arXiv subjects

Wei Cheng

Publications and source records attributed to Wei Cheng.

At least 37 records · Page 2Linked to original sources

arXiv2026

Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs

In this work, we investigate how Large Language Models (LLMs) adapt their internal representations when encountering inputs of increasing difficulty, quantified as the degree of out-of-distribution (OOD) shift. We reveal a consistent and quantifiable phenomenon: as task difficulty increases, whether through harder reasoning questions, longer contexts, or adding answer choices, the last hidden states of LLMs become substantially sparser. In short, \textbf{\textit{the farther the shift, the sparser the representations}}. This sparsity--difficulty relation is observable across diverse models and domains, suggesting that language models respond to unfamiliar or complex inputs by concentrating computation into specialized subspaces in the last hidden state. Through a series of controlled analyses with a learning dynamic explanation, we demonstrate that this sparsity is not incidental but an adaptive mechanism for stabilizing reasoning under OOD. Leveraging this insight, we design \textit{Sparsity-Guided Curriculum In-Context Learning (SG-ICL)}, a strategy that explicitly uses representation sparsity to schedule few-shot demonstrations, leading to considerable performance enhancements. Our study provides new mechanistic insights into how LLMs internalize OOD challenges. The source code is available at the URL: https://github.com/MingyuJ666/sparsityLLM.

cs.CL↗

arXiv2026

ConsistWorld: Evidence Routing for Consistent Multi-Agent World Models

Autoregressive video world models enable temporally coherent generation for a single observer. Extending them to multiple agents requires consistency across independently controlled views and temporal gaps under causal streaming. We present ConsistWorld, a multi-agent world model that generates camera-controlled video streams of a static scene from one shared image. We formulate consistency as routing evidence from committed multi-agent history and concurrently generated peer views to the tokens being generated. Pose Conditioned Memory Retrieval selects relevant historical observations from all agents, recovering evidence beyond the recent context window. Visibility-Gated Peer Sharing regulates current peer information according to estimated historical coverage and current-view overlap. Together, they determine which historical observations enter the context and where concurrent peer information contributes, supporting long-term recall and coordinated exploration. Both mechanisms use camera geometry and maintain a bounded active context for a fixed agent count and retrieval budget. Experiments on evidence sharing cases and video length and agent number generalizations show that ConsistWorld achieves a strong cross-time and cross-agent consistency while preserving competitive generation quality.

cs.CV↗

arXiv2026

WithEveryone: Unified Planning and Identity Grounding for Group Image Generation

Identity-preserving image generation becomes increasingly unreliable when a scene must contain many specified people. Beyond retaining each identity, the model must bind every reference to a distinct person and location, while training-time identity losses must establish correspondence among several noisy predicted faces. We introduce WithEveryone, a unified framework for generating group images up to ten reference identities. WithEveryone injects each selected identity as an addressed token, predicts a structured identity--layout plan, and renders the plan as a visual condition. Its key objective, Layout-Grounded ID Loss, uses annotated face regions to supervise the intended identities directly, avoiding unstable embedding-based face matching; ID Representation Forcing additionally trains a prediction for each identity before image synthesis. On an identity-disjoint benchmark, WithEveryone achieves the highest target-context identity similarity, improving face similarity from 0.462 for GPT-Image-2 to 0.499, while reducing copy-paste artifacts from 0.169 to 0.055. It further covers 97.3\% of the requested identities with a duplicate rate of only 2.8\%. These results show that explicit identity--layout grounding enables identity-preserving generation to scale to larger groups without relying on direct reference-face copying.

cs.CV↗

arXiv2026

Beyond the Survey: A Systematic Empirical Study of Detection and Association in Visual MOT

This paper presents a comprehensive experimental evaluation and detailed analysis of state-of-the-art multi-object tracking algorithms, with an emphasis on quantifying the individual contributions of detection and association components to overall tracking performance. Unlike existing surveys that primarily offer theoretical categorizations or taxonomies of tracking methods, our work adopts a rigorous experimental perspective grounded in publicly available implementations, providing practical guidance for researchers and practitioners in method selection and system design. We introduce a unified pipeline diagram that consolidates the core components across the two main branches of visual multi-object tracking: tracking-by-detection and end-to-end deep learning paradigms, and systematically analyze the object detection, feature extraction, and data association modules. Through extensive empirical studies on standard benchmarks, including MOT16, MOT17, MOT20, SportsMOT, DanceTrack, and CrowdTrack datasets, we reveal critical insights: (1) detection quality dominates association strategy performance, with detector improvements yielding more than 10% gains compared to less than 5% from refined association strategies; (2) modern deep learning detectors paired with specialized re-identification models significantly outperform joint detection and embedding approaches; and (3) transformer-based end-to-end methods exhibit greater robustness to detection quality variations but at a substantial computational cost. Our findings from extensive experiments provide key insights into component-level effects in MOT, particularly the dominant influence of detection quality relative to association, while offering practical insights for designing and optimizing MOT systems under varying performance and robustness requirements. Code and experimental setups are available at github.com/linh-gist/VisualMOT.

cs.CV↗

arXiv2026

Prior-Averaged Ranking of Low-Order Monomial Potentials in Low-Temperature Warm Inflation

We compare the relative prior-averaged weights of the monomial potentials $V_p(ϕ)=λ_pϕ^p/p$, with $p=2,3,4$, in low-temperature warm inflation with the dissipative coefficient fixed to $Υ=C_ϕT^3/ϕ^2$. The dissipative form, background equations, primordial-spectrum prescription, and compressed observable likelihood are held fixed while the monomial power is varied. For each branch, the warm background equations, including radiation backreaction, are solved numerically, and the broadened likelihood for $(A_s,n_s,r_{0.05})$ is integrated over the adopted prior domain to obtain $Z_{\rm eff}^{(A_s,n_s,r)}$. For the reference configuration $N_*=55$, $σ_r=0.005$, and the viability-conditioned prior A, we obtain $Δ\ln Z_{\rm eff}(p=2)=-32.18$ and $Δ\ln Z_{\rm eff}(p=3)=-6.99$ relative to $p=4$. The same ranking is retained for the specific variations of $N_*$, the prior domains, the random seeds, and the smoothing width of the upper-bound penalty examined here. A representative quartic trajectory gives $n_s=0.96420$, $r_{0.05}=0.02663$, $Q_*=4.68\times10^{-3}$, and $T_*/H_*=10.67$. Its scalar spectrum is evaluated using the assumed Bose--Einstein occupation prescription. A component test at this representative point indicates that the occupation term contributes more to the scalar enhancement than the fitted dissipative growth factor. Thus, within the compressed-likelihood setup, the adopted thermal-occupation prescription, and the adopted viability-conditioned priors, the prior-averaged ranking is $p=4>p=3\gg p=2$.

astro-ph.CO↗

arXiv2026

MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters

Time series forecasting (TSF) is evolving toward multimodal and agentic settings, yet using foundation models remains uneconomical in resource-constrained scenarios, where compact, specialized forecasters are more desirable. However, lightweight forecasters typically require substantial training data, limiting their use in domains with scarce, slowly accumulated, or privacy-sensitive time series. To address this dilemma, we investigate the challenging problem of few-shot learning for lightweight forecasters. We propose MetaCaster, a meta-harness-optimized multi-agent framework that uses agentic data generation to automatically train specialized lightweight forecasters from only a few examples and textual contexts. Our work highlights a new TSF paradigm in which agents act not as forecasters but as intermediary engineers that prepare efficient, task-specific forecasters for deployment. Experiments on 18 datasets, 23 state-of-the-art lightweight forecasters, and 14 baselines demonstrate that MetaCaster achieves both data efficiency and computational efficiency while maintaining high-quality TSF performance.

cs.LG↗

arXiv2026

On the cut locus of Hamilton--Jacobi equations I: structure and propagation via the touching approach

For a semiconcave function with linear modulus, we introduce the cut locus through a touching approach and prove that it coincides with the variational cut locus defined via the Lax--Oleinik semigroup, independently of the Hamiltonian. Cut points are characterized by the emptiness of the proximal subdifferential of $ϕ$, which we use as the analytic criterion behind the touching definition. We introduce the degree of regularity $R_ϕ$, which measures the $C^{1,1}$ deviation of $ϕ$, and prove the quantitative relation $R_ϕ(x)\sim 1/τ_{ϕ,H}(x)$ with the cut time function, valid pointwise up to explicit truncation constants. From this relation we derive a separation estimate for calibrated curves, showing that no conjugate points occur before the cut locus. This estimate supplies the tools for the propagation results. Cut points propagate globally along generalized characteristics for the evolutionary Hamilton--Jacobi equation, and Alexandrov points propagate forward along calibrated curves, with the second derivative satisfying a matrix Riccati equation. We also show that the cut locus is a Lebesgue null set and give a streamlined proof of Alexandrov's theorem. Finally, we give necessary and sufficient conditions for the cut locus of a weak KAM solution to be closed, in terms of the cut time function, the $C^{1,1}$ support, and the degree of regularity.

math.AP↗

arXiv2026

Aphanta: Diagnosing Task-Aligned Image-Edited Intermediates for Multimodal Reasoning

Explicit visual intermediates can help multimodal large language models (MLLMs) externalize spatial evidence and updated visual states, but their utility depends on whether an image editor can faithfully realize the required transformation. We introduce \textbf{Aphanta}, an automated task-discovery and closed-loop diagnostic framework for the MLLM -> image editor -> MLLM pipeline. Aphanta evaluates three conditions---direct reasoning, reasoning with an editor-generated intermediate, and reasoning with an idealized reference intermediate---to separate potential visual headroom from the practical utility of current editors. Across 20 candidate tasks and multiple editor--MLLM combinations, we find that utility is strongly task-conditioned. Gains concentrate in visual cue injection, grounding, and counterfactual state realization, whereas intermediates requiring symbol-sensitive construction or structural extrapolation are substantially less reliable. On the selected positive-task subset, our consolidated Qwen pipeline improves the mean task score from 0.343 to 0.445 ($+10.2$ points; $+29.7\%$ relative), while the full study also retains filtered and unsuccessful tasks to expose the boundary. These results position image editing as a specialized visual workspace rather than a universal reasoning mechanism, and establish Aphanta as a reusable protocol for measuring task--representation alignment, editor realization, and downstream pipeline utility.

cs.CV↗

arXiv2026

TopoAgent: A Structure-Aware Perception-to-Reasoning Framework for Diagram-to-Graph Topology Extraction with Large Vision-Language Models

Diagram-to-graph topology extraction aims to extract a graph of entities and their connections from a structural diagram. This task remains challenging for current vision-language models because it requires both fine-grained perceptual grounding and topology-aware reasoning with global consistency. We present TopoBench-180, a human-verified benchmark for diagram-to-graph topology extraction, and TopoAgent, a structure-aware perception-to-reasoning framework for reliable topology extraction using large vision-language models. TopoBench-180 contains 180 structural diagrams spanning Web-style and Network-style categories, paired with canonical graph annotations. TopoAgent progressively extracts the target graph by combining grounded perception, global structural priors, canonical node inventory construction, node-centric local-to-global relation reasoning, and topological consistency enforcement. Experiments on TopoBench-180 show that TopoAgent outperforms strong vision-language model baselines and recent visual reasoning frameworks, especially on edge extraction. More broadly, this work fills an important gap in multimodal structured understanding by establishing a benchmark and framework for diagram-to-graph topology extraction. The benchmark and associated resources will be publicly released at https://huggingface.co/datasets/WayneGuo0011/TopoBench-180.

cs.CV↗

arXiv2026

Towards A Unified Information Bottleneck Framework for Time Series Explanations

Explaining deep learning models operating on time series data is crucial in various applications that require transparent and interpretable insights into model behavior. {Existing explanation methods generally fall into two categories: attribution-based explanations, which identify the temporal regions most responsible for a prediction, and counterfactual explanations, which reveal how an input should be modified to alter the model's decision.} {Despite valuable insights, these two fields are largely studied independently. This disconnect leaves attribution methods lacking causal validation, while counterfactual methods suffer from severe instability, producing adversarial-like noise instead of meaningful explanations.} In this work, we revisit time-series explainability from an information-theoretic perspective and show that existing explainers are vulnerable to trivial solutions and distributional shifts. To address these limitations, we propose a unified objective function for explainable time series learning that bridges attribution and counterfactual reasoning within a single framework. Building upon the Information Bottleneck principle, our formulation explicitly prevents trivial explanations and out-of-distribution counterfactuals. {Based on this objective function, we introduce {\modelname}, a novel explanation framework that learns a parametric transformation network to construct explanation-embedded instances, where preserved information yields attribution explanations and controlled information removal produces stable counterfactual explanations.} We evaluate {\modelname} on synthetic and real-world benchmarks against state-of-the-art baselines. Extensive quantitative and qualitative results show that {\modelname} consistently outperforms competing methods, yielding faithful attributions and stable counterfactual explanations.

cs.LG↗

arXiv2026

Generalized Hamiltonian gradient flow of contact type I: regularity of the fundamental solutions

This paper studies generalized Hamiltonian gradient flows for contact-type Hamilton--Jacobi equations, adopting the variational framework of Herglotz's principle and its fundamental solution $h_L(t,x,y,u)$. Two main results are presented. First, precise first-order sensitivity relations are derived, linking derivatives of $h_L$ to dual arcs of minimizing trajectories. Second, quantitative second-order estimates show that $h_L$ is locally semiconcave and, over short time, semiconvex, indeed uniformly convex in certain variables. These regularity properties follow from a detailed variational analysis of minimizing trajectories. The work establishes a foundation for intrinsic methods in analyzing singularity propagation and generalized gradient flows in contact context, with implications for weak KAM theory, optimal transport, and regularity of viscosity solutions.

math.AP↗

arXiv2026

On the Effect of Sampling Diversity in Scaling LLM Inference

Large language model (LLM) scaling inference is key to unlocking greater performance, and leveraging diversity has proven an effective way to enhance it. Motivated by the observed relationship between solution accuracy and meaningful response diversity, we systematically study the effect of prompt diversity in scaling inference. We theoretically explain why diversified sampling improves Best-of-$N$ scaling, showing that responses generated from diverse prompts after Best-of-$N$ selection exhibit significantly lower error rates than those produced from stationary prompts. Building on this analysis, we derive a diversity-fidelity trade-off principle, that guides the design of sampling strategies introducing diversity. From this guidance, we instantiate a family of effective perturbation styles. We theoretically and empirically characterize \textbf{when} diversified exploration remains effective, demonstrating that it works under a variety of conditions, and we further show that under majority voting, diversity may vanish. Finally, we systematically evaluate the effectiveness of sampling diversity and show that, when applied appropriately in different contexts, meaningful perturbations yield stronger, task-dependent gains as diversity increases. Overall, this work provides a systematic analysis that offers a theoretical and empirical foundation for understanding how sampling diversity affects LLM inference-time scaling.

cs.LG↗

arXiv2026

Breaking Customized LLMs for Coding: Automated Red Teaming for Instruction Backdoor Attacks

LLM customization platforms allow users to build task-specific models for code intelligence tasks by embedding instructions into system prompts, without modifying the underlying model parameters. While these platforms lower the barrier to developing customized LLMs, they also introduce a new attack surface: instruction backdoor attacks, in which adversaries implant hidden malicious behaviors into customized instructions. However, existing attacks suffer from two key limitations. First, they often rely on explicit trigger patterns readily detected by platform-side or user-side inspection. Second, they require substantial manual effort to craft task-specific backdoored instructions, limiting their scalability. In this paper, we propose ARIA, an automated red-teaming framework for crafting covert and effective backdoored instructions against customized LLMs. ARIA leverages an attacker LLM to iteratively generate and refine backdoored instructions, guided by structured feedback from the target LLM along three dimensions: stealthiness, clean-task utility, and backdoor effectiveness. We evaluate ARIA on three code intelligence tasks, using four representative LLMs, and compare it with three baseline attacks. Experimental results show that ARIA achieves the highest attack success rate of 0.945, while maintaining the best clean-task utility across all tasks. ARIA also generalizes well across programming languages and remains robust to generation temperature. Furthermore, ARIA significantly outperforms existing attacks in evading platform-side and user-side detection, achieving a false negative rate of up to 1.000, and stays effective against existing defense methods, demonstrating its strong generalizability and robustness.

cs.CR↗

arXiv2026

iMontage: Unified, Versatile, Highly Dynamic Many-to-many Image Generation

Pre-trained video models learn powerful priors for generating high-quality, temporally coherent content. While these models excel at temporal coherence, their dynamics are often constrained by the continuous nature of their training data. We hypothesize that by injecting the rich and unconstrained content diversity from image data into this coherent temporal framework, we can generate image sets that feature both natural transitions and a far more expansive dynamic range. To this end, we introduce iMontage, a unified framework designed to repurpose a powerful video model into an all-in-one image generator. The framework consumes and produces variable-length image sets, unifying a wide array of image generation and editing tasks. To achieve this, we propose an elegant and minimally invasive adaptation strategy, complemented by a tailored data curation process and training paradigm. This approach allows the model to acquire broad image manipulation capabilities without corrupting its invaluable original motion priors. iMontage excels across several mainstream many-in-many-out tasks, not only maintaining strong cross-image contextual consistency but also generating scenes with extraordinary dynamics that surpass conventional scopes. Find our homepage at: https://kr1sjfu.github.io/iMontage-web/.

cs.CV↗

arXiv2026

MMPhysVideo: Physically Plausible Video Generation Through Joint RGB-Perception Modeling

Despite advancements in generating visually stunning content, video diffusion models (VDMs) often yield physically inconsistent results due to pixel-only reconstruction. To address this, we propose MMPhysVideo, the first study to enhance physical plausibility in video generation through joint multimodal modeling. We recast perceptual cues, specifically semantics, geometry, and spatio-temporal trajectories, into a unified pseudo-RGB format, enabling VDMs to directly capture complex physical dynamics. To mitigate cross-modal interference, we propose a Bidirectionally Controlled Teacher architecture, which utilizes parallel branches to fully decouple RGB and perception processing and adopts two zero-initialized control links to gradually establish pixel-wise consistency. For inference efficiency, the teacher's physical prior is distilled into a single-stream student model via representation alignment. Furthermore, we present MMPhysPipe, an end-to-end data curation and annotation pipeline tailored for constructing physics-rich multimodal datasets. MMPhysPipe employs a vision-language model (VLM) guided by a chain-of-visual-evidence rule to pinpoint physical subjects, enabling expert models to extract multi-granular perceptual information. Without additional inference costs, MMPhysVideo consistently improves physical plausibility of advanced models on the Videophy and PhyGenbench benchmarks and achieves superior performance among existing methods.

cs.CV↗

arXiv2026

Physics-structured cooperative neural network for baseline-free nuclear mass modeling

Machine learning approaches can improve nuclear mass modelling, but the most accurate strategies often depend on a theoretical mass baseline or hand-crafted physics features. We test whether a modular architecture encoding selected nuclear-structure priors improves baseline-free direct prediction and yields informative branch diagnostics. The Cooperative Neural Network (CoNN) implements this approach through four form-constrained branches: a smooth macroscopic network, discrete embeddings, a two-dimensional regional grid, and a parity-aware network. It extracts complementary patterns from (Z, N) through these branches and sums their outputs to predict binding energies without a theoretical mass-model baseline. Thus, the model retains physics priors while reducing its reliance on engineered input features. On AME2020, CoNN reaches a root-mean-square deviation (RMSD) of 0.269 MeV for 3558 nuclei, compared with 0.836 MeV for a parameter-matched unstructured MLP. It also gives RMSDs of 0.419 MeV on a held-out interpolation subset and 0.728 MeV on 122 nuclei newly measured since AME2016. The learned branch outputs show recognizable physical patterns, including embedding shell-kink signatures at major magic numbers and odd-even staggering along isotopic chains. These results identify architecture-level priors as a practical route to baseline-free mass prediction, with learned components that help diagnose both nuclear-structure patterns and extrapolation limits.

nucl-th↗

arXiv2026

Towards Structurally Explainable Machine-Generated Text Detection: A Graph-Perspective Framework

Despite the success of machine-generated text detectors, the black-box nature remains a critical limitation. Traditional explainability methods rely on token-level saliency, insufficient to reveal the high-order structural dependencies that distinguish LLM outputs. In this paper, we propose \textsc{LM$^2$otifs}, a principled framework that shifts detection from linear sequences to graph-structured manifolds. We first provide a theoretical grounding based on probabilistic graphical models, demonstrating that detection performance is more distinguishable in the graph-topological space. Driven by this theory, \textsc{LM$^2$otifs} transforms text into lexical co-occurrence graphs to preserve latent structural fingerprints. The framework employs Graph Neural Networks for robust detection and utilizes graph-specific explainers to extract interpretable motifs. Crucially, our experiments reveal that these structural motifs achieve higher faithfulness compared to traditional methods. This empirical evidence confirms the existence of high-order structural explanations that linear methods fail to capture. Experimental results show that \textsc{LM$^2$otifs} achieves state-of-the-art performance while providing multi-level \textit{distinct linguistic fingerprints} that are more faithful to the model's decision.

cs.CL↗

arXiv2026

Information Bottleneck Learning for Faithful Time Series Forecasting Explanations

As forecasts increasingly drive decisions in fields such as energy, transportation, and healthcare, understanding the historical data behind these predictions has become as crucial as the predictions themselves. Although existing interpretable-by-design forecasters reveal their internal structures, they offer no guarantee that these structures faithfully reflect the underlying evidence driving the predictions. In contrast, while faithfulness-oriented methods explicitly verify model behavior, they are almost exclusively designed for post-hoc classification tasks. To bridge this gap, we propose IB-Forecast, an inherently interpretable multivariate time-series forecasting framework. It decomposes forecasting into a learned periodic component and a residual component computed with explainable masks over input tokens. With a budget-constrained information bottleneck, end-to-end optimization enables users to directly control explanation sparsity. With a rigorous faithfulness evaluation protocol, extensive experiments demonstrate that IB-Forecast matches the forecasting error of leading black-box models while providing faithful explanations at no additional inference cost. Furthermore, under a matched sparsity budget, these native explanations consistently surpass gradient-based, occlusion-based, and optimization-based baselines across all evaluated datasets. Ultimately, whereas the native explanations of existing interpretable forecasters exhibit poor faithfulness, IB-Forecast guarantees high explanation fidelity, requiring only 14-20% of the observations to deliver low-error predictions.

cs.LG↗