Searcharxiv⌕ Search

arXiv subjects

Wanyu Zhang

Publications and source records attributed to Wanyu Zhang.

12 recordsLinked to original sources

TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual Modulation

Large language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.

cs.CR↗

POEM: Output-Relative Block-Sparse Attention for Long-Context Prefilling

Large language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.

cs.CR↗

Calibrated Transfer Learning for Survival Prediction and Treatment Inference

Historical survival data can improve prediction in a cohort with few events, but prognostic differences can make borrowing harmful. To borrow while adapting, we propose calibrated transfer learning for the Cox model, in which a historical risk function enters the target analysis as a fixed offset with a prespecified correction for changing covariate effects and a separate baseline hazard. Calibration removes historical error in the correction space exactly, so only the remaining error governs the prediction rate, for any bounded learner. Accurate prediction does not, however, ensure valid treatment inference. In randomized trials, the remaining error biases the conditional treatment coefficient, because a treatment effect changes the prognostic composition of later risk sets. We characterize this bias and its sharp worst-case direction, and we introduce a cross-fitted projected score that removes it at first order and leaves only a second-order bias when a fixed error lies in its dictionary. A historical learning bound separates approximation, sampling, and optimization error. In simulations, calibration improves prediction when the historical risk is well learned, and the projected score restores interval coverage under drift outside the correction, also at a hazard ratio of 0.7. In a retrospective transfer between the POPLAR and OAK lung cancer trials, selected corrections reduce the integrated Brier score by 4.3\% relative to ridge-penalized clinical Cox with 100 training patients, without improving treatment precision.

stat.ME↗

Adaptive Sampling of Costly Outcomes in Randomized Clinical Trials

In some randomized trials the primary outcome is costly or slow to measure, while auxiliary variables that predict it are available for everyone. The outcome can then be measured in a probability sample. We study an adaptive design in which the statistician who selects outcomes to measure is blinded to treatment assignment. Outcomes from an initial random sample are used to fit pooled models for the outcome and its residual variance. These set the sampling probabilities for the remaining participants and may be refitted as outcomes accumulate. After unblinding, arm means are estimated by augmented inverse probability weighting. The estimator is unbiased for the complete-data treatment difference for any working models, and a martingale central limit theorem gives Wald intervals under repeated updating. We bound the variance lost by estimating the sampling rule. The bound is linear in the error of the fitted residual variance, quadratic when the optimal probabilities are not truncated, and relates the initial sample size to the learning rate of the models. Relative to designs using treatment assignment, the blinded design loses a term due to unequal residual variances in the two arms and a term due to the conditional treatment effect, which is second order near the null. In simulations, coverage was near nominal. Adaptive sampling was 12% to 34% more efficient than simple random sampling and needed 10% to 26% fewer measured outcomes for the same precision. In a resampling study of an antifungal trial, adaptive sampling reduced the sampling variance by 39%.

stat.ME↗

Stochastic Gradient Methods with Online Scaling

This paper introduces Stochastic Online Scaled Gradient Methods (SOSGM), a generalization of the recently developed adaptive preconditioning framework in arXiv:2505.23081 and arXiv:2509.11007 to stochastic optimization. Under standard assumptions, we establish convergence guarantees for SOSGM using large batchsize or variance reduction. SOSGM is compatible with popular diagonal and/or low-rank preconditioners as well as heavy-ball momentum, while maintaining memory and computation cost comparable to Adam. Extensive numerical experiments demonstrate the strong empirical performance of SOSGM. Using a diagonal preconditioner, SOSGM and its variants substantially outperform existing adaptive first-order methods across a range of statistical learning tasks.

math.OC↗

Operator Splitting Methods with Online Scaling

This paper develops a principled framework for automatically tuning preconditioners in operator splitting methods, including forward-backward splitting, Douglas-Rachford splitting, and the alternating direction method of multipliers. We interpret these splitting methods as preconditioned gradient descent applied to their corresponding envelope functions and use online learning to adaptively learn the preconditioner during optimization. Our framework achieves asymptotically faster convergence rates and accelerates the solution of important optimization problems, including Lasso, $\ell_1$-regularized logistic regression, and quadratic programming. Extensive numerical experiments validate our theoretical results and demonstrate the practical effectiveness of the proposed approach.

math.OC↗

Enhanced PDHG for Linear Programming with Online Preconditioning

We present an online preconditioning technique for the primal-dual hybrid gradient (PDHG) algorithm for linear programming (LP). The method adaptively updates primal and dual preconditioners using an online optimization framework. To improve its practical performance, we introduce several algorithmic enhancements, including using normalized online loss functions and updating preconditioners infrequently. We implement the technique on top of vanilla PDHG and the GPU-based LP solver cuPDLP.jl, and benchmark its performance on standard LP datasets. Our numerical experiments demonstrate that online preconditioning effectively reduces both iteration counts and overall solving time.

math.OC↗

Data-driven Mixed Integer Optimization through Probabilistic Multi-variable Branching

In this paper, we propose a Pre-trained Mixed Integer Optimization framework (PreMIO) that accelerates online mixed integer program (MIP) solving with offline datasets and machine learning models. Our method is based on a data-driven multi-variable cardinality branching procedure that splits the MIP feasible region using hyperplanes chosen by the concentration inequalities. Unlike most previous ML+MIP approaches that either require complicated implementation or suffer from a lack of theoretical justification, our method is simple, flexible, provable, and explainable. Numerical experiments on both classical OR benchmark datasets and real-life instances validate the efficiency of our proposed method.

math.OC↗

Orientation Determination of Cryo-EM Images Using Block Stochastic Riemannian Subgradient Methods

The determination of molecular orientations is crucial for the three-dimensional reconstruction of Cryo-EM images. Traditionally addressed using the common-line method, this challenge is reformulated as a self-consistency error minimization problem constrained to rotation groups. In this paper, we consider the least-squared deviation (LUD) formulation and employ a Riemannian subgradient method to effectively solve the orientation determination problem. To enhance computational efficiency, a block stochastic version of the method is proposed, and its convergence properties are rigorously established. Extensive numerical evaluations reveal that our method not only achieves accuracy comparable to that of state-of-the-art methods but also delivers an average 20-fold speedup. Additionally, we implement a modified formulation and algorithm specifically designed to address scenarios characterized by very low SNR.

math.OC↗

A Robust Anchor-based Method for Multi-Camera Pedestrian Localization

This paper addresses the problem of vision-based pedestrian localization, which estimates a pedestrian's location using images and camera parameters. In practice, however, calibrated camera parameters often deviate from the ground truth, leading to inaccuracies in localization. To address this issue, we propose an anchor-based method that leverages fixed-position anchors to reduce the impact of camera parameter errors. We provide a theoretical analysis that demonstrates the robustness of our approach. Experiments conducted on simulated, real-world, and public datasets show that our method significantly improves localization accuracy and remains resilient to noise in camera parameters, compared to methods without anchors.

cs.CV↗

Restarted Primal-Dual Hybrid Conjugate Gradient Method for Large-Scale Quadratic Programming

Convex quadratic programming (QP) is an essential class of optimization problems with broad applications across various fields. Traditional QP solvers, typically based on simplex or barrier methods, face significant scalability challenges. In response to these limitations, recent research has shifted towards matrix-free first-order methods to enhance scalability in QP. Among these, the restarted accelerated primal-dual hybrid gradient (rAPDHG) method, proposed by Lu, has gained notable attention due to its linear convergence rate to an optimal solution and its straightforward implementation on Graphics Processing Units (GPUs). Building on this framework, this paper introduces a restarted primal-dual hybrid conjugate gradient (PDHCG) method, which incorporates conjugate gradient (CG) techniques to address the primal subproblems inexactly. We demonstrate that PDHCG maintains a linear convergence rate with an improved convergence constant and is also straightforward to implement on GPUs. Extensive numerical experiments on both synthetic and real-world datasets demonstrate that our method significantly reduces the number of iterations required to achieve the desired accuracy compared to rAPDHG. Additionally, the GPU implementation of our method achieves state-of-the-art performance on large-scale problems. In most large-scale scenarios, our method is approximately 5 times faster than rAPDHG and about 100 times faster than other existing methods. These results highlight the substantial potential of the proposed PDHCG method to greatly improve both the efficiency and scalability of solving complex quadratic programming challenges.

math.OC↗

Who Will Support My Project? Interactive Search of Potential Crowdfunding Investors Through InSearch

Crowdfunding provides project founders with a convenient way to reach online investors. However, it is challenging for founders to find the most potential investors and successfully raise money for their projects on crowdfunding platforms. A few machine learning based methods have been proposed to recommend investors' interest in a specific crowdfunding project, but they fail to provide project founders with explanations in detail for these recommendations, thereby leading to an erosion of trust in predicted investors. To help crowdfunding founders find truly interested investors, we conducted semi-structured interviews with four crowdfunding experts and presents inSearch, a visual analytic system. inSearch allows founders to search for investors interactively on crowdfunding platforms. It supports an effective overview of potential investors by leveraging a Graph Neural Network to model investor preferences. Besides, it enables interactive exploration and comparison of the temporal evolution of different investors' investment details.

cs.HC↗