Searcharxiv⌕ Search

arXiv subjects

Jiaxin Gao

Publications and source records attributed to Jiaxin Gao.

At least 19 recordsLinked to original sources
arXiv2026

Beyond Token-Local Imitation: Reward-Compatible Temporal Credit Assignment for On-Policy Distillation

On-policy distillation (OPD) has emerged as an effective approach for large language model post-training, yet existing objectives face a trade-off between objective fidelity and optimization stability. Token-level OPD provides stable but local supervision, whereas sequence-level OPD captures future credit at the cost of horizon-dependent variance. We establish a unified temporal-credit view of these formulations, showing that practical token-level OPD can be interpreted as a temporal approximation to the sequence-level reverse-KL gradient. Building on this connection, we propose $γ$OPD (GammaOPD), which uses discounted temporal credit assignment to balance long-horizon supervision and optimization stability, while admitting a horizon-independent variance bound. We further develop a reward-compatible bounded mixing (RBM) mechanism for $γ$OPD that balances verifiable outcome feedback with the discounted OPD advantage to move beyond purely teacher-dependent optimization. Experiments on mathematical and code reasoning demonstrate consistent improvements over existing OPD methods across vanilla, size-mismatched, and multi-teacher distillation settings.

cs.LG↗
arXiv2026

AutoPSO: A Meta-framework for Automated Particle Swarm Optimization

Particle swarm optimization (PSO) is a widely used metaheuristic, prized for its simplicity and small parameter set. Although decades of research have produced numerous PSO variants that improve performance by modifying key components (e.g., parameter schedules, swarm topologies, or updating rules), two fundamental challenges persist. First, most existing approaches are problem-specific and hand-crafted, leading to poor cross-task generalization and forcing practitioners to navigate an impractically large design space, which also hinders systematic reuse of prior effective mechanisms. Second, mainstream implementations remain CPU-bound, constraining scalability and substantially increasing computational cost in real-world applications. To address these challenges, we propose AutoPSO, a highly automated meta-framework for constructing customized PSO algorithms. AutoPSO formulates PSO-based optimization as a bi-level process: an outer search explores the joint space of effective PSO components, while an inner loop instantiates candidate variants to solve the target task and provide feedback. The outer search operates over a curated, open-design component pool, supporting flexible replacement of the component set and the outer optimizer. Crucially, by leveraging EvoX for population tensorization and batched evaluations, AutoPSO can efficiently assess thousands of particles within practical time budgets. Comprehensive experiments on numerical benchmarks and neuroevolution robotic control tasks demonstrate that AutoPSO consistently discovers novel PSO variants that significantly outperform strong baselines. Ablation and scalability studies further highlight the contribution of individual algorithmic components and confirm that AutoPSO achieves increasing performance gains with larger swarm sizes. Code is available at {https://github.com/EMI-Group/autopso}.

cs.NE↗
arXiv2026

Learning with Bilevel-Minimax Optimization for Efficient and Reliable Transfer Attacks

Transfer-based adversarial attacks craft adversarial examples using surrogate models to mislead black-box victim models. Beyond perturbation generation, transferability is fundamentally governed by the coupling of initialization, surrogate adaptation, and gradient dynamics. We revisit this challenge from a bilevel-minimax perspective and propose BMAT (Bilevel-Minimax Adversarial Transfer). The bilevel formulation captures the dependency between initialization and perturbation, while the inner minimax problem promotes surrogate robustness for cross-architecture generalization. Algorithmically, we develop an integrated bottom-up solver that combines a Soft Weight Modulator and an Implicit Gradient Approximator to enable ternary coupling among initialization, surrogate adaptation, and perturbation optimization. We further provide theoretical insights into the optimization dynamics of the proposed bilevel-minimax framework. Extensive experiments on classification and segmentation benchmarks show that BMAT outperforms more than 10 strong baselines across more than 30 victim models, improving both intra- and cross-architecture transfer and yielding up to a 2x reduction in mIoU. Code is available at https://github.com/callous-youth/BMAT.

cs.LG↗
arXiv2026

FAST: A Framework for Aligned Sampling and Training in Parallel Reinforcement Learning for Autonomous Driving

Deep reinforcement learning is pivotal for closed-loop autonomous driving yet remains constrained by severe bottlenecks in sampling efficiency. Standard parallel sampling mitigates this but suffers from the straggler effect, where the premature termination of a single environment necessitates a synchronized batch re-initialization, leading to suboptimal sample utilization and prohibitive re-initialization latency. To address this, we propose FAST, a synchronous parallel framework tailored for closed-loop simulation. Specifically, FAST employs Dynamic Parallel Sampling Alignment (DPSA) to maintain vectorization synchronization by extending terminated episodes via virtual continuation, thereby decoupling the sampling loop from individual terminations. By dynamically triggering global truncation based on the termination rate of parallel clips, FAST effectively eliminates the bottleneck of premature resets without sacrificing data diversity. Furthermore, to strictly preserve theoretical consistency, we incorporate a Scaled Mask-Padding Optimization (SMPO) that leverages validity masking and adaptive loss normalization to nullify the bias from auxiliary padding data. Empirical evaluations demonstrate that FAST achieves at least a 1.78 times wall-clock speedup over the single-clip baseline while preserving statistical unbiasedness.

cs.LG↗
arXiv2026

KGS-GCN: Kinematics-Driven Gaussian Splatting and Probabilistic Topology for Skeleton-Based Action Recognition

Skeleton-based action recognition is widely applied in sensor-based systems, including human-computer interaction and intelligent surveillance. However, typical sensors produce sparse and discrete joint coordinates, often leading to the loss of fine-grained spatiotemporal information during dynamic movements. Furthermore, predefined physical topologies restrict modeling potential long-range dependencies. To address these challenges, we propose KGS-GCN, which integrates kinematics-driven Gaussian splatting and probabilistic topology within a graph convolutional network. A Gaussian splatting module constructs anisotropic covariance matrices by extracting instantaneous joint velocity vectors, rendering sparse skeleton sequences into multi-view continuous heatmaps rich in spatiotemporal semantics. Additionally, a probabilistic topology construction strategy transcends physical connectivity limitations by utilizing the Bhattacharyya distance to quantify statistical correlations between joint Gaussian distributions, generating an adaptive prior adjacency matrix. Finally, the lightweight multi-view rendering branch and topological GCN backbone are unified through a visual context gating mechanism, enabling seamless fusion of continuous dynamic cues with structural priors while maintaining high computational efficiency, requiring only 1.4M parameters and 1.3 GFLOPs. Extensive experiments on multiple benchmark datasets demonstrate that KGS-GCN significantly enhances the modeling of complex spatiotemporal dynamics and achieves competitive performance at low computational cost, establishing an efficient paradigm for improving the perceptual robustness of low-fidelity sensor data.

cs.CV↗
arXiv2026

Learning to Evolve for Optimization via Stability-Inducing Neural Unrolling

Evolutionary algorithms serve as a powerful paradigm for tackling optimization challenges, yet their reliance on manually engineered heuristics inherently limits their adaptability across diverse landscapes. However, the transition from the hand-crafted heuristics to data-driven algorithms faces a fundamental dilemma: achieving neural \emph{plasticity} without sacrificing algorithmic stability. Although learned optimizers offer high adaptivity, their unconstrained update rules often result in unstable dynamics and brittle generalization on unseen landscapes. To address this challenge, this paper proposes Learning to Evolve (L2E), a bilevel meta-optimization framework that learns evolutionary search via stability-inducing neural unrolling. First, L2E reformulates population evolution as an unrolled fixed-point iteration via a structured neural operator. In this design, the inner loop imposes a stability-biased update structure, while the outer loop meta-trains the operator to produce effective search trajectories across tasks. Second, to balance global exploration with local refinement, a gradient-derived composite solver adaptively fuses learned evolutionary proposals with proxy numerical guidance in a differentiable manner. Extensive experiments on synthetic benchmarks and real-world control tasks demonstrate that L2E achieves substantial optimization performance, scales to high-dimensional problems, and exhibits robust zero-shot transfer across diverse test distributions.

cs.NE↗
arXiv2026

Evaluating and Mitigating LLM-as-a-judge Bias in Communication Systems

Large Language Models (LLMs) are increasingly being used to autonomously evaluate the quality of content in communication systems, e.g., to assess responses in telecom customer support chatbots. However, the impartiality of these AI "judges" is not guaranteed, and any biases in their evaluation criteria could skew outcomes and undermine user trust. In this paper, we systematically investigate judgment biases across 6 LLM-as-a-judge models spanning both prompt-based and fine-tuned judges under the pointwise scoring setting, encompassing 11 types of biases that cover both implicit and explicit forms. We observed that state-of-the-art LLM judges demonstrate robustness to biased inputs, generally assigning them lower scores than the corresponding clean samples. We further found that fine-tuning an LLM on high-scoring yet biased responses can significantly degrade its performance, highlighting the risk of training on biased data. We also discovered that the judged scores correlate with task difficulty: a challenging dataset like GPQA yields lower average scores, whereas an open-ended reasoning dataset (e.g., JudgeLM-val) sees higher average scores. Finally, we proposed four potential mitigation strategies to ensure fair and reliable AI judging in practical communication scenarios.

cs.AI↗
arXiv2026

Plato's Form: Toward Backdoor Defense-as-a-Service for LLMs with Prototype Representations

Large language models (LLMs) are increasingly deployed in security-sensitive applications, yet remain vulnerable to backdoor attacks. However, existing backdoor defenses are difficult to operationalize for Backdoor Defense-as-a-Service (BDaaS), as they require unrealistic side information (e.g., downstream clean data, known triggers/targets, or task domain specifics), and lack reusable, scalable purification across diverse backdoored models. In this paper, we present PROTOPURIFY, a backdoor purification framework via parameter edits under minimal assumptions. PROTOPURIFY first builds a backdoor vector pool from clean and backdoored model pairs, aggregates vectors into candidate prototypes, and selects the most aligned candidate for the target model via similarity matching. PROTOPURIFY then identifies a boundary layer through layer-wise prototype alignment and performs targeted purification by suppressing prototype-aligned components in the affected layers, achieving fine-grained mitigation with minimal impact on benign utility. Designed as a BDaaS-ready primitive, PROTOPURIFY supports reusability, customizability, interpretability, and runtime efficiency. Experiments across various LLMs on both classification and generation tasks show that PROTOPURIFY consistently outperforms 6 representative defenses against 6 diverse attacks, including single-trigger, multi-trigger, and triggerless backdoor settings. PROTOPURIFY reduces ASR to below 10%, and even as low as 1.6% in some cases, while incurring less than a 3% drop in clean utility. PROTOPURIFY further demonstrates robustness against adaptive backdoor variants and stability on non-backdoored models.

cs.CR↗
arXiv2026

Bootstrap Off-policy with World Model

Online planning has proven effective in reinforcement learning (RL) for improving sample efficiency and final performance. However, using planning for environment interaction inevitably introduces a divergence between the collected data and the policy's actual behaviors, degrading both model learning and policy improvement. To address this, we propose BOOM (Bootstrap Off-policy with WOrld Model), a framework that tightly integrates planning and off-policy learning through a bootstrap loop: the policy initializes the planner, and the planner refines actions to bootstrap the policy through behavior alignment. This loop is supported by a jointly learned world model, which enables the planner to simulate future trajectories and provides value targets to facilitate policy improvement. The core of BOOM is a likelihood-free alignment loss that bootstraps the policy using the planner's non-parametric action distribution, combined with a soft value-weighted mechanism that prioritizes high-return behaviors and mitigates variability in the planner's action quality within the replay buffer. Experiments on the high-dimensional DeepMind Control Suite and Humanoid-Bench show that BOOM achieves state-of-the-art results in both training stability and final performance. The code is accessible at https://github.com/molumitu/BOOM_MBRL.

cs.LG↗
arXiv2025

Lethe: Purifying Backdoored Large Language Models with Knowledge Dilution

Large language models (LLMs) have seen significant advancements, achieving superior performance in various Natural Language Processing (NLP) tasks. However, they remain vulnerable to backdoor attacks, where models behave normally for standard queries but generate harmful responses or unintended output when specific triggers are activated. Existing backdoor defenses either lack comprehensiveness, focusing on narrow trigger settings, detection-only mechanisms, and limited domains, or fail to withstand advanced scenarios like model-editing-based, multi-trigger, and triggerless attacks. In this paper, we present LETHE, a novel method to eliminate backdoor behaviors from LLMs through knowledge dilution using both internal and external mechanisms. Internally, LETHE leverages a lightweight dataset to train a clean model, which is then merged with the backdoored model to neutralize malicious behaviors by diluting the backdoor impact within the model's parametric memory. Externally, LETHE incorporates benign and semantically relevant evidence into the prompt to distract LLM's attention from backdoor features. Experimental results on classification and generation domains across 5 widely used LLMs demonstrate that LETHE outperforms 8 state-of-the-art defense baselines against 8 backdoor attacks. LETHE reduces the attack success rate of advanced backdoor attacks by up to 98% while maintaining model utility. Furthermore, LETHE has proven to be cost-efficient and robust against adaptive backdoor attacks.

cs.CL↗
arXiv2025

Distributional Soft Actor-Critic with Three Refinements

Reinforcement learning (RL) has shown remarkable success in solving complex decision-making and control tasks. However, many model-free RL algorithms experience performance degradation due to inaccurate value estimation, particularly the overestimation of Q-values, which can lead to suboptimal policies. To address this issue, we previously proposed the Distributional Soft Actor-Critic (DSAC or DSACv1), an off-policy RL algorithm that enhances value estimation accuracy by learning a continuous Gaussian value distribution. Despite its effectiveness, DSACv1 faces challenges such as training instability and sensitivity to reward scaling, caused by high variance in critic gradients due to return randomness. In this paper, we introduce three key refinements to DSACv1 to overcome these limitations and further improve Q-value estimation accuracy: expected value substitution, twin value distribution learning, and variance-based critic gradient adjustment. The enhanced algorithm, termed DSAC with Three refinements (DSAC-T or DSACv2), is systematically evaluated across a diverse set of benchmark tasks. Without the need for task-specific hyperparameter tuning, DSAC-T consistently matches or outperforms leading model-free RL algorithms, including SAC, TD3, DDPG, TRPO, and PPO, in all tested environments. Additionally, DSAC-T ensures a stable learning process and maintains robust performance across varying reward scales. Its effectiveness is further demonstrated through real-world application in controlling a wheeled robot, highlighting its potential for deployment in practical robotic tasks.

cs.LG↗
arXiv2025

Towards Action Hijacking of Large Language Model-based Agent

Recently, applications powered by Large Language Models (LLMs) have made significant strides in tackling complex tasks. By harnessing the advanced reasoning capabilities and extensive knowledge embedded in LLMs, these applications can generate detailed action plans that are subsequently executed by external tools. Furthermore, the integration of retrieval-augmented generation (RAG) enhances performance by incorporating up-to-date, domain-specific knowledge into the planning and execution processes. This approach has seen widespread adoption across various sectors, including healthcare, finance, and software development. Meanwhile, there are also growing concerns regarding the security of LLM-based applications. Researchers have disclosed various attacks, represented by jailbreak and prompt injection, to hijack the output actions of these applications. Existing attacks mainly focus on crafting semantically harmful prompts, and their validity could diminish when security filters are employed. In this paper, we introduce AI$\mathbf{^2}$, a novel attack to manipulate the action plans of LLM-based applications. Different from existing solutions, the innovation of AI$\mathbf{^2}$ lies in leveraging the knowledge from the application's database to facilitate the construction of malicious but semantically-harmless prompts. To this end, it first collects action-aware knowledge from the victim application. Based on such knowledge, the attacker can generate misleading input, which can mislead the LLM to generate harmful action plans, while bypassing possible detection mechanisms easily. Our evaluations on three real-world applications demonstrate the effectiveness of AI$\mathbf{^2}$: it achieves an average attack success rate of 84.30\% with the best of 99.70\%. Besides, it gets an average bypass rate of 92.7\% against common safety filters and 59.45\% against dedicated defense.

cs.CR↗
arXiv2024

Revisiting PCA for time series reduction in temporal dimension

Revisiting PCA for Time Series Reduction in Temporal Dimension; Jiaxin Gao, Wenbo Hu, Yuntian Chen; Deep learning has significantly advanced time series analysis (TSA), enabling the extraction of complex patterns for tasks like classification, forecasting, and regression. Although dimensionality reduction has traditionally focused on the variable space-achieving notable success in minimizing data redundancy and computational complexity-less attention has been paid to reducing the temporal dimension. In this study, we revisit Principal Component Analysis (PCA), a classical dimensionality reduction technique, to explore its utility in temporal dimension reduction for time series data. It is generally thought that applying PCA to the temporal dimension would disrupt temporal dependencies, leading to limited exploration in this area. However, our theoretical analysis and extensive experiments demonstrate that applying PCA to sliding series windows not only maintains model performance, but also enhances computational efficiency. In auto-regressive forecasting, the temporal structure is partially preserved through windowing, and PCA is applied within these windows to denoise the time series while retaining their statistical information. By preprocessing time-series data with PCA, we reduce the temporal dimensionality before feeding it into TSA models such as Linear, Transformer, CNN, and RNN architectures. This approach accelerates training and inference and reduces resource consumption. Notably, PCA improves Informer training and inference speed by up to 40% and decreases GPU memory usage of TimesNet by 30%, without sacrificing model accuracy. Comparative analysis against other reduction methods further highlights the effectiveness of PCA in improving the efficiency of TSA models.

cs.LG↗
arXiv2024

Auto-Regressive Moving Diffusion Models for Time Series Forecasting

Time series forecasting (TSF) is essential in various domains, and recent advancements in diffusion-based TSF models have shown considerable promise. However, these models typically adopt traditional diffusion patterns, treating TSF as a noise-based conditional generation task. This approach neglects the inherent continuous sequential nature of time series, leading to a fundamental misalignment between diffusion mechanisms and the TSF objective, thereby severely impairing performance. To bridge this misalignment, and inspired by the classic Auto-Regressive Moving Average (ARMA) theory, which views time series as continuous sequential progressions evolving from previous data points, we propose a novel Auto-Regressive Moving Diffusion (ARMD) model to first achieve the continuous sequential diffusion-based TSF. Unlike previous methods that start from white Gaussian noise, our model employs chain-based diffusion with priors, accurately modeling the evolution of time series and leveraging intermediate state information to improve forecasting accuracy and stability. Specifically, our approach reinterprets the diffusion process by considering future series as the initial state and historical series as the final state, with intermediate series generated using a sliding-based technique during the forward process. This design aligns the diffusion model's sampling procedure with the forecasting objective, resulting in an unconditional, continuous sequential diffusion TSF model. Extensive experiments conducted on seven widely used datasets demonstrate that our model achieves state-of-the-art performance, significantly outperforming existing diffusion-based TSF models. Our code is available on GitHub: https://github.com/daxin007/ARMD.

cs.LG↗
arXiv2024

Neutralizing Backdoors through Information Conflicts for Large Language Models

Large language models (LLMs) have seen significant advancements, achieving superior performance in various Natural Language Processing (NLP) tasks, from understanding to reasoning. However, they remain vulnerable to backdoor attacks, where models behave normally for standard queries but generate harmful responses or unintended output when specific triggers are activated. Existing backdoor defenses often suffer from drawbacks that they either focus on detection without removal, rely on rigid assumptions about trigger properties, or prove to be ineffective against advanced attacks like multi-trigger backdoors. In this paper, we present a novel method to eliminate backdoor behaviors from LLMs through the construction of information conflicts using both internal and external mechanisms. Internally, we leverage a lightweight dataset to train a conflict model, which is then merged with the backdoored model to neutralize malicious behaviors by embedding contradictory information within the model's parametric memory. Externally, we incorporate convincing contradictory evidence into the prompt to challenge the model's internal backdoor knowledge. Experimental results on classification and conversational tasks across 4 widely used LLMs demonstrate that our method outperforms 8 state-of-the-art backdoor defense baselines. We can reduce the attack success rate of advanced backdoor attacks by up to 98% while maintaining over 90% clean data accuracy. Furthermore, our method has proven to be robust against adaptive backdoor attacks. The code will be open-sourced upon publication.

cs.CL↗
arXiv2024

Deep-sub-cycle ultrafast optical pulses

Sub-cycle optical pulse is of great importance for ultrafast science and technology. While a narrower pulse can offer a higher temporal resolution, so far the pulse width has not reached the limit of half an optical cycle. Here we propose to break the half-cycle limit via inverse Compton scattering in a deep-subwavelength-confined infrared or THz optical driving field, which can be converted into a deep-sub-cycle (i.e., less than half a cycle) ultrafast pulse by relativistic electrons passing through. Quantitatively, by using a deep-subwavelength-confined 0.4-THz 1-ps pulsed driving field and a 3-MeV 1-fs electron bunch, we obtain a $\sim$ 0.1-cycle femtosecond pulse with a peak frequency of $\sim$ 26 THz ($\sim$ 38 fs in optical cycle) and a pulse width of $\sim$ 3.6 fs; with the same driving field and a 30-MeV 50-as electron bunch, we obtain a 0.17-cycle attosecond pulse with a peak frequency of $\sim$ 2.0 PHz ($\sim$ 500 as in optical cycle) and a pulse width of $\sim$ 86 as. Below the optical damage threshold of the field-confinement material, the single-pulse photon number can exceed $10^6$. Such pulses may open opportunities for studying light-matter interaction on the deep-sub-cycle level, and pave a way to unprecedented optical technology ranging from temporal super-resolution optical microscopy and spectroscopy to unconventional atom/molecule polarization and manipulation.

physics.optics↗
arXiv2024

Cross-variable Linear Integrated ENhanced Transformer for Photovoltaic power forecasting

Photovoltaic (PV) power forecasting plays a crucial role in optimizing the operation and planning of PV systems, thereby enabling efficient energy management and grid integration. However, un certainties caused by fluctuating weather conditions and complex interactions between different variables pose significant challenges to accurate PV power forecasting. In this study, we propose PV-Client (Cross-variable Linear Integrated ENhanced Transformer for Photovoltaic power forecasting) to address these challenges and enhance PV power forecasting accuracy. PV-Client employs an ENhanced Transformer module to capture complex interactions of various features in PV systems, and utilizes a linear module to learn trend information in PV power. Diverging from conventional time series-based Transformer models that use cross-time Attention to learn dependencies between different time steps, the Enhanced Transformer module integrates cross-variable Attention to capture dependencies between PV power and weather factors. Furthermore, PV-Client streamlines the embedding and position encoding layers by replacing the Decoder module with a projection layer. Experimental results on three real-world PV power datasets affirm PV-Client's state-of-the-art (SOTA) performance in PV power forecasting. Specifically, PV-Client surpasses the second-best model GRU by 5.3% in MSE metrics and 0.9% in accuracy metrics at the Jingang Station. Similarly, PV-Client outperforms the second-best model SVR by 10.1% in MSE metrics and 0.2% in accuracy metrics at the Xinqingnian Station, and PV-Client exhibits superior performance compared to the second-best model SVR with enhancements of 3.4% in MSE metrics and 0.9% in accuracy metrics at the Hongxing Station.

cs.LG↗
arXiv2024

Advancing Unsupervised Low-light Image Enhancement: Noise Estimation, Illumination Interpolation, and Self-Regulation

Contemporary Low-Light Image Enhancement (LLIE) techniques have made notable advancements in preserving image details and enhancing contrast, achieving commendable results on specific datasets. Nevertheless, these approaches encounter persistent challenges in efficiently mitigating dynamic noise and accommodating diverse low-light scenarios. Insufficient constraints on complex pixel-wise mapping learning lead to overfitting to specific types of noise and artifacts associated with low-light conditions, reducing effectiveness in variable lighting scenarios. To this end, we first propose a method for estimating the noise level in low light images in a quick and accurate way. This facilitates precise denoising, prevents over-smoothing, and adapts to dynamic noise patterns. Subsequently, we devise a Learnable Illumination Interpolator (LII), which employs learnlable interpolation operations between the input and unit vector to satisfy general constraints between illumination and input. Finally, we introduce a self-regularization loss that incorporates intrinsic image properties and essential visual attributes to guide the output towards meeting human visual expectations. Comprehensive experiments validate the competitiveness of our proposed algorithm in both qualitative and quantitative assessments. Notably, our noise estimation method, with linear time complexity and suitable for various denoisers, significantly improves both denoising and enhancement performance. Benefiting from this, our approach achieves a 0.675dB PSNR improvement on the LOL dataset and 0.818dB on the MIT dataset on LLIE task, even compared to supervised methods. The source code is available at \href{https://doi.org/10.5281/zenodo.11463142}{this DOI repository} and the specific code for noise estimation can be found at \href{https://github.com/GoogolplexGoodenough/noise_estimate}{this separate GitHub link}.

cs.CV↗