Searcharxiv⌕ Search

arXiv subjects

Xiaoqian Wang

Publications and source records attributed to Xiaoqian Wang.

At least 19 recordsLinked to original sources

Beyond Modality Preference: Controlling Multimodal Arbitration under Modality Conflict in MLLMs

Multimodal large language models (MLLMs) integrate information across modalities, but become unreliable when those modalities disagree. Under conflicting inputs, MLLMs may favor modalities that are not the most reliable for the task, thus motivating inference-time control of modality arbitration. We identify a small set of attention heads, termed arbitration heads, that causally influence whether predictions follow visual or textual evidence. Building on this, we introduce a lightweight, training-free intervention that biases these heads toward a designated trusted modality at inference time. The identified heads are stable across calibration set sizes and generalize across settings, and steering them enables effective bidirectional control of modality reliance across diverse benchmarks. Our findings identify an effective control mechanism to align MLLM predictions with task-specified modality priorities. The code is available at https://github.com/zhaoyingpan/modality-arbitration.

cs.CV↗

Guiding Data Allocation for Robust Subpopulation Generalization

Machine learning models often achieve high average accuracy in the training data while performing poorly on test data due to subpopulation shift. A common data-centric solution is to enforce balanced subgroup proportions, but full balance can be costly, infeasible, and not always necessary. In this work, we study whether balanced training data is the unique optimal configuration for robust subpopulation generalization. We extend a Gaussian feature-model framework to theoretically characterize the performance landscape across general training subgroup distributions. We show that multiple configurations, including substantially imbalanced ones, can perform comparably to the balanced one. We further derive gradient-based directions for adjusting subgroup proportions to improve adjusted accuracy and worst-group accuracy more efficiently. These findings suggest that strategically allocating data offers a more flexible and principled path to robust performance than simple balancing. The code is available at https://github.com/zhaoyingpan/subpop-data-allocation.

cs.LG↗

Online conformal inference for multi-step time series forecasting

We study the construction of distribution-free prediction intervals for multi-step time series forecasting. In this setting, forecast errors at different horizons are serially dependent, and feedback is delayed, since the error of an h-step-ahead forecast becomes available only after h periods. We propose autocorrelated multi-step conformal prediction (AcMCP), an online procedure that accommodates delayed feedback and explicitly incorporates the dependence among multi-step forecast errors into conformal calibration. To characterize this dependence, we derive first-order representations of the optimal forecast errors under a general nonstationary autoregressive data-generating process, with explicit bounds on remainder terms. AcMCP guarantees long-run coverage of at least the nominal level at every horizon, without requiring exchangeability or any assumption on how the data distribution evolves over time. We also establish a finite-sample bound on the empirical coverage shortfall, which depends on the forecast horizon and the evaluation sample size. Notably, these guarantees do not rely on the accuracy of the linear approximations. Simulations and real-data applications show that AcMCP adapts to time-varying dynamics and maintains rolling coverage close to the nominal level. By exploiting the dependence among multi-step errors, it often yields narrower intervals than its counterpart without this dependence, with similar coverage, particularly at longer horizons.

stat.ME↗

Rethinking SAE Evaluation: An Atomic Interpretable Unit Framework Reveals Hidden Polysemanticity and Redundancy

Sparse autoencoders (SAEs) were originally motivated by discovering monosemantic features, a property that characterizes individual features but does not specify what the learned feature set should look like as a whole. As SAEs have increasingly been adopted for attribution, circuit discovery, and intervention, researchers have begun evaluating metrics that describe global properties of the feature set, such as sparsity, dead feature rates, reconstruction fidelity, and disentanglement, a diverse set that extends far beyond monosemanticity. However, this implicit shift beyond monosemanticity has not been formalized as a clear, well-defined objective that comes before these evaluation metrics. Without such an objective, existing metrics remain ad hoc, with their necessity unjustified. In this work, we formalize this ultimate objective by introducing a new paradigm for SAE evaluation: requiring that SAE features, and more generally dictionary-based decompositions of model representations, form an atomic interpretable unit basis, a coordinate system for semantically meaningful explanation. Under this paradigm, we organize existing and new measurements such as DAME (Dominance-Aware Monosemanticity Evaluation) as proxy metrics and introduce a comprehensive benchmarking framework. Evaluating hundreds of SAEs across multiple architectures and model families, we uncover failure modes in which some SAEs appear monosemantic under existing evaluation metrics while remaining functionally polysemantic: the most dominant features, those contributing most to model computation, are systematically less interpretable and more redundant. These failure modes warrant attention from SAE researchers as targets for improved training and architectures, and motivate dominance-aware evaluation as a standard check for practitioners before SAE features are deployed for attribution, circuit tracing, or intervention.

cs.LG↗

conformalForecast: An R package for sequential conformal prediction in time series

Reliable uncertainty quantification is central to modern forecasting. Conformal prediction offers model-agnostic intervals with coverage guarantee under exchangeability, but existing software often targets general prediction problems or individual time series procedures rather than integrated workflows for multistep forecasting applications. We introduce conformalForecast, an open-source R package for constructing, updating, and evaluating conformal prediction intervals for multistep time series forecasts. The package separates point forecasting from calibration, accepts forecasts from diverse sources through a common interface, and preserves horizon-specific forecast errors and calibration states. It implements several conformal calibration schemes spanning empirical-quantile, feedback-based, and dependence-aware approaches, together with tools for assessing coverage, interval width, and interval scores over time and across horizons. Fitted objects can be extended as new observations arrive while retaining previously issued intervals and avoiding recomputation at earlier forecast origins. An application to short-term electricity price forecasting across five Australian regions illustrates the workflow. Aggregate coverage is generally close to the nominal level, while rolling coverage and interval efficiency vary across methods and horizons. In the reported experiment, incremental updating reproduces full-reconstruction intervals under identical forecasting and calibration settings while substantially reducing computation. conformalForecast provides a reproducible and extensible framework linking conformal methodology with sequential forecasting practice.

stat.CO↗

Can Language Models Learn to Reject Their Own Bad Reasoning Steps?

Verifier-guided decoding can prevent harmful reasoning steps from contaminating subsequent generation, but typically relies on an external learned verifier. We ask whether a language model can instead reject its own bad reasoning steps. We define a prefix's recoverability as the probability that the frozen generator can complete it correctly. Diagnostics show that adjacent recoverability changes are often difficult to resolve with practical Monte Carlo budgets, while same-prefix candidates exhibit a sparse low-recoverability tail. We introduce Self-Step Rejection (SSR), which trains a lightweight LoRA acceptance gate on the generator backbone while keeping the base model frozen. SSR uses confidence-qualified first-passage supervision: steps before the first resolved crossing of a root-relative recoverability barrier are accepted, the crossing step is rejected, and unresolved steps and suffixes are excluded. Training combines pointwise classification, same-prefix pairwise learning, and group-relative policy refinement using final-answer correctness. At inference, SSR accepts candidates or resamples from the unchanged prefix under rejection budgets, without an external learned verifier. Across three reasoning models and five mathematical reasoning benchmarks, SSR improves macro-average accuracy over single-pass decoding by 5.4--10.1 points using 1.21--1.40x as many generated tokens, and achieves the highest macro-average accuracy among evaluated step-level methods. Full-solution scaling methods require 4.47--8.27x the single-pass token cost for comparable performance.

cs.CL↗

Efficient conformal prediction intervals for time series: Online PID-Expert aggregation

For a given point forecaster, proportional-integral-derivative (PID) calibration configurations can attain similar overall coverage yet produce different interval widths. We introduce PID-Expert, an online aggregation method for improving the efficiency of conformal prediction intervals for time series. PID-Expert combines thresholds from a fixed library of PID calibrators, each evolving under its own coverage feedback. Aggregation weights depend on normalized interval width and miscoverage, while a shared multiplier adapts the miscoverage penalty using feedback from the reported interval. We establish local regret bounds for weighted expert losses under the realized multiplier sequence and, separately, a pathwise upper bound on time-averaged aggregate miscoverage, with control in expectation and almost surely under a stability condition. Across four simulation settings and two real-data applications, PID-Expert produces narrower mean intervals than a prespecified Conformal PID benchmark while keeping overall empirical coverage close to the nominal level. Compared with expert selection and equal averaging, aggregation generally provides a more balanced coverage--width trade-off across forecasting settings. Rolling analyses further reveal local coverage--efficiency trade-offs, particularly following abrupt distributional shifts. Overall, PID-Expert reduces reliance on a single PID configuration while improving interval efficiency.

stat.ME↗

Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems

Heterogeneous multi-agent systems combine models with different capabilities through a common communication interface. Exchanging internal states directly requires translating between model-specific representations and controlling intermediate computation. We introduce the Vision Wormhole, which repurposes the visual input interface of Vision-Language Models (VLMs) for continuous communication between frozen heterogeneous agents. A Universal Visual Codec encodes each sender's latent rollout into a fixed-size message, maps it through a shared reference space, and decodes received messages into the receiver's image-token span. Per-model codecs and affine reference maps form a hub-and-spoke architecture with $O(N)$ components for $N$ models. Each model learns its codec independently through self-distillation on anchor texts, and shared-anchor alignment enables reuse across communication partners. Across four VLM families, six team configurations, and nine reasoning benchmarks, Vision Wormhole improves accuracy by 6.0 percentage points on average over text-mediated MAS and achieves a 1.69$\times$ geometric-mean speedup in batch-normalized end-to-end runtime.

cs.CL↗

Do Proactive Agents Need an LLM to Decide When to Act?

Proactive assistants continuously decide when to intervene and what context should support the intervention. Large language model (LLM) pipelines repeatedly interpret activity histories to make these decisions, paying an inference cost even when the assistant remains silent. We show that a small graph model can handle both decisions and improve the language agents it controls. Our key insight is that user activity has a native graph structure: events involve persistent entities whose recurrence connects interactions over time. Triggering and context selection map directly to predictions on event and entity nodes. We introduce a temporal-graph-learning (TGL) controller that learns these predictions jointly and supplies both outputs in one forward pass. The downstream language agent generates suggestions on triggered events using the activity history and scored entities. At 11.13 ms per event on a GPU server, TGL achieves the highest AUCs among nine trigger architectures and gives approximately $4$--$7\times$ trigger-stage speedups over the two single-forward LLM triggers. A shared TGL model improves F1 across all 14 downstream backbones by a mean of 16.7 points. The controller also runs at 13.99 ms on a consumer laptop with an approximately 220 MiB BF16 resident footprint, bringing effective proactive control to on-device deployment.

cs.CL↗

Uncertainty-Aware Hybrid Retrieval for Long-Document RAG

Retrieval augmented generation (RAG) depends critically on the quality and granularity of retrieved evidence. Large retrieval units preserve context but often introduce irrelevant content, which can dilute answer bearing evidence and worsen long context utilization. Fine-grained units are more compact, but they may be difficult to retrieve reliably because short chunks can lack semantic, lexical, or bridging cues needed to match the query. We propose Uncertainty-aware Multi-Granularity RAG (UMG-RAG), a training-free hybrid retrieval framework that treats chunk granularity as query-specific reliability estimation. Instead of training a new retriever or modifying the generator, UMG-RAG uses existing dense and sparse retrievers as complementary experts across multiple chunk granularities. For each query, it converts each expert-granularity score list into an evidence distribution, estimates reliability from distribution entropy, and fuses candidates according to query-specific semantic, lexical, and granularity confidence. We further introduce UMGP-RAG, a parent promotion variant that uses fine-grained hits to locate relevant evidence while returning broader non-redundant parent chunks for local coherence. Experiments on question answering benchmarks show that uncertainty-aware fusion and parent promotion improve generation quality while maintaining a lightweight, plug-and-play retrieval pipeline.

cs.AI↗

When the Same Coefficients Reach Different Places: Asymmetric Realizability in Transplanting Tokenizers across Large Language Models

Tokenizer transplant in cross-vocabulary model composition reconstructs donor-only embedding rows as weighted combinations over shared lexical anchors and reuses those coefficients on the base. We identify a structural geometric property of this reconstruction: the same coefficient vector reaches different sets in the donor and base anchor spans, an \emph{asymmetric realizability} gap. Across 65 donor-base pairs under OMP, with cross-operator validation on CLP, WECHSEL, and FOCUS, we construct \textit{breaker tokens}: single coefficient vectors that remain statistically inert in the donor anchor span while producing a high-salience reconstruction in the base. The same Gemma-2-2B donor checkpoint admits this construction against 13 different downstream bases drawn from five model families. The planted direction passes weight-merging with a clean reference unchanged. In a deployer case study, standard LoRA fine-tuning suppresses the breaker primarily on prompts whose distribution matches the training corpus and is not a sufficient mitigation against this attack family in our setting. The tested spectral filters miss the asymmetry. We discuss potential misuse in the open-weight composition supply chain.

cs.LG↗

The Cost of Context: Mitigating Textual Bias in Multimodal Retrieval-Augmented Generation

While Multimodal Large Language Models (MLLMs) are increasingly integrated with Retrieval-Augmented Generation (RAG) to mitigate hallucinations, the introduction of external documents can conceal severe failure modes at the instance level. We identify and formalize the phenomenon of recorruption, where the introduction of even perfectly accurate "oracle" context causes a capable model to abandon an initially correct prediction. Through a mechanistic diagnosis of internal attention matrices, we show that recorruption is driven by a two-fold attentional collapse: (1) visual blindness, characterized by the systemic suppression of visual attention mass ($M_{vis}$) and sharpness ($S_{vis}$), and (2) a structural positional bias that forces the model to prioritize boundary tokens over semantic relevance. Our analysis reveals an Illusion of Success, demonstrating that many seemingly correct RAG outcomes are merely positional coincidences where the model's textual copying bias happens to align with the ground-truth location. To address these vulnerabilities, we propose Bottleneck Attention Intervention for Recovery (BAIR), a parameter-free, inference-time framework that restores visual saliency and applies position-aware penalties to textual distractors. Across medical factuality, social fairness, and geospatial benchmarks, BAIR successfully restores multimodal grounding and improves diagnostic reliability without requiring model retraining or fine-tuning.

cs.CL↗

On the Effects of Adversarial Perturbations on Distribution Robustness

Adversarial robustness refers to a model's ability to resist perturbation of inputs, while distribution robustness evaluates the performance of the model under data shifts. Although both aim to ensure reliable performance, prior work has revealed a tradeoff in distribution and adversarial robustness. Specifically, adversarial training might increase reliance on spurious features, which can harm distribution robustness, especially the performance on some underrepresented subgroups. We present a theoretical analysis of adversarial and distribution robustness that provides a tractable surrogate for per-step adversarial training by studying models trained on perturbed data. In addition to the tradeoff, our work further identified a nuanced phenomenon that $\ell_\infty$ perturbations on data with moderate bias can yield an increase in distribution robustness. Moreover, the gain in distribution robustness remains on highly skewed data when simplicity bias induces reliance on the core feature, characterized as greater feature separability. Our theoretical analysis extends the understanding of the tradeoff by highlighting the interplay of the tradeoff and the feature separability. Despite the tradeoff that persists in many cases, overlooking the role of feature separability may lead to misleading conclusions about robustness.

cs.LG↗

Distributed Online Economic Dispatch with Time-Varying Coupled Inequality Constraints

We investigate the distributed online economic dispatch problem for power systems with time-varying coupled inequality constraints. The problem is formulated as a distributed online optimization problem in a multi-agent system. At each time step, each agent only observes its own instantaneous objective function and local inequality constraints; agents make decisions online and cooperate to minimize the sum of the time-varying objectives while satisfying the global coupled constraints. To solve the problem, we propose an algorithm based on the primal-dual approach combined with constraint-tracking. Under appropriate assumptions that the objective and constraint functions are convex, their gradients are uniformly bounded, and the path length of the optimal solution sequence grows sublinearly, we analyze theoretical properties of the proposed algorithm and prove that both the dynamic regret and the constraint violation are sublinear with time horizon T. Finally, we evaluate the proposed algorithm on a time-varying economic dispatch problem in power systems using both synthetic data and Australian Energy Market data. The results demonstrate that the proposed algorithm performs effectively in terms of tracking performance, constraint satisfaction, and adaptation to time-varying disturbances, thereby providing a practical and theoretically well-supported solution for real-time distributed economic dispatch.

math.OC↗

SUV: Scalable Large Language Model Copyright Compliance with Regularized Selective Unlearning

Large Language Models (LLMs) have transformed natural language processing by learning from massive datasets, yet this rapid progress has also drawn legal scrutiny, as the ability to unintentionally generate copyrighted content has already prompted several prominent lawsuits. In this work, we introduce SUV (Selective Unlearning for Verbatim data), a selective unlearning framework designed to prevent LLM from memorizing copyrighted content while preserving its overall utility. In detail, the proposed method constructs a dataset that captures instances of copyrighted infringement cases by the targeted LLM. With the dataset, we unlearn the content from the LLM by means of Direct Preference Optimization (DPO), which replaces the verbatim copyrighted content with plausible and coherent alternatives. Since DPO may hinder the LLM's performance in other unrelated tasks, we integrate gradient projection and Fisher information regularization to mitigate the degradation. We validate our approach using a large-scale dataset of 500 famous books (predominantly copyrighted works) and demonstrate that SUV significantly reduces verbatim memorization with negligible impact on the performance on unrelated tasks. Extensive experiments on both our dataset and public benchmarks confirm the scalability and efficacy of our approach, offering a promising solution for mitigating copyright risks in real-world LLM applications.

cs.CL↗

Inland Waterway Object Detection in Multi-environment: Dataset and Approach

The success of deep learning in intelligent ship visual perception relies heavily on rich image data. However, dedicated datasets for inland waterway vessels remain scarce, limiting the adaptability of visual perception systems in complex environments. Inland waterways, characterized by narrow channels, variable weather, and urban interference, pose significant challenges to object detection systems based on existing datasets. To address these issues, this paper introduces the Multi-environment Inland Waterway Vessel Dataset (MEIWVD), comprising 32,478 high-quality images from diverse scenarios, including sunny, rainy, foggy, and artificial lighting conditions. MEIWVD covers common vessel types in the Yangtze River Basin, emphasizing diversity, sample independence, environmental complexity, and multi-scale characteristics, making it a robust benchmark for vessel detection. Leveraging MEIWVD, this paper proposes a scene-guided image enhancement module to improve water surface images based on environmental conditions adaptively. Additionally, a parameter-limited dilated convolution enhances the representation of vessel features, while a multi-scale dilated residual fusion method integrates multi-scale features for better detection. Experiments show that MEIWVD provides a more rigorous benchmark for object detection algorithms, and the proposed methods significantly improve detector performance, especially in complex multi-environment scenarios.

cs.CV↗

A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks

Recent advancements in Vision-Language Models (VLMs) have enabled complex multimodal tasks by processing text and image data simultaneously, significantly enhancing the field of artificial intelligence. However, these models often exhibit biases that can skew outputs towards societal stereotypes, thus necessitating debiasing strategies. Existing debiasing methods focus narrowly on specific modalities or tasks, and require extensive retraining. To address these limitations, this paper introduces Selective Feature Imputation for Debiasing (SFID), a novel methodology that integrates feature pruning and low confidence imputation (LCI) to effectively reduce biases in VLMs. SFID is versatile, maintaining the semantic integrity of outputs and costly effective by eliminating the need for retraining. Our experimental results demonstrate SFID's effectiveness across various VLMs tasks including zero-shot classification, text-to-image retrieval, image captioning, and text-to-image generation, by significantly reducing gender biases without compromising performance. This approach not only enhances the fairness of VLMs applications but also preserves their efficiency and utility across diverse scenarios.

cs.CV↗

A liquid crystal geometric phase hyperbolic lens with a positive focal length for arbitrary circularly polarized incidence

This paper presents the design and experimental validation of a liquid crystal geometric phase hyperbolic lens (LCHL) with positive focal lengths for arbitrary circularly polarized light. Utilizing Pancharatnam-Berry phase modulation, the lens enables isotropic focusing for both left- and right-handed circular polarization. We fabricated and characterized two lenses with different focal lengths, and their diffraction patterns were analyzed experimentally. The results exhibit strong agreement with theoretical simulations, highlighting the lens's capability for precise optical field modulation. The proposed LCHL demonstrates significant potential for detecting weak polarization states, making it a promising tool for advanced applications in biomedical imaging, optical vortex generation, and multifocal optical systems.

physics.optics↗