arXiv subjects

Lei Feng

Publications and source records attributed to Lei Feng.

At least 37 records · Page 2Linked to original sources

arXiv2026

Test-Time Attention Purification for Backdoored Large Vision Language Models

Despite the strong multimodal performance, large vision-language models (LVLMs) are vulnerable during fine-tuning to backdoor attacks, where adversaries insert trigger-embedded samples into the training data to implant behaviors that can be maliciously activated at test time. Existing defenses typically rely on retraining backdoored parameters (e.g., adapters or LoRA modules) with clean data, which is computationally expensive and often degrades model performance. In this work, we provide a new mechanistic understanding of backdoor behaviors in LVLMs: the trigger does not influence prediction through low-level visual patterns, but through abnormal cross-modal attention redistribution, where trigger-bearing visual tokens steal attention away from the textual context - a phenomenon we term attention stealing. Motivated by this, we propose CleanSight, a training-free, plug-and-play defense that operates purely at test time. CleanSight (i) detects poisoned inputs based on the relative visual-text attention ratio in selected cross-modal fusion layers, and (ii) purifies the input by selectively pruning the suspicious high-attention visual tokens to neutralize the backdoor activation. Extensive experiments show that CleanSight significantly outperforms existing pixel-based purification defenses across diverse datasets and backdoor attack types, while preserving the model's utility on both clean and poisoned samples.

cs.CV↗

arXiv2026

FastBUS: A Fast Bayesian Framework for Unified Weakly-Supervised Learning

Machine Learning often involves various imprecise labels, leading to diverse weakly supervised settings. While recent methods aim for universal handling, they usually suffer from complex manual pre-work, ignore the relationships between associated labels, or are unable to batch process due to computational design flaws, resulting in long running times. To address these limitations, we propose a novel general framework that efficiently infers latent true label distributions across various weak supervisions. Our key idea is to express the label brute-force search process as a probabilistic transition of label variables, compressing diverse weakly supervised DFS tree structures into a shared Bayesian network. From this, we derived a latent probability calculation algorithm based on generalized belief propagation and proposed two joint acceleration strategies: 1) introducing a low-rank assumption to approximate the transition matrix, reducing time complexity; 2) designing an end-to-end state evolution module to learn batch-scale transition matrices, facilitating multi-category batch processing. In addition, the equivalence of our method with the EM algorithm in most scenarios is further demonstrated. Extensive experiments show that our method achieves SOTA results under most weakly supervised settings, and achieves up to hundreds of times faster acceleration in running time compared to other general methods.

cs.LG↗

arXiv2026

Hierarchy-of-Groups Policy Optimization for Long-Horizon Agentic Tasks

Group-based reinforcement learning (RL), such as GRPO, has advanced the capabilities of large language models on long-horizon agentic tasks. To enable more fine-grained policy updates, recent research has increasingly shifted toward stepwise group-based policy optimization, which treats each step in a rollout trajectory independently while using a memory module to retain historical context. However, we find a key issue in estimating stepwise relative advantages, namely context inconsistency, where steps within the same group may differ in their historical contexts. Empirically, we reveal that this issue can lead to severely biased advantage estimation, thereby degrading policy optimization significantly. To address the issue, in this paper, we propose Hierarchy-of-Groups Policy Optimization (HGPO) for long-horizon agentic tasks. Specifically, within a group of rollout trajectories, HGPO assigns each step to multiple hierarchical groups according to the consistency of historical contexts. Then, for each step, HGPO computes distinct advantages within each group and aggregates them with an adaptive weighting scheme. In this way, HGPO can achieve a favorable bias-variance trade-off in stepwise advantage estimation, without extra models or rollouts. Evaluations on two challenging agentic tasks, ALFWorld and WebShop with Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct, show that HGPO significantly outperforms existing agentic RL methods under the same computational constraints. Code is available at https://github.com/langfengQ/verl-agent/tree/master/recipe/hgpo.

cs.LG↗

arXiv2026

Phase-Aware Mixture of Experts for Agentic Reinforcement Learning

Reinforcement learning (RL) has equipped LLM agents with a strong ability to solve complex tasks. However, existing RL methods normally use a \emph{single} policy network, causing \emph{simplicity bias} where simple tasks occupy most parameters and dominate gradient updates, leaving insufficient capacity for complex tasks. A plausible remedy could be employing the Mixture-of-Experts (MoE) architecture in the policy network, as MoE allows different parameters (experts) to specialize in different tasks, preventing simple tasks from dominating all parameters. However, a key limitation of traditional MoE is its token-level routing, where the router assigns each token to specialized experts, which fragments phase-consistent patterns into scattered expert assignments and thus undermines expert specialization. In this paper, we propose \textbf{Phase-Aware Mixture of Experts (PA-MoE)}. It first features a lightweight \emph{phase router} that learns latent phase boundaries directly from the RL objective without pre-defining phase categories. Then, the phase router allocates temporally consistent assignments to the same expert, allowing experts to preserve phase-specific expertise. Experimental results demonstrate the effectiveness of our proposed PA-MoE.

cs.AI↗

arXiv2026

When More Experts Hurt: Underfitting in Multi-Expert Learning to Defer

Learning to Defer (L2D) enables a classifier to abstain from predictions and defer to an expert, and has recently been extended to multi-expert settings. In this work, we show that multi-expert L2D is fundamentally more challenging than the single-expert case. With multiple experts, the classifier's underfitting becomes inherent, which seriously degrades prediction performance, whereas in the single-expert setting it arises only under specific conditions. We theoretically reveal that this stems from an intrinsic expert identifiability issue: learning which expert to trust from a diverse pool, a problem absent in the single-expert case and renders existing underfitting remedies failed. To tackle this issue, we propose PiCCE (Pick the Confident and Correct Expert), a surrogate-based method that adaptively identifies a reliable expert based on empirical evidence. PiCCE effectively reduces multi-expert L2D to a single-expert-like learning problem, thereby resolving multi expert underfitting. We further prove its statistical consistency and ability to recover class probabilities and expert accuracies. Extensive experiments across diverse settings, including real-world expert scenarios, validate our theoretical results and demonstrate improved performance.

cs.LG↗

arXiv2026

PhGPO: Pheromone-Guided Policy Optimization for Long-Horizon Tool Planning

Recent advancements in Large Language Model (LLM) agents have demonstrated strong capabilities in executing complex tasks through tool use. However, long-horizon multi-step tool planning is challenging, because the exploration space suffers from a combinatorial explosion. In this scenario, even when a correct tool-use path is found, it is usually considered an immediate reward for current training, which would not provide any reusable information for subsequent training. In this paper, we argue that historically successful trajectories contain reusable tool-transition patterns, which can be leveraged throughout the whole training process. Inspired by ant colony optimization where historically successful paths can be reflected by the pheromone, we propose Pheromone-Guided Policy Optimization (PhGPO), which learns a trajectory-based transition pattern (i.e., pheromone) from historical trajectories and then uses the learned pheromone to guide policy optimization. This learned pheromone provides explicit and reusable guidance that steers policy optimization toward historically successful tool transitions, thereby improving long-horizon tool planning. Comprehensive experimental results demonstrate the effectiveness of our proposed PhGPO.

cs.AI↗

arXiv2026

Mitigating Mismatch within Reference-based Preference Optimization

Direct Preference Optimization (DPO) has become the de facto standard for offline preference alignment of large language models, but its reliance on a reference policy introduces a critical tension. DPO weighs each update relative to a reference, which stabilizes the training by regularizing the updates within a trusted region. This reliance becomes problematic for pessimistic pairs, where the reference model prefers the rejected response. For these pairs, DPO prematurely attenuates the gradient as soon as the policy margin ($\Delta_\theta$) merely beats the reference margin ($\Delta_{\mathrm{ref}}$) even if the policy is still wrong ($\Delta_{\theta}<0$). We name this failure premature satisfaction, which is a concrete form of the training-inference mismatch. Reference-free objectives remove this mismatch by optimizing the absolute margin, but at the cost of discarding the stabilizing signal of the reference. We mitigate this tension with Hybrid-DPO (HyPO), a drop-in modification to DPO that applies reference conditionally: HyPO behaves exactly like DPO when the reference is optimistic or neutral, and it treats the reference as neutral when it is pessimistic by replacing $\Delta_\theta-\Delta_{\mathrm{ref}}$ with $\Delta_\theta-\max\{0,\Delta_{\mathrm{ref}}\}$. This one-line change strictly strengthens per-example learning signals on pessimistic pairs while preserving DPO's objective form and computational cost. By conditionally debiasing the pessimistic reference signal, HyPO mitigates premature satisfaction; empirically, across preference alignment, HyPO improves inference-aligned metrics and achieves higher pairwise win rates. Our results provide evidence that direct preference alignment could be enhanced by conditionally debiasing the reference signal, rather than discarding it.

cs.LG↗

arXiv2026

Online Causal Kalman Filtering for Stable and Effective Policy Optimization

Reinforcement learning for large language models suffers from high-variance token-level importance sampling (IS) ratios, which would destabilize policy optimization at scale. To improve stability, recent methods typically use a fixed sequence-level IS ratio for all tokens in a sequence or adjust each token's IS ratio separately, thereby neglecting temporal off-policy derivation across tokens in a sequence. In this paper, we first empirically identify that local off-policy deviation is structurally inconsistent at the token level, which may distort policy-gradient updates across adjacent tokens and lead to training collapse. To address the issue, we propose Online Causal Kalman Filtering for stable and effective Policy Optimization (KPO). Concretely, we model the desired IS ratio as a latent state that evolves across tokens and apply a Kalman filter to update this state online and autoregressively based on the states of past tokens, regardless of future tokens. The resulting filtered IS ratios preserve token-wise local structure-aware variation while strongly smoothing noise spikes, yielding more stable and effective policy updates. Experimentally, KPO achieves superior results on challenging math reasoning datasets compared with state-of-the-art counterparts.

cs.CL↗

arXiv2026

Synchrotron self-Compton process for constraining sub-GeV dark matter in Omega Centauri via SKA

The search for the particle identity of dark matter (DM) continues to be a primary objective in modern physics. In this field, the sub-GeV mass range of DM detection remains a crucial yet challenging window. We investigate synchrotron self-Compton (SSC) emission from electrons and positrons produced by MeV-scale DM annihilation as a novel indirect detection channel. Focusing on the globular cluster Omega Centauri and the sensitivity of the Square Kilometre Array, we derive constraints on the annihilation cross section reaching $\langle\sigma v\rangle \sim 10^{-29}\,\rm{cm}^{3}\,\rm{s}^{-1}$ in the tens-of-MeV range. Furthermore, constraints can reach below $\langle\sigma v\rangle \sim 10^{-30}\,\rm{cm}^{3}\,\rm{s}^{-1}$ under favorable parameter choices. Although the derived limits depend on the uncertain propagation parameters, the SSC channel remains competitive with existing indirect constraints over a representative range of astrophysical assumptions, establishing SSC emission as a promising probe of sub-GeV DM.

astro-ph.CO↗

arXiv2026

Do All Individual Layers Help? An Empirical Study of Task-Interfering Layers in Vision-Language Models

Current VLMs have demonstrated capabilities across a wide range of multimodal tasks. Typically, in a pretrained VLM, all layers are engaged by default to make predictions on downstream tasks. We find that intervening on a single layer, such as by zeroing its parameters, can improve the performance on certain tasks, indicating that some layers hinder rather than help downstream tasks. We systematically investigate how individual layers influence different tasks via layer intervention. Specifically, we measure the change in performance relative to the base model after intervening on each layer and observe improvements when bypassing specific layers. This improvement can be generalizable across models and datasets, indicating the presence of Task-Interfering Layers that harm downstream tasks' performance. We introduce Task-Layer Interaction Vector, which quantifies the effect of intervening on each layer of a VLM given a task. These task-interfering layers exhibit task-specific sensitivity patterns: tasks requiring similar capabilities show consistent response trends under layer interventions, as evidenced by the high similarity in their task-layer interaction vectors. Inspired by these findings, we propose TaLo (Task-Adaptive Layer Knockout), a training-free, test-time adaptation method that dynamically identifies and bypasses the most interfering layer for a given task. Without parameter updates, TaLo improves performance across various models and datasets, including boosting Qwen-VL's accuracy on the Maps task in ScienceQA by up to 16.6%. Our work reveals an unexpected form of modularity in pretrained VLMs and provides a plug-and-play, training-free mechanism to unlock hidden capabilities at inference time. The source code will be publicly available.

cs.AI↗

arXiv2026

Guided by Trajectories: Repairing and Rewarding Tool-Use Trajectories for Tool-Integrated Reasoning

Tool-Integrated Reasoning (TIR) enables large language models (LLMs) to solve complex tasks by interacting with external tools, yet existing approaches depend on high-quality synthesized trajectories selected by scoring functions and sparse outcome-based rewards, providing limited and biased supervision for learning TIR. To address these challenges, in this paper, we propose AutoTraj, a two-stage framework that automatically learns TIR by repairing and rewarding tool-use trajectories. Specifically, in the supervised fine-tuning (SFT) stage, AutoTraj generates multiple candidate tool-use trajectories for each query and evaluates them along multiple dimensions. High-quality trajectories are directly retained, while low-quality ones are repaired using a LLM (i.e., LLM-as-Repairer). The resulting repaired and high-quality trajectories form a synthetic SFT dataset, while each repaired trajectory paired with its original low-quality counterpart constitutes a dataset for trajectory preference modeling. In the reinforcement learning (RL) stage, based on the preference dataset, we train a trajectory-level reward model to assess the quality of reasoning paths and combine it with outcome and format rewards, thereby explicitly guiding the optimization toward reliable TIR behaviors. Experiments on real-world benchmarks demonstrate the effectiveness of AutoTraj in TIR.

cs.AI↗

arXiv2025

New Constraints on Cosmic-ray boosted Sub-GeV Dark Matter via Light Mediators

Traditional direct detection experiments lack the sensitivity to probe the sub-GeV dark matter (DM), primarily due to the low energy of the expected nuclear recoils. In this work, we investigate cosmic-ray (CR) upscattering as a mechanism to accelerate DM particles to detectable velocities in underground experiments. By analyzing four models of DM-nucleon interactions -- namely scalar, vector, pseudoscalar, and axial-vector mediators -- we derive constraints on the coupling parameters using data from the LZ, XENON, and Borexino experiments, covering mediator mass from $10^{-6}$ to $1$ GeV. As the mediator mass varies, the shift in dominance between momentum transfer and mediator mass leads to a turnover in the constraints around $10^{-2}$--$10^{-3}~\mathrm{GeV}$. Our results extend the reach of direct detection into the sub-GeV window and clarify the critical role of momentum dependence in light-mediator scenarios.

astro-ph.HE↗

arXiv2025

LLM-Empowered Agentic AI for QoE-Aware Network Slicing Management in Industrial IoT

The Industrial Internet of Things (IIoT) requires networks that deliver ultra-low latency, high reliability, and cost efficiency, which traditional optimization methods and deep reinforcement learning (DRL)-based approaches struggle to provide under dynamic and heterogeneous workloads. To address this gap, large language model (LLM)-empowered agentic AI has emerged as a promising paradigm, integrating reasoning, planning, and adaptation to enable QoE-aware network management. In this paper, we explore the integration of agentic AI into QoE-aware network slicing for IIoT. We first review the network slicing management architecture, QoE metrics for IIoT applications, and the challenges of dynamically managing heterogeneous network slices, while highlighting the motivations and advantages of adopting agentic AI. We then present the workflow of agentic AI-based slicing management, illustrating the full lifecycle of AI agents from processing slice requests to constructing slice instances and performing dynamic adjustments. Furthermore, we propose an LLM-empowered agentic AI approach for slicing management, which integrates a retrieval-augmented generation (RAG) module for semantic intent inference, a DRL-based orchestrator for slicing configuration, and an incremental memory mechanism for continual learning and adaptation. Through a case study on heterogeneous slice management, we demonstrate that the proposed approach significantly outperforms other baselines in balancing latency, reliability, and cost, and achieves up to a 19% improvement in slice availability ratio.

cs.NI↗

arXiv2025

An Invariant Latent Space Perspective on Language Model Inversion

Language model inversion (LMI), i.e., recovering hidden prompts from outputs, emerges as a concrete threat to user privacy and system security. We recast LMI as reusing the LLM's own latent space and propose the Invariant Latent Space Hypothesis (ILSH): (1) diverse outputs from the same source prompt should preserve consistent semantics (source invariance), and (2) input<->output cyclic mappings should be self-consistent within a shared latent space (cyclic invariance). Accordingly, we present Inv^2A, which treats the LLM as an invariant decoder and learns only a lightweight inverse encoder that maps outputs to a denoised pseudo-representation. When multiple outputs are available, they are sparsely concatenated at the representation layer to increase information density. Training proceeds in two stages: contrastive alignment (source invariance) and supervised reinforcement (cyclic invariance). An optional training-free neighborhood search can refine local performance. Across 9 datasets covering user and system prompt scenarios, Inv^2A outperforms baselines by an average of 4.77% BLEU score while reducing dependence on large inverse corpora. Our analysis further shows that prevalent defenses provide limited protection, underscoring the need for stronger strategies. The source code and data involved in this paper can be found in https://github.com/yyy01/Invariant_Attacker.

cs.LG↗

arXiv2025

Exploring Weak-to-Strong Generalization for CLIP-based Classification

Aligning large-scale commercial models with user intent is crucial to preventing harmful outputs. Current methods rely on human supervision but become impractical as model complexity increases. When models surpass human knowledge, providing accurate feedback becomes challenging and inefficient. A novel solution proposed recently is using a weaker model to supervise a stronger model. This concept leverages the ability of weaker models to perform evaluations, thereby reducing the workload on human supervisors. Previous work has shown the effectiveness of weak-to-strong generalization in the context of language-only models. Extending this concept to vision-language models leverages these insights, adapting the proven benefits to a multi-modal context. In our study, we explore weak-to-strong generalization for CLIP-based classification. We propose a method, class prototype learning (CPL), which aims to enhance the classification capabilities of the CLIP model, by learning more representative prototypes for each category. Our findings indicate that, despite using a simple loss function under weak supervision, CPL yields robust improvements in targeted scenarios, particularly when pretraining is limited. Extensive experiments demonstrate that our approach is effective under these settings, achieving a 3.67% improvement over strong baseline methods.

cs.CV↗

arXiv2025

Charge-dependent spectral softenings of primary cosmic-rays below the knee

In most particle acceleration or propagation theories, the characteristic features of the cosmic ray spectra due to acceleration limits or propagation phase changes are charge dependent. Alternatively, the interaction scenario would expect mass dependent spectral features in general. The observational verification of which relation takes effect in nature is still lack due to the difficulty of measuring the spectra of individual particles up to very high energies. Here we report direct measurements of the carbon, oxygen, and iron spectra from ~20 gigavolts to ~100 teravolts (~60 teravolts for iron) with 9 years of on-orbit data collected by the Dark Matter Particle Explorer. Distinct spectral softenings have been directly detected in these spectra for the first time. Combined with the updated proton and helium spectra, the spectral softening appears universally at a rigidity of ~15 teravolts. A nuclei mass dependent softening is rejected at a confidence level of >99.999%. Possible interpretations of these results, including a nearby cosmic ray source and other models such as the propagation effect, are discussed.

astro-ph.HE↗

arXiv2025

Universal Relations in Long-range Quantum Spin Chains

Understanding the emergence of novel collective behaviors in strongly interacting systems lies at the heart of quantum many-body physics. Valuable insight comes from examining how few-body correlations manifest in many-body systems, embodying the ``from few to many'' philosophy. An intriguing example is the set of universal relations in ultracold atomic gases, which connect a wide range of observables to a single quantity known as the contact. In this Letter, we demonstrate that universal relations manifest in a distinct class of quantum many-body systems, long-range quantum spin chains, which belong to a completely new universality class. Using effective field theory and the operator product expansion, we establish connections between the asymptotic behavior of equal-time spin correlation functions, the dynamical structure factor, and the contact density. The theoretical predictions for equal-time correlators are explicitly verified through numerical simulations based on matrix product states. Our results could be readily tested in state-of-the-art trapped-ion systems.

cond-mat.quant-gas↗

arXiv2025

Rethinking Toxicity Evaluation in Large Language Models: A Multi-Label Perspective

Large language models (LLMs) have achieved impressive results across a range of natural language processing tasks, but their potential to generate harmful content has raised serious safety concerns. Current toxicity detectors primarily rely on single-label benchmarks, which cannot adequately capture the inherently ambiguous and multi-dimensional nature of real-world toxic prompts. This limitation results in biased evaluations, including missed toxic detections and false positives, undermining the reliability of existing detectors. Additionally, gathering comprehensive multi-label annotations across fine-grained toxicity categories is prohibitively costly, further hindering effective evaluation and development. To tackle these issues, we introduce three novel multi-label benchmarks for toxicity detection: \textbf{Q-A-MLL}, \textbf{R-A-MLL}, and \textbf{H-X-MLL}, derived from public toxicity datasets and annotated according to a detailed 15-category taxonomy. We further provide a theoretical proof that, on our released datasets, training with pseudo-labels yields better performance than directly learning from single-label supervision. In addition, we develop a pseudo-label-based toxicity detection method. Extensive experimental results show that our approach significantly surpasses advanced baselines, including GPT-4o and DeepSeek, thus enabling more accurate and reliable evaluation of multi-label toxicity in LLM-generated content.

cs.CL↗