Selected work

Representative Papers

Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class-Incremental Learning

Jul 08, 2024arXiv.org

Few-shot class-incremental learning (FSCIL) confronts a dual challenge: static architectures suffer from overfitting to base classes, while dynamic architectures incur excessive parameter growth. To address this, we propose Dual-SSM—a dual selective state space model framework. It introduces a class-sensitive selective scanning mechanism to decouple feature evolution between base and novel classes, and integrates sequence modeling–driven incremental feature alignment with dynamic weight projection to enable parameter-adaptive expansion and efficient knowledge consolidation. Evaluated on miniImageNet, CUB-200, and CIFAR-100, Dual-SSM consistently surpasses existing state-of-the-art methods. It significantly mitigates catastrophic forgetting and enhances few-shot generalization for novel classes. By jointly achieving parameter efficiency and architectural scalability, Dual-SSM establishes a new paradigm for FSCIL that balances lightweight design with extensibility.

6 citationsRead paper

Machine Learning-Guided Memory Optimization for DLRM Inference on Tiered Memory

Mar 01, 2025International Symposium on High-Performance Computer Architecture

Deep learning recommendation models (DLRMs) require terabyte-scale embedding memory, and while hierarchical memory offers cost efficiency, its irregular access patterns severely degrade embedding placement and cache efficiency. To address this, we propose RecMG—a novel system that decouples cache admission decisions from prefetching prediction into two independently trainable models for the first time. We introduce a differentiable loss function explicitly modeling long reuse distances and infrequent embedding accesses, significantly improving prefetching accuracy. RecMG integrates vectorized access pattern learning, hierarchical-memory-aware scheduling, and dynamic prefetching policies. Experimental results show that RecMG reduces on-demand embedding loads by 1.5–2.8× over state-of-the-art baselines and achieves up to a 43% reduction in end-to-end inference latency in industrial-scale deployments.

4 citationsRead paper

DiMo-GUI: Advancing Test-time Scaling in GUI Grounding via Modality-Aware Visual Reasoning

Jun 11, 2025arXiv.org

Natural language query-based GUI element localization faces challenges including visual element diversity, spatial clutter, and linguistic ambiguity. This paper proposes a training-free modality-separation and dynamic focusing framework: it decouples textual and icon representations, leverages a general-purpose vision-language model, and integrates candidate region generation with a progressive zooming mechanism to enable independent, modality-specific reasoning. Furthermore, hierarchical visual focusing is introduced, wherein spatial attention dynamically refines the localization region. To our knowledge, this is the first approach achieving modality-aware semantic alignment and fine-grained localization for GUI elements without any task-specific training. Evaluated on standard GUI benchmarks, our method significantly outperforms existing baselines, demonstrating both the effectiveness and generalizability of the co-designed modality separation and region focusing strategy.

2 citationsRead paper

Video Prediction Transformers without Recurrence or Convolution

Oct 07, 2024

To address the high computational cost of RNNs and the limited receptive field and poor generalization of CNNs in video prediction, this work proposes PredFormer—a purely Transformer-based architecture. PredFormer eliminates both recurrent and convolutional components, introducing two key innovations: a gated Transformer mechanism and a novel three-dimensional (3D) self-attention module, enabling end-to-end spatiotemporal modeling of frame sequences. Furthermore, it establishes the first 3D attention analysis framework specifically designed for video temporal modeling. Extensive experiments demonstrate that PredFormer achieves state-of-the-art performance on four benchmark datasets—Kinetics-600, BAIR, KTH, and UCF101—while significantly reducing computational complexity. This work provides the first empirical validation of the superiority and efficiency of a fully Transformer-based, RNN-free and CNN-free design for video prediction.

2 citationsRead paper

Leveraging LLMs for Mission Planning in Precision Agriculture

Jun 11, 2025

To address the challenge of non-expert users in precision agriculture struggling to specify complex data-collection tasks for autonomous robots, this paper proposes an end-to-end natural language task planning framework powered by large language models (LLMs). Methodologically, it introduces the first deep integration of ChatGPT into an agricultural robot’s closed-loop control system, establishing a three-layer “semantic–standard–execution” mapping: natural language instructions are parsed into IEEE-standardized task specifications and subsequently dispatched via ROS 2 nodes for robotic execution. To compensate for LLMs’ limitations in spatial reasoning and multi-objective path planning, the framework incorporates a spatial semantic parsing module and a route optimization module. Experimental evaluation in real-world farmland environments demonstrates a 92% task parsing accuracy and an 89% task execution success rate, significantly enhancing task specification usability, reusability, and cross-platform interoperability.

1 citations1 influentialRead paper

Recent publications

Latest Papers

Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Jul 29, 2026

This work challenges the prevailing assumption that large language model (LLM) agents inherently function without structured long-term memory, presenting the first systematic investigation into file systems as a memory substrate. The authors introduce a unified memory architecture comprising three collaborating agent types—manager, searcher, and executor—that jointly maintain a directory-tree-based Markdown storage system augmented with sandboxed shell tools, dedicated memory interfaces, and chunked retrieval strategies. Through comprehensive evaluation, they assess how organizational schemes, tooling, and model capabilities jointly influence memory coherence and task performance. Results demonstrate that well-structured memory organization can reduce retrieval costs by approximately 50%, yet most agents struggle to sustain such structure over time. Notably, changes to the toolset exert an impact on memory morphology comparable to switching the underlying LLM, underscoring that memory architecture constitutes a design space rather than a fixed default.

0 citationsRead paper

ProgFormer: Hierarchical Voxel Diffusion Transformer for Longitudinal Brain MRI Prediction

Jul 29, 2026

Existing longitudinal brain MRI prediction methods struggle to simultaneously preserve global structural stability and fine-grained details of local pathological changes. To address this, this work proposes a hierarchical voxel diffusion Transformer architecture featuring dual pathways: a coarse path models the overall brain structure and temporal context, while a fine path refines voxel-level local details using the coarse representation as a spatiotemporal anchor. The model jointly estimates a velocity field in voxel space and employs Euler integration for end-to-end future MRI synthesis. This approach introduces, for the first time, a dual-path hierarchical mechanism that directly models longitudinal dynamics in voxel space, circumventing reconstruction losses inherent in latent-space methods. By integrating conditional flow matching, it co-optimizes global consistency and local sensitivity. Comprehensive experiments on ADNI, AIBL, and OASIS benchmarks demonstrate significant improvements over state-of-the-art methods in both pairwise and trajectory-based evaluations.

0 citationsRead paper

Audio-Zero: Label-Free Self-Evolution for Fine-Grained Audio Reasoning

Jul 22, 2026

This work addresses the limitations of existing large audio language models, which struggle with fine-grained audio reasoning—such as understanding event order, repetition, and duration—and often rely on costly annotations or yield only coarse-grained semantics. To overcome these challenges, the authors propose Audio-Zero, the first self-evolving framework that operates without external labels. It constructs unsupervised contrastive audio pairs and introduces an auditory self-play mechanism: the model autonomously generates auditory cues and identifies “anomalous listeners” based on cue inconsistencies, driving reinforcement learning through verifiable intrinsic rewards. Evaluated on Qwen2-Audio-7B-Instruct and Qwen2.5-Omni-7B, this approach significantly improves performance on fine-grained reasoning benchmarks—including TREA, MMAU Test-mini, and MMAR—while preserving strong general audio understanding capabilities.

0 citationsRead paper

What Do They See? Interpreting Complex Road Scenarios Through the Eyes of Vision-Language-Action Models for Safe and Trustworthy Autonomous Vehicle Learning

Jul 18, 2026

This work addresses the opacity of end-to-end autonomous driving models in complex scenarios, which hinders safety assessment. To this end, the authors propose CVAA, a counterfactual ablation analysis framework that uniquely integrates generative image inpainting with mechanistic interpretability. By removing individual objects from front-view images to generate counterfactual samples, they construct the Counter-nuScenes dataset and analyze corresponding changes in model intermediate representations. Experiments on Alpamayo-1 and nuScenes demonstrate that CVAA can reveal whether the model’s causal dependencies on vehicles and pedestrians align with human expectations, while also uncovering anomalous responses to ostensibly irrelevant objects. This approach offers a novel pathway toward interpretable and trustworthy autonomous driving systems.

0 citationsRead paper

Controlling Motion Transfer in Diffusion Transformers via Attention Heads

Jul 13, 2026

This work addresses the challenge of aligning motion transfer with target semantics and spatial structure in video diffusion transformers by proposing a training-free, controllable motion transfer method. By analyzing the functional specialization of attention heads, the approach decouples motion and spatial structure at the attention layer for the first time, identifying dedicated motion and structure heads. Leveraging semantic correspondence guidance and selective feature injection, it enables precise control over motion transfer without requiring any parameter updates. The method offers strong interpretability and generates high-fidelity videos that maintain temporal coherence, accurately reproduce reference motion, and preserve the structural consistency of the target content.

0 citationsRead paper