Selected work

Representative Papers

Learning to Synthesize Compatible Fashion Items Using Semantic Alignment and Collocation Classification: An Outfit Generation Framework

Sep 15, 2022IEEE Transactions on Neural Networks and Learning Systems

This work addresses the challenging problem of complete outfit generation conditioned on a single garment and target-region masks—a key task in fashion design automation. We propose OutfitGAN, an end-to-end generative framework that synthesizes compatible tops, bottoms, footwear, and accessories given an input garment and spatially localized masks. Methodologically, we introduce two novel components: (i) a Semantic Alignment Module (SAM) that models fine-grained cross-garment semantic correspondences, and (ii) a Compatibility Classification Module (CCM) that explicitly enforces style and semantic coherence. Our multi-stage GAN architecture integrates semantic segmentation guidance, feature-level alignment losses, compatibility-aware adversarial supervision, and mask-conditioned generation control. Evaluated on a large-scale dataset of 20,000 real-world outfits, OutfitGAN achieves state-of-the-art performance across image fidelity, perceptual realism, and outfit compatibility metrics. It enables high-fidelity, diverse, and interactive fashion editing.

13 citationsRead paper

Diffusion Language Models are Super Data Learners

Nov 05, 2025

The performance advantages of diffusion language models (DLMs) under data-constrained regimes remain unclear. Method: This work systematically investigates DLM training dynamics and generalization with limited unique tokens, introducing a triple-gain mechanism: (i) arbitrary-order sequence modeling, (ii) high-density computation via iterative bidirectional denoising, and (iii) intrinsic Monte Carlo sampling augmentation—collectively overcoming data-efficiency bottlenecks. Training employs repeated pretraining, input/parameter noise injection, and standard diffusion objectives. Results: A 1.7B-parameter DLM surpasses same-scale autoregressive (AR) models using only 10B unique Python tokens; a 1B-parameter DLM achieves 56.2% accuracy on HellaSwag and 33.7% on MMLU using merely 1B tokens—significantly outperforming AR baselines trained on comparable data volumes. This study provides the first empirical validation of DLMs’ sustained competitive advantage in the small-data, large-model paradigm.

6 citationsRead paper

LightTransfer: Your Long-Context LLM is Secretly a Hybrid Model with Effortless Adaptation

Oct 17, 2024

To address the excessive KV cache memory overhead in large language models (LLMs) for long-context processing, this paper proposes a training-free, lightweight architectural transformation method. It first defines and identifies “lazy layers” in Transformer architectures—layers exhibiting low sensitivity to attention span—and replaces them with streaming attention mechanisms to construct a hybrid attention architecture. The approach enables plug-and-play recompilation of pretrained models, supporting zero-shot long-context understanding and minimal fine-tuning (e.g., o1-style long-reasoning). Core techniques include layer-wise importance analysis and seamless streaming attention integration. Evaluated on LongBench, the method incurs <1.5% performance degradation while achieving 2.17× throughput improvement; on the AIME24 mathematics benchmark, it attains 53.3% accuracy. Key contributions are: (i) formal introduction of the “lazy layer” concept; (ii) realization of training-free architectural adaptation; and (iii) simultaneous achievement of high efficiency and strong generalization across diverse long-context tasks.

2 citationsRead paper

Rethinking the Trust Region in LLM Reinforcement Learning

Feb 04, 2026

This work addresses the instability and inefficiency in training large vocabulary language models with Proximal Policy Optimization (PPO), which stems from its ratio-clipping mechanism excessively suppressing updates to low-probability tokens while inadequately constraining high-probability ones. To overcome this limitation, the authors propose Divergence Proximal Policy Optimization (DPPO), a theoretically grounded approach that replaces heuristic clipping with direct estimates of policy divergence—such as total variation (TV) or Kullback–Leibler (KL) divergence—to enforce trust-region constraints. Efficient computation is achieved through Binary and Top-K sparse approximations of the divergence measures. Empirical results across multiple tasks demonstrate that DPPO substantially improves training stability and sample efficiency, offering a more robust optimization framework for reinforcement learning–based fine-tuning of large language models.

1 citationsRead paper

Revisiting Parameter Server in LLM Post-Training

Jan 27, 2026

This work addresses the synchronization bottleneck in Fully Sharded Data Parallel (FSDP) during large language model (LLM) post-training, where variable sequence lengths cause inefficient All-Gather and Reduce-Scatter operations, degrading device utilization. To overcome this, the authors introduce a parameter server paradigm into FSDP and propose an On-Demand Communication (ODC) mechanism that replaces collective communication with point-to-point exchanges. ODC enables only one synchronization per mini-batch and supports micro-batch-level dynamic load balancing. Implemented on top of PyTorch FSDP with integrated dynamic load distribution and communication scheduling, ODC significantly improves device utilization and training throughput across diverse LLM post-training tasks, achieving up to a 36% speedup over standard FSDP.

1 citationsRead paper

Recent publications

Latest Papers

Lapis: Laplacian Spiking Attention via First-Spike Timing and Membrane Leakage

Aug 12, 2026

This work addresses the limitation of existing spiking vision Transformers, which adopt dense-network-style dot-product self-attention and thereby neglect the temporal information inherent in spike trains. To overcome this, the authors propose Lapis, a novel attention mechanism that directly incorporates first-spike timing into attention computation for the first time. Specifically, Lapis computes affinity via a Laplacian kernel applied to the L1 distance between first-spike latency vectors of queries and keys, aligning naturally with the membrane potential decay dynamics of leaky integrate-and-fire neurons and enabling multiplication-free, efficient attention. Combined with power-of-two rounding, row-wise normalization reduces to bit-shift operations. The method achieves 96.56% Top-1 accuracy on CIFAR-10 (only 0.53% lower than baseline) and reduces attention-path energy consumption by 14.5× on ImageNet-1K, attaining 83.25% accuracy with 3.28 mJ per image at 6-bit precision.

0 citationsRead paper

Scaling GUI Agents with Visual State Transitions

Jul 27, 2026

This work addresses the limited generalization of GUI agents in cross-platform tasks by proposing State Transition Pretraining (STP), which introduces state transition modeling as a novel dimension for GUI understanding. STP jointly optimizes inverse and forward dynamics to continually pretrain a unified multimodal model on large-scale visual state transition data, thereby learning action-anchored visual representations and a dynamic world model of GUIs. When combined with instruction fine-tuning, STP significantly outperforms baseline methods that rely solely on trajectory fine-tuning across multiple benchmarks—including AgentNetBench, AndroidControl, and GUIOdyssey—and demonstrates consistent performance gains with increasing data scale, confirming its effectiveness and scalability.

0 citationsRead paper

LEAP: Trajectory-Level Evaluation of LLMs in Iterative Scientific Design

May 14, 2026

This work addresses a critical limitation in existing scientific design benchmarks, which evaluate large language models (LLMs) solely based on final performance after a fixed number of iterations while ignoring the optimization trajectory. To remedy this, the authors propose LEAPBench, a novel evaluation framework that introduces trajectory efficiency as a core assessment dimension. LEAPBench incorporates area-under-the-curve (AUC) of the optimization trajectory as a metric, establishes Bayesian optimization baselines, integrates literature-anchored validation, and leverages trajectory-based scores as trainable reward signals in offline reinforcement learning. Experiments across 55 tasks reveal that trajectory-aware evaluation alters the ranking of best-performing models in 53% of cases; offline reinforcement learning improves performance on 14 out of 21 held-out tasks; and overall, LLMs fail to surpass classical Bayesian optimization baselines. This study underscores the inadequacy of endpoint-only evaluation and establishes a more reliable paradigm for iterative scientific design.

0 citationsRead paper

Scalable Token-Level Hallucination Detection in Large Language Models

May 12, 2026

This work addresses the challenge of fine-grained hallucinations in large language models during complex reasoning, a problem inadequately tackled by existing detection methods due to their reliance on coarse-grained step segmentation and limited scalability. The authors propose TokenHD, a novel framework that enables token-level hallucination detection without requiring predefined step boundaries. TokenHD leverages a scalable hallucination data synthesis engine to generate large-scale annotated training data and employs an importance-weighted training strategy to end-to-end identify hallucinated tokens in free-form text. Remarkably, a lightweight detector with only 0.6B parameters outperforms reasoning models as large as 32B and demonstrates consistently improved performance across model scales from 0.6B to 8B. The framework exhibits strong generalization and scalability across diverse domains.

0 citationsRead paper

ShiftLIF: Efficient Multi-Level Spiking Neurons with Power-of-Two Quantization

May 03, 2026

This work addresses the limited representational capacity of traditional leaky integrate-and-fire (LIF) neurons, which support only binary spikes, and the high computational overhead or reliance on uniform quantization in existing multi-level spiking schemes. The authors propose ShiftLIF, a novel approach that employs logarithmically spaced, power-of-two quantization levels for membrane potentials, enabling fine-grained, low-magnitude multi-level spike encoding. Crucially, ShiftLIF facilitates multiplier-free synaptic computation through bit-shift and accumulation operations. This design achieves a remarkable balance between accuracy and energy efficiency: it matches or exceeds state-of-the-art performance across ten diverse datasets spanning wireless, acoustic, motion, and visual sensing modalities, while maintaining synaptic energy consumption nearly equivalent to that of standard binary LIF neurons.

0 citationsRead paper