arXiv subjects

Zhaokun Wang

Publications and source records attributed to Zhaokun Wang.

16 recordsLinked to original sources

arXiv2026

Unlearning Is Not Just Erasing: Temporal Decoupling via Generation Inequality

Large language models (LLMs) require effective unlearning to address privacy regulations and safety concerns. However, achieving precise forgetting without compromising general utility remains challenging. Existing sequence- and token-level methods penalize target outputs without modeling their context-dependent retrieval paths, which can disrupt linguistic structure or suppress benign knowledge. We present ADU, a fine-grained, training-based framework that shifts unlearning from token erasure to contextual attention-pathway decoupling. Exploiting the functional distinction between local and global attention heads, ADU identifies preplan positions that retrieve persistent sensitive anchors and fixes their candidate paths under the original model. It then trains attention-projection adapters to suppress attention mass along these paths while preserving local-attention structure and retain-set language modeling. Post-training activation exchange tests whether the modified attention-output module transmits the learned forgetting effect. ADU achieves the strongest aggregate performance among evaluated baselines on the TOFU and WMDP benchmarks, including a Forget Quality of (0.93) on TOFU. It preserves 87--98% of model utility (92.9% on average versus 81.9% for baselines) while reducing side effects in benign contexts.

cs.CL↗

arXiv2026

Power-law-anchored residual learning for H-mode energy confinement time in tokamaks: interpolation and parameter-defined extrapolation

Reliable prediction of the energy confinement time is essential for magnetic-confinement fusion. Conventional power-law scalings provide constrained extrapolation trends but cannot represent complex nonlinearities, whereas neural networks interpolate accurately but may behave unpredictably outside the training distribution. We propose a unified power-law-anchored residual-learning framework in which a frozen empirical power-law scaling supplies the global trend and a nonlinear model learns only the systematic residual in logarithmic space. PLR-KAN is developed as the primary implementation, while a parameter-matched PLR-MLP serves as a controlled architecture replacement. Using the ITPA DB5.2.3 H-mode confinement database, we evaluate interpolation and parameter-defined held-out cohorts over ten complete training pipelines. PLR-KAN retains near-best interpolation accuracy, achieving R2=0.9671+/-0.0027, while substantially improving the stability of direct KAN under parameter-defined distribution shifts. It outperforms direct KAN across all five non-epsilon single-parameter-defined cohorts and the core-five joint cohort, reaching R2=0.9263+/-0.0157 in the latter. Results from PLR-MLP further demonstrate that the benefit of power-law anchoring is not specific to KAN, although the effectiveness of residual transfer remains architecture and direction dependent. As an exploratory extension, a Mahalanobis-distance-based prediction-time gate improves stability in selected shifted regions but is not universally beneficial and cannot compensate for missing device or physics-regime coverage. Overall, power-law-anchored residual learning provides a practical balance between nonlinear interpolation capability and empirically constrained extrapolation behavior.

physics.plasm-ph↗

arXiv2026

AdapShot: Adaptive Many-Shot In-Context Learning with Semantic-Aware KV Cache Reuse

Many-Shot In-Context Learning (ICL) has emerged as a promising paradigm, leveraging extensive examples to unlock the reasoning potential of Large Language Models (LLMs). However, existing methods typically rely on a predetermined, fixed number of shots. This static approach often fails to adapt to the varying difficulty of different queries, leading to either insufficient context or interference from noise. Furthermore, the prohibitive computational and memory costs of long contexts severely limit Many-Shot's feasibility. To address the above limitations, we propose AdapShot, which dynamically optimizes shot counts and leverages KV cache reuse for efficient inference. Specifically, we design a probe-based evaluation mechanism that utilizes output entropy to determine the optimal number of shots. To bypass the redundant prefilling computation during both the probing and inference phases, we incorporate a semantics-aware KV cache reuse strategy. Within this reuse strategy, to address positional encoding incompatibilities, we introduce a decoupling and re-encoding method that enables the flexible reordering of cached key-value pairs. Extensive experiments demonstrate that AdapShot achieves an average performance gain of around 10% and a 4.64x speedup compared to state-of-the-art DBSA.

cs.AI↗

arXiv2026

CAP: Controllable Alignment Prompting for Unlearning in LLMs

Large language models (LLMs) trained on unfiltered corpora inherently risk retaining sensitive information, necessitating selective knowledge unlearning for regulatory compliance and ethical safety. However, existing parameter-modifying methods face fundamental limitations: high computational costs, uncontrollable forgetting boundaries, and strict dependency on model weight access. These constraints render them impractical for closed-source models, yet current non-invasive alternatives remain unsystematic and reliant on empirical experience. To address these challenges, we propose the Controllable Alignment Prompting for Unlearning (CAP) framework, an end-to-end prompt-driven unlearning paradigm. CAP decouples unlearning into a learnable prompt optimization process via reinforcement learning, where a prompt generator collaborates with the LLM to suppress target knowledge while preserving general capabilities selectively. This approach enables reversible knowledge restoration through prompt revocation. Extensive experiments demonstrate that CAP achieves precise, controllable unlearning without updating model parameters, establishing a dynamic alignment mechanism that overcomes the transferability limitations of prior methods.

cs.LG↗

arXiv2026

ALTER: Asymmetric LoRA for Token-Entropy-Guided Unlearning of LLMs

Large language models (LLMs) have advanced to encompass extensive knowledge across diverse domains. Yet controlling what a LLMs should not know is important for ensuring alignment and thus safe use. However, effective unlearning in LLMs is difficult due to the fuzzy boundary between knowledge retention and forgetting. This challenge is exacerbated by entangled parameter spaces from continuous multi-domain training, often resulting in collateral damage, especially under aggressive unlearning strategies. Furthermore, the computational overhead required to optimize State-of-the-Art (SOTA) models with billions of parameters poses an additional barrier. In this work, we present ALTER, a lightweight unlearning framework for LLMs to address both the challenges of knowledge entanglement and unlearning efficiency. ALTER operates through two phases: (I) high entropy tokens are captured and learned via the shared A matrix in LoRA, followed by (II) an asymmetric LoRA architecture that achieves a specified forgetting objective by parameter isolation and unlearning tokens within the target subdomains. Serving as a new research direction for achieving unlearning via token-level isolation in the asymmetric framework. ALTER achieves SOTA performance on TOFU, WMDP, and MUSE benchmarks with over 95% forget quality and shows minimal side effects through preserving foundational tokens. By decoupling unlearning from LLMs' billion-scale parameters, this framework delivers excellent efficiency while preserving over 90% of model utility, exceeding baseline preservation rates of 47.8-83.6%.

cs.CL↗

arXiv2025

Noise-Robustness Through Noise: A Framework combining Asymmetric LoRA with Poisoning MoE

Current parameter-efficient fine-tuning methods for adapting pre-trained language models to downstream tasks are susceptible to interference from noisy data. Conventional noise-handling approaches either rely on laborious data pre-processing or employ model architecture modifications prone to error accumulation. In contrast to existing noise-process paradigms, we propose a noise-robust adaptation method via asymmetric LoRA poisoning experts (LoPE), a novel framework that enhances model robustness to noise only with generated noisy data. Drawing inspiration from the mixture-of-experts architecture, LoPE strategically integrates a dedicated poisoning expert in an asymmetric LoRA configuration. Through a two-stage paradigm, LoPE performs noise injection on the poisoning expert during fine-tuning to enhance its noise discrimination and processing ability. During inference, we selectively mask the dedicated poisoning expert to leverage purified knowledge acquired by normal experts for noise-robust output. Extensive experiments demonstrate that LoPE achieves strong performance and robustness purely through the low-cost noise injection, which completely eliminates the requirement of data cleaning.

cs.LG↗

arXiv2025

LT-PINN: Lagrangian Topology-conscious Physics-informed Neural Network for Boundary-focused Engineering Optimization

Physics-informed neural networks (PINNs) have emerged as a powerful meshless tool for topology optimization, capable of simultaneously determining optimal topologies and physical solutions. However, conventional PINNs rely on density-based topology descriptions, which necessitate manual interpolation and limit their applicability to complex geometries. To address this, we propose Lagrangian topology-conscious PINNs (LT-PINNs), a novel framework for boundary-focused engineering optimization. By parameterizing the control variables of topology boundary curves as learnable parameters, LT-PINNs eliminate the need for manual interpolation and enable precise boundary determination. We further introduce specialized boundary condition loss function and topology loss function to ensure sharp and accurate boundary representations, even for intricate topologies. The accuracy and robustness of LT-PINNs are validated via two types of partial differential equations (PDEs), including elastic equation with Dirichlet boundary conditions and Laplace's equation with Neumann boundary conditions. Furthermore, we demonstrate effectiveness of LT-PINNs on more complex time-dependent and time-independent flow problems without relying on measurement data, and showcase their engineering application potential in flow velocity rearrangement, transforming a uniform upstream velocity into a sine-shaped downstream profile. The results demonstrate (1) LT-PINNs achieve substantial reductions in relative L2 errors compared with the state-of-art density topology-oriented PINNs (DT-PINNs), (2) LT-PINNs can handle arbitrary boundary conditions, making them suitable for a wide range of PDEs, and (3) LT-PINNs can infer clear topology boundaries without manual interpolation, especially for complex topologies.

cs.LG↗

arXiv2025

HASH-RAG: Bridging Deep Hashing with Retriever for Efficient, Fine Retrieval and Augmented Generation

Retrieval-Augmented Generation (RAG) encounters efficiency challenges when scaling to massive knowledge bases while preserving contextual relevance. We propose Hash-RAG, a framework that integrates deep hashing techniques with systematic optimizations to address these limitations. Our queries directly learn binary hash codes from knowledgebase code, eliminating intermediate feature extraction steps, and significantly reducing storage and computational overhead. Building upon this hash-based efficient retrieval framework, we establish the foundation for fine-grained chunking. Consequently, we design a Prompt-Guided Chunk-to-Context (PGCC) module that leverages retrieved hash-indexed propositions and their original document segments through prompt engineering to enhance the LLM's contextual awareness. Experimental evaluations on NQ, TriviaQA, and HotpotQA datasets demonstrate that our approach achieves a 90% reduction in retrieval time compared to conventional methods while maintaining considerate recall performance. Additionally, The proposed system outperforms retrieval/non-retrieval baselines by 1.4-4.3% in EM scores.

cs.IR↗

arXiv2025

Accelerating Adaptive Retrieval Augmented Generation via Instruction-Driven Representation Reduction of Retrieval Overlaps

Retrieval-augmented generation (RAG) has emerged as a pivotal method for expanding the knowledge of large language models. To handle complex queries more effectively, researchers developed Adaptive-RAG (A-RAG) to enhance the generated quality through multiple interactions with external knowledge bases. Despite its effectiveness, A-RAG exacerbates the pre-existing efficiency challenges inherent in RAG, which are attributable to its reliance on multiple iterations of generation. Existing A-RAG approaches process all retrieved contents from scratch. However, they ignore the situation where there is a significant overlap in the content of the retrieval results across rounds. The overlapping content is redundantly represented, which leads to a large proportion of repeated computations, thus affecting the overall efficiency. To address this issue, this paper introduces a model-agnostic approach that can be generally applied to A-RAG methods, which is dedicated to reducing the redundant representation process caused by the overlapping of retrieval results. Specifically, we use cache access and parallel generation to speed up the prefilling and decoding stages respectively. Additionally, we also propose an instruction-driven module to further guide the model to more effectively attend to each part of the content in a more suitable way for LLMs. Experiments show that our approach achieves 2.79 and 2.33 times significant acceleration on average for prefilling and decoding respectively while maintaining equal generation quality.

cs.AI↗

arXiv2025

In situ measurement of electrical resistivity evolution during dynamic compression of copper

We report a novel experimental methodology for in situ measurement of electrical resistivity changes in T2 copper during dynamic compression utilizing a split Hopkinson pressure bar. The effects of adiabatic temperature rise and specimen shape deformation on the resistance were carefully accounted, which allowed one to isolate the contribution of microstructure changes such as dislocation evolution, defect generation, and lattice distortion. The latter allows for a real-time relationship between strain and electrical resistivity to be tracked. The experimental findings are also supplemented by molecular dynamics simulations that provide details about the process of microstructure evolution under dynamic loading. Up to now, very few in situ measurements has been carried out for changes in electrical resistivity during dynamic deformation, thus establishing a direct link for resistivity-strain which has important implications toward the understanding of plastic deformation and industrial application guidance.

cond-mat.mtrl-sci↗

arXiv2024

Multi Loss-based Feature Fusion and Top Two Voting Ensemble Decision Strategy for Facial Expression Recognition in the Wild

Facial expression recognition (FER) in the wild is a challenging task affected by the image quality and has attracted broad interest in computer vision. There is no research using feature fusion and ensemble strategy for FER simultaneously. Different from previous studies, this paper applies both internal feature fusion for a single model and feature fusion among multiple networks, as well as the ensemble strategy. This paper proposes one novel single model named R18+FAML, as well as one ensemble model named R18+FAML-FGA-T2V to improve the performance of the FER in the wild. Based on the structure of ResNet18 (R18), R18+FAML combines internal Feature fusion and three Attention blocks using Multiple Loss functions (FAML) to improve the diversity of the feature extraction. To improve the performance of R18+FAML, we propose a Feature fusion among networks based on the Genetic Algorithm (FGA), which can fuse the convolution kernels for feature extraction of multiple networks. On the basis of R18+FAML and FGA, we propose one ensemble strategy, i.e., the Top Two Voting (T2V) to support the classification of FER, which can consider more classification information comprehensively. Combining the above strategies, R18+FAML-FGA-T2V can focus on the main expression-aware areas. Extensive experiments demonstrate that our single model R18+FAML and the ensemble model R18+FAML-FGA-T2V achieve the accuracies of $\left( 90.32, 62.17, 65.83 \right)\%$ and $\left( 91.59, 63.27, 66.63 \right)\%$ on three challenging unbalanced FER datasets RAF-DB, AffectNet-8 and AffectNet-7 respectively, both outperforming the state-of-the-art results.

cs.CV↗

arXiv2023

Simulation of magnetic hyperthermia cancer treatment near a blood vessel

In this study, we conduct a study on magnetic hyperthermia treatment when a vessel is located near the tumor. The holistic framework is established to solve the process of tumor treatment. The interstitial tissue fluid, MNP distribution, temperature profile, and nanofluids are involved in the simulation. The study evaluates the cancer treatment efficacy by cumulative-equivalent-minutes-at-43 centigrade (CEM43), a widely accepted thermal dose. The influence of the nearby blood vessel is investigated, and parameter studies about the distance to the tumor, the width of the blood vessel, and the vessel direction are also conducted. After that, the effects of the fluid-structure interaction of moving vessel boundaries and blood rheology are discussed. The results demonstrate the cooling effect of a nearby blood vessel, and such effect reduces with the augment of the distance between the tumor and blood vessel. The combination of downward gravity and the cool effect from the lower horizontal vessel leads to the best performance with 97.77% ablation in the tumor and 0.87% injury in healthy tissue at distance d = 4 mm, but the cases of the vertical vessel are relatively poor. The vessel width and blood rheology affect the treatment by velocity gradient near the vessel wall. Additionally, the moving boundary almost has no impact on treatment efficacy. The simulation tool has been well-validated and the outcomes can provide a useful reference to magnetic hyperthermia treatment.

physics.flu-dyn↗

arXiv2023

Simulation of tumor ablation in hyperthermia cancer treatment: A parametric study

A holistic simulation framework is established on magnetic hyperthermia modeling to solve the treatment process of tumor, which is surrounded by a healthy tissue block. The interstitial tissue fluid, MNP distribution, temperature profile, and nanofluids are involved in the simulation. Study evaluates the cancer treatment efficacy by cumulative-equivalent-minutes-at-43 centigrade (CEM43), a widely accepted thermal dose coming from the cell death curve. Results are separated into the conditions of with or without gravity effect in the computational domain, where two baseline case are investigated and compared. An optimal treatment time 46.55 min happens in the baseline case without gravity, but the situation deteriorates with gravity effect where the time for totally killing tumor cells prolongs 36.11% and meanwhile causing 21.32% ablation in healthy tissue. For the cases without gravity, parameter study of Lewis number and Heat source number are conducted and the variation of optimal treatment time are both fitting to the inverse functions. For the case considering the gravity, parameters Buoyancy ratio and Darcy ratio are investigated and their influence on totally killing tumor cells and the injury on healthy tissue are matching with the parabolic functions. The results are beneficial to the prediction of various conditions, and provides useful guide to the magnetic hyperthermia treatment.

physics.med-ph↗

arXiv2022

Compositionally Complex Perovskite Oxides as a New Class of Li-Ion Solid Electrolytes

Compositionally complex ceramics (CCCs), including high-entropy ceramics (HECs) as a subclass, offer new opportunities of materials discovery beyond the traditional methodology of searching new stoichiometric compounds. Herein, we establish new strategies of tailoring CCCs via a seamless combination of (1) non-equimolar compositional designs and (2) controlling microstructures and interfaces. Using oxide solid electrolytes for all-solid-state batteries as an exemplar, we validate these new strategies via discovering a new class of compositionally complex perovskite oxides (CCPOs) to show the possibility of improving ionic conductivities beyond the limit of conventional doping. As an example (amongst the 28 CCPOs examined), we demonstrate that the ionic conductivity can be improved by >60% in (Li0.375Sr0.4375)(Ta0.375Nb0.375Zr0.125Hf0.125)O3-δ, in comparison with the state-of-art (Li0.375Sr0.4375)(Ta0.75Zr0.25)O3-δ (LSTZ) baseline, via maintaining comparable electrochemical stability. Furthermore, the ionic conductivity can be improved by another >70% via grain boundary (GB) engineering, achieving >270% of the LSTZ baseline. This work suggests transformative new strategies for designing and tailoring HECs and CCCs, thereby opening a new window for discovering materials for energy storage and many other applications.

cond-mat.mtrl-sci↗

arXiv2020

Compact optical polarization-insensitive zoom metalens-doublet

Metasurface-based lenses (metalenses) offer specific conceptual advantages compared to ordinary refractive lenses. For example, it is possible to tune the focal length of a metalens doublet by varying the relative angle between the two metalenses while fixing their distance, leading to an extremely compact zoom lens. An improved polarization-insensitive design based on silicon-nanocylinders on silica substrates is presented. This design is realized and characterized experimentally at 1550 nm wavelength. By varying the relative angle between the metalenses in steps of 10 degrees, tuning of the doublet focal length is demonstrated from -54 mm to -+3 mm to +54 mm. This results in a zoom factor of an imaging system varying between 1 and 18. For positive focal lengths, the doublet focusing efficiency has a minimum of 34% and a maximum of 83%. Experiment and theory are in very good agreement.

physics.optics↗

arXiv2018

Generalized Hartmann-Shack array of dielectric metalens sub-arrays for polarimetric beam profiling

To define and characterize optical systems, obtaining information on the amplitude, phase, and polarization profile of optical beams is of utmost importance. Polarimetry using bulk optics is well established to characterize the polarization state. Recently, metasurfaces and metalenses have successfully been introduced as compact optical components. Here, we take the metasurface concept to the system level by realizing arrays of metalens 2*3 sub-arrays, allowing to determine the polarization profile of an optical beam. We use silicon-based metalenses with a numerical aperture of 0.32 and a mean measured diffraction efficiency in transmission mode of 28% at 1550 nm wavelength. Together with a standard camera recording the array foci, our system is extremely compact and allows for real-time beam diagnostics by inspecting the foci amplitudes. By further analyzing the foci displacements in the spirit of a Hartmann-Shack wavefront sensor, we can simultaneously detect phase-gradient profiles. As application examples, we diagnose the polarization profiles of a radially polarized beam, an azimuthally polarized beam, and of a vortex beam.

physics.optics↗