems from context fragility: while AR's stable causal history allows models to naturally rediscover subword patterns, the MDM objective destroys the local contiguity required to efficiently resolve semantics from raw bytes. Our findings from controlled permutation experiments suggest that future modality-agnostic designs must incorporate alternative structural biases to maintain viable scaling trajectories in the byte regime.

cs.LG↗

arXiv2026

RealCam: Real-Time Novel-View Video Generation with Interactive Camera Control

Camera-controlled video-to-video (V2V) generation enables dynamic viewpoint synthesis from monocular footage, holding immense potential for interactive filmmaking and live broadcasting. However, existing implicit synthesis methods fundamentally rely on non-causal, full-sequence processing and rigid prefix-style temporal concatenation. This architectural paradigm mandates bidirectional attention, resulting in prohibitive computational latency, quadratic complexity scaling, and inherent incompatibility with real-time streaming or variable-length inputs. To overcome these limitations, we introduce \texttt{RealCam}, a novel autoregressive framework for interactive, real-time camera-controlled V2V generation. We first design a high-fidelity teacher model grounded in a \textbf{Cross-frame In-context Learning} paradigm. By interleaving source and target frames into synchronized contextual pairs, our design inherently enables length-agnostic generalization and naturally facilitates causal adaptation, breaking the rigid prefix bottleneck. We then distill this teacher into a few-step causal student via Self-Forcing with Distribution Matching Distillation, enabling efficient, on-the-fly streaming synthesis. Furthermore, to mitigate severe loop inconsistency in closed-loop trajectories, we propose \textbf{Loop-Closed Data Augmentation (LoopAug)}, a novel paradigm that synthesizes globally consistent loop sequences from existing multiview datasets. Extensive experiments demonstrate that \texttt{RealCam} achieves state-of-the-art visual fidelity and temporal consistency while enabling truly interactive camera control with orders-of-magnitude faster inference than existing paradigms. Our project page is at https://xyc-fly.github.io/RealCam/.

cs.CV↗

arXiv2026

DustNET: enabling machine learning and AI models of dusty plasmas

Dusty plasmas are ubiquitous throughout the universe, spanning laboratory and industrial plasmas, fusion devices, planetary environments, cometary comae, and interstellar media. Despite decades of research, many aspects of their behavior remain poorly understood within a unified framework. While numerous theoretical and numerical models describe specific phenomena, such as dust charging, transport, waves, and self-organization, fully predictive models across the wide range of spatial and temporal scales in both laboratory and natural systems remain elusive. Conventional plasma descriptions rely on coupled differential equations for particle densities, momenta, and energies, but their solutions are often limited by computational cost, numerical uncertainties, and incomplete knowledge of boundary conditions and transport processes. Recent advances in machine learning (ML), particularly deep neural networks, offer new opportunities to complement traditional physics-based modeling. Here we review ML and artificial intelligence (AI) approaches, termed bottom-up data-driven methods, for dusty plasma research. Central to this effort is Dust Neural nEtworks Technology (DustNET), a community-driven dataset initiative inspired by ImageNet, integrating experimental, simulation, and synthetic data to enable predictive modeling, uncertainty quantification, and multi-scale analysis. DustNET-trained models may also be deployed in real-time experimental settings under edge computing constraints. Combined with emerging multi-modal AI foundation models and autonomous agents, this framework provides a pathway toward a unified, physics-informed understanding of dusty plasmas across laboratory, industrial, space, and astrophysical environments.

physics.plasm-ph↗

arXiv2026

Closed Orbits and Descents for Enhanced Standard Representations of Classical Groups

Let $G=\mathrm{GL}_n(\mathbb{F})$, $\mathrm{O}_n(\mathbb{F})$, or $\mathrm{Sp}_{2n}(\mathbb{F})$ be one of the classical groups over an algebraically closed field $\mathbb{F}$ of characteristic $0$, let $\breve{G}$ be the MVW-extension of $G$, and let $\mathfrak{g}$ be the Lie algebra of $G$. In this paper, we classify the closed orbits in the enhanced standard representation $\mathfrak{g}\times E$ of $G$, where $E$ is the natural representation if $G=\mathrm{O}_n(\mathbb{F})$ or $\mathrm{Sp}_{2n}(\mathbb{F})$, and is the direct sum of the natural representation and its dual if $G=\mathrm{GL}_n(\mathbb{F})$. Additionally, for every closed $G$-orbit in $\mathfrak{g}\times E$, we prove that it is $\breve{G}$-stable, and determine explicitly the corresponding stabilizer group as well as the action on the normal space.

math.RT↗

arXiv2026

DiT as Real-Time Rerenderer: Streaming Video Stylization with Autoregressive Diffusion Transformer

Recent advances in video generation models has significantly accelerated video generation and related downstream tasks. Among these, video stylization holds important research value in areas such as immersive applications and artistic creation, attracting widespread attention. However, existing diffusion-based video stylization methods struggle to maintain stability and consistency when processing long videos, and their high computational cost and multi-step denoising make them difficult to apply in practical scenarios. In this work, we propose RTR-DiT (DiT as Real-Time Rerenderer), a steaming video stylization framework built upon Diffusion Transformer. We first fine-tune a bidirectional teacher model on a curated video stylization dataset, supporting both text-guided and reference-guided video stylization tasks, and subsequently distill it into a few-step autoregressive model via post-training with Self Forcing and Distribution Matching Distillation. Furthermore, we propose a reference-preserving KV cache update strategy that not only enables stable and consistent processing of long videos, but also supports real-time switching between text prompts and reference images. Experimental results show that RTR-DiT outperforms existing methods in both text-guided and reference-guided video stylization tasks, in terms of quantitative metrics and visual quality, and demonstrates excellent performance in real-time long video stylization and interactive style-switching applications.

cs.CV↗

arXiv2026

Shuffle the Context: RoPE-Perturbed Self-Distillation for Long-Context Adaptation

Large language models (LLMs) increasingly operate in settings that require reliable long-context understanding, such as retrieval-augmented generation and multi-document reasoning. A common strategy is to fine-tune pretrained short-context models at the target sequence length. However, we find that standard long-context adaptation can remain brittle: model accuracy depends strongly on the absolute placement of relevant evidence, exhibiting high positional variance even when controlling for task format and difficulty. We propose RoPE-Perturbed Self-Distillation, a training regularizer that improves positional robustness. The core idea is to form alternative "views" of the same training sequence by perturbing its RoPE indices -- effectively moving parts of the context to different positions -- and to train the model to produce consistent predictions across views via self-distillation. This encourages reliance on semantic signals instead of brittle position dependencies. Experiments on long-context adaptation of Llama-3-8B and Qwen-3-4B demonstrate consistent gains on long-context benchmarks, including up to 12.04% improvement on RULER-64K for Llama-3-8B and 2.71% on RULER-256K for Qwen-3-4B after SFT, alongside improved length extrapolation beyond the training context window.

cs.CL↗

arXiv2026

Global regularity and sharp decay rates to the 1D hypo-viscous compressible Navier-Stokes equations

In this paper, we study the global regularity and sharp decay rates for the isentropic hypo-viscous compressible Navier-Stokes equations in 1D. Firstly, we prove the global stability for the small initial data near a stable equilibrium. Especially, we establish the global critical regularity in the Sobolev space $H^β$ with $\frac{1}{2}<β<1$. Furthermore, by bootstrap argument, Fourier splitting method and energy method, we then establish the optimal time decay rates under the extra low-frequency smallness assumption. We find the $L^2$ energy is self-closed, which motivates us to obtain the existence of global large solutions for initial data with high regularity. By a pure energy method, we also derive the optimal time decay rates when $\frac{1}{2}\leβ<\frac{3}{4}$. We find a phenomenon that $\|(a,u)\|_{L^2}$ still decays even if the initial data does not possess $L^2$ smallness. Notably, the low-frequency smallness assumption is removed in the case with $\frac{1}{2}\leβ<\frac{3}{4}$.

math.AP↗

arXiv2026

Tap-to-Adapt: Learning User-Aligned Response Timing for Speech Agents

Response timing judgment is a critical component of interactive speech agents. Although there exists substantial prior work on turn modeling and voice wake-up, there is a lack of research on response timing judgments continuously aligned with user intent. To address this, we propose the Tap-to-Adapt framework, which enables users to naturally activate or interrupt the agent via tap interactions to construct online learning labels for response timing models. Under this framework, Dilated TCN and a sequential replay strategy play significant roles, as demonstrated through data-driven experiments and user studies. Additionally, we develop an evaluation and continuous data mining system tailored for the Tap-to-Adapt framework, through which we have collected approximately 20,000 samples from the user studies involving 20 participants.

cs.HC↗

arXiv2026

EMGauss: Continuous Slice-to-3D Reconstruction via Dynamic Gaussian Modeling in Volume Electron Microscopy

Volume electron microscopy (vEM) enables nanoscale 3D imaging of biological structures but remains constrained by acquisition trade-offs, leading to anisotropic volumes with limited axial resolution. Existing deep learning methods seek to restore isotropy by leveraging lateral priors, yet their assumptions break down for morphologically anisotropic structures. We present EMGauss, a general framework for 3D reconstruction from planar scanned 2D slices with applications in vEM, which circumvents the inherent limitations of isotropy-based approaches. Our key innovation is to reframe slice-to-3D reconstruction as a 3D dynamic scene rendering problem based on Gaussian splatting, where the progression of axial slices is modeled as the temporal evolution of 2D Gaussian point clouds. To enhance fidelity in data-sparse regimes, we incorporate a Teacher-Student bootstrapping mechanism that uses high-confidence predictions on unobserved slices as pseudo-supervisory signals. Compared with diffusion- and GAN-based reconstruction methods, EMGauss substantially improves interpolation quality, enables continuous slice synthesis, and eliminates the need for large-scale pretraining. Beyond vEM, it potentially provides a generalizable slice-to-3D solution across diverse imaging domains.

cs.CV↗

arXiv2026

Single-ion phonon laser in the quantum regime

The quantum phonon laser state is a vibrational state generated by phonon coherent amplification based on quantum mechanics. Its core is coherent excitation and manipulation of phonon quantum states by controlling phonon dynamics. This technology breaks classical limits of traditional phonon lasers, offering new methods for quantum information. Previous research on quantum phonon lasers focused on quantum van der Pol oscillators. As typical nonlinear quantum systems, they show significant value in trapped-ion systems. These breakthroughs extend nonlinear dynamics into the quantum domain and provide platforms for exploring quantum nonlinear phenomena. Although realized in two-ion systems, practical applications remain challenging. This paper explores how a single trapped ion generates quantum phonon laser states using a three-level model. By solving the quantum master equation numerically, steady-state characteristics are analyzed, focusing on quantum statistics including the Wigner function and second-order correlation function. An experimental scheme is proposed based on a single trapped 40Ca+ ion, using bichromatic blue-sideband and red-sideband lasers to generate quantum phonon laser states. By introducing the characteristic function of motional states, precise quantum state tomography is achieved. Additionally, a two-level model discusses the phonon laser threshold effect. However, the three-level model shows significantly different thresholds and more accurately describes the quantum phonon laser's physical mechanisms.

quant-ph↗

arXiv2026

The Market Consequences of Perceived Strategic Generosity: An Empirical Examination of NFT Charity Fundraisers

Crypto donations now represent a significant fraction of charitable giving worldwide. Nonfungible token (NFT) charity fundraisers, which involve the sale of NFTs of artistic works with the proceeds donated to philanthropic causes, have emerged as a novel development in this space. A unique aspect of NFT charity fundraisers is the significant potential for donors to reap financial gains from the rising value of purchased NFTs. Questions may arise about donors' motivations in these charity fundraisers, potentially resulting in a negative social image. NFT charity fundraisers thus offer a unique opportunity to understand the economic consequences of a donor's social image. We investigate these effects in the context of a large NFT charity fundraiser. We identify the causal effect of purchasing an NFT within the charity fundraiser on a donor's later market outcomes by leveraging random variation in transaction processing times on the blockchain. Further, we demonstrate a clear pattern of heterogeneity based on an individual's decision to relist (versus hold) the purchased charity NFTs (a sign of perceived strategic generosity) and based on an individual's social exposure within the NFT marketplace. We show that charity-NFT 're-listers' experience significant penalties in the market regarding the prices they can command for their other NFTs, particularly among those who are more socially exposed. Finally, we report the results of a scenario-based online experiment, which again support our findings, highlighting that the re-listing a charity NFT for sale at a profit leads others to perceive their initial donation as strategic generosity and reduces those others' willingness to purchase NFTs from the donor. Our study underscores the growing importance of digital visibility and traceability, features that characterize crypto-philanthropy, and online philanthropy more broadly.

econ.GN↗

arXiv2026

Cautious Weight Decay

We introduce Cautious Weight Decay (CWD), a one-line, optimizer-agnostic modification that applies weight decay only to parameter coordinates whose signs align with the optimizer update. Unlike standard decoupled decay, which implicitly optimizes a regularized or constrained objective, CWD preserves the original loss and admits a bilevel interpretation: it induces sliding-mode behavior upon reaching the stationary manifold, allowing it to search for locally Pareto-optimal stationary points of the unmodified objective. In practice, CWD is a drop-in change for optimizers such as AdamW, Lion, and Muon, requiring no new hyperparameters or additional tuning. For language model pre-training and ImageNet classification, CWD consistently improves final loss and accuracy at million- to billion-parameter scales.

cs.LG↗

arXiv2026

How to Disclose? Strategic AI Disclosure in Crowdfunding

As artificial intelligence (AI) increasingly integrates into crowdfunding practices, strategic disclosure of AI involvement has become critical. Yet, empirical insights into how different disclosure strategies influence investor decisions remain limited. Drawing on signaling theory and Aristotle's rhetorical framework, we examine how mandatory AI disclosure affects crowdfunding performance and how substantive signals (degree of AI involvement) and rhetorical signals (logos/explicitness, ethos/authenticity, pathos/emotional tone) moderate these effects. Leveraging Kickstarter's mandatory AI disclosure policy as a natural experiment and four supplementary online experiments, we find that mandatory AI disclosure significantly reduces crowdfunding performance: funds raised decline by 39.8% and backer counts by 23.9% for AI-involved projects. However, this adverse effect is systematically moderated by disclosure strategy. Greater AI involvement amplifies the negative effects of AI disclosure, while high authenticity and high explicitness mitigate them. Interestingly, excessive positive emotional tone (a strategy creators might intuitively adopt to counteract AI skepticism) backfires and exacerbates negative outcomes. Supplementary randomized experiments identify two underlying mechanisms: perceived creator competence and AI washing concerns. Substantive signals primarily affect competence judgments, whereas rhetorical signals operate through varied pathways: either mediator alone or both in sequence. These findings provide theoretical and practical insights for entrepreneurs, platforms, and policymakers strategically managing AI transparency in high-stakes investment contexts.

cs.HC↗