Selected work

Representative Papers

OFA-MAS: One-for-All Multi-Agent System Topology Design based on Mixture-of-Experts Graph Generative Models

Jan 19, 2026

This work addresses the limited generalizability of existing multi-agent systems (MAS), which rely on task-specific collaboration topologies and cannot share structural knowledge across tasks. To overcome this, we propose OFA-TAD, a novel framework that pioneers a “one-model-for-all-tasks” paradigm by dynamically generating sparse, adaptive collaboration graphs from arbitrary natural language task descriptions using a single unified model. Our approach integrates a Task-Aware Graph State Encoder (TAGSE), a Mixture-of-Experts (MoE) graph generation architecture, and a three-stage training strategy—comprising unconditional pretraining, conditional pretraining with large-model-generated data, and empirical graph fine-tuning. Evaluated across six diverse benchmarks, OFA-TAD significantly outperforms task-specific methods, enabling unified cross-domain topology generation and efficient knowledge transfer.

1 citationsRead paper

Uncovering and Mitigating Positional Blind Spots in Vision-Language-Action Models

Aug 02, 2026

This work addresses positional blind spots (PBS) in vision-language-action (VLA) models for robotic manipulation—regions in the workspace where failure rates sharply increase due to minor, task-irrelevant changes in object positions, even when instructions and scenes remain unchanged. The study systematically demonstrates the prevalence and spatial clustering of PBS across five state-of-the-art VLA models. To mitigate this issue without requiring internal model access, the authors propose a black-box, two-stage framework: first, high-risk regions are precisely localized via grid-based sampling and a one-sided likelihood ratio test; then, local demonstration data from these regions are collected to fine-tune the policy using LoRA. Experiments on two benchmarks show PBS failure rates as high as 0.58; the proposed method achieves a blind spot detection F1 score of 0.678—significantly outperforming baselines—and reduces overall failure rates by 40.00%–85.19% while preserving global performance stability.

0 citationsRead paper

UASPL: Uncertainty-Aware Self-Paced Learning with Evidential Neural Networks

Jul 07, 2026

This work addresses a key limitation of traditional self-paced learning (SPL), which relies solely on loss values to distinguish easy from hard samples and often misclassifies unreliable samples as “easy,” thereby compromising training stability. To overcome this issue, the study introduces uncertainty estimation into the SPL framework for the first time, proposing an uncertainty-aware SPL method grounded in evidential deep learning and subjective logic. By incorporating prediction reliability into the sample selection process through an interpretable and scalable general loss function, the approach enables robust curriculum learning that progresses from easy to hard samples. Extensive experiments demonstrate that the proposed method significantly outperforms existing SPL techniques across multiple datasets, achieving superior performance in classification accuracy, generalization capability, and model interpretability.

0 citationsRead paper

The Turning Point of 3D Plant Phenotyping: 3D Foundation Models Enable Minute-to-Second Cross-Crop Reconstruction and Beyond

Jul 02, 2026

Traditional 3D plant phenotyping struggles with sparse viewpoints and self-occlusion in low-cost imagery—such as smartphone videos—resulting in inefficient and inaccurate reconstructions. This work proposes the first cross-crop phenotyping framework based on 3D foundation models (3DFMs), establishing an end-to-end pipeline that transforms sparse inputs into organ-level semantic segmentation and metrically accurate 3D reconstructions. The approach integrates feedforward geometry recovery, geometry-constrained 3D Gaussian splatting, iterative view synthesis, 2D-to-3D semantic transfer, and scale recovery. Evaluated on 26 plant sequences, the method reduces average reconstruction time from 6.52 minutes to 1.58 seconds while preserving high-fidelity geometry and precise phenotypic measurements, substantially enhancing throughput and scalability.

0 citationsRead paper

Invisible Manipulation Channels in AI-Assisted Financial Advisory: Implications for Market Integrity and Regulatory Design

Jun 14, 2026

This study reveals a stealthy manipulation vulnerability in the inference phase of large language models (LLMs) deployed as financial advisors, wherein adversaries can subtly bias investment recommendations through targeted sampling-layer perturbations while preserving output compliance and evading state-of-the-art watermarking and black-box detection mechanisms. To counter this threat, we propose a hardware-level defense integrating a quantum random number generator (QRNG) with a trusted execution environment (TEE). Experimental results demonstrate that the attack increases the frequency of directional keywords by 1.8–1.9× while remaining undetected by all six evaluated detectors. In contrast, our QRNG+TEE solution completely neutralizes the manipulation, reducing its success rate to the natural baseline, thereby confirming both the generality of the vulnerability and the efficacy of the proposed defense.

0 citationsRead paper

Recent publications

Latest Papers

Uncovering and Mitigating Positional Blind Spots in Vision-Language-Action Models

Aug 02, 2026

This work addresses positional blind spots (PBS) in vision-language-action (VLA) models for robotic manipulation—regions in the workspace where failure rates sharply increase due to minor, task-irrelevant changes in object positions, even when instructions and scenes remain unchanged. The study systematically demonstrates the prevalence and spatial clustering of PBS across five state-of-the-art VLA models. To mitigate this issue without requiring internal model access, the authors propose a black-box, two-stage framework: first, high-risk regions are precisely localized via grid-based sampling and a one-sided likelihood ratio test; then, local demonstration data from these regions are collected to fine-tune the policy using LoRA. Experiments on two benchmarks show PBS failure rates as high as 0.58; the proposed method achieves a blind spot detection F1 score of 0.678—significantly outperforming baselines—and reduces overall failure rates by 40.00%–85.19% while preserving global performance stability.

0 citationsRead paper

UASPL: Uncertainty-Aware Self-Paced Learning with Evidential Neural Networks

Jul 07, 2026

This work addresses a key limitation of traditional self-paced learning (SPL), which relies solely on loss values to distinguish easy from hard samples and often misclassifies unreliable samples as “easy,” thereby compromising training stability. To overcome this issue, the study introduces uncertainty estimation into the SPL framework for the first time, proposing an uncertainty-aware SPL method grounded in evidential deep learning and subjective logic. By incorporating prediction reliability into the sample selection process through an interpretable and scalable general loss function, the approach enables robust curriculum learning that progresses from easy to hard samples. Extensive experiments demonstrate that the proposed method significantly outperforms existing SPL techniques across multiple datasets, achieving superior performance in classification accuracy, generalization capability, and model interpretability.

0 citationsRead paper

The Turning Point of 3D Plant Phenotyping: 3D Foundation Models Enable Minute-to-Second Cross-Crop Reconstruction and Beyond

Jul 02, 2026

Traditional 3D plant phenotyping struggles with sparse viewpoints and self-occlusion in low-cost imagery—such as smartphone videos—resulting in inefficient and inaccurate reconstructions. This work proposes the first cross-crop phenotyping framework based on 3D foundation models (3DFMs), establishing an end-to-end pipeline that transforms sparse inputs into organ-level semantic segmentation and metrically accurate 3D reconstructions. The approach integrates feedforward geometry recovery, geometry-constrained 3D Gaussian splatting, iterative view synthesis, 2D-to-3D semantic transfer, and scale recovery. Evaluated on 26 plant sequences, the method reduces average reconstruction time from 6.52 minutes to 1.58 seconds while preserving high-fidelity geometry and precise phenotypic measurements, substantially enhancing throughput and scalability.

0 citationsRead paper

Invisible Manipulation Channels in AI-Assisted Financial Advisory: Implications for Market Integrity and Regulatory Design

Jun 14, 2026

This study reveals a stealthy manipulation vulnerability in the inference phase of large language models (LLMs) deployed as financial advisors, wherein adversaries can subtly bias investment recommendations through targeted sampling-layer perturbations while preserving output compliance and evading state-of-the-art watermarking and black-box detection mechanisms. To counter this threat, we propose a hardware-level defense integrating a quantum random number generator (QRNG) with a trusted execution environment (TEE). Experimental results demonstrate that the attack increases the frequency of directional keywords by 1.8–1.9× while remaining undetected by all six evaluated detectors. In contrast, our QRNG+TEE solution completely neutralizes the manipulation, reducing its success rate to the natural baseline, thereby confirming both the generality of the vulnerability and the efficacy of the proposed defense.

0 citationsRead paper

StarOR: Synergizing Tree Search and Test-Time Reinforcement Learning for Optimization Modeling

Jun 13, 2026

This work addresses key limitations of existing learning-based optimization modeling approaches, which rely heavily on large annotated datasets, exhibit limited generalization, and are prone to early error propagation due to one-shot formulation generation. To overcome these challenges, the authors propose StarOR, a novel framework that uniquely integrates Monte Carlo Tree Search (MCTS) with test-time reinforcement learning. StarOR dynamically adapts its policy by updating LoRA adapters and introduces an unsupervised, multi-dimensional reward mechanism to provide fine-grained intermediate feedback. Crucially, it leverages sibling nodes to construct local comparison sets, transforming search exploration into instance-level policy optimization and enabling label-free credit assignment. Evaluated across five optimization benchmarks, StarOR achieves significant performance gains over current methods and state-of-the-art large language models, even when using a 4B-parameter backbone.

0 citationsRead paper