Zhonghao Chen

Liu Yu, Zhonghao Chen, Ping Kuang et al.

Object hallucination remains a critical challenge in Large Vision-Language Models (LVLMs), where models generate content inconsistent with visual inputs. Existing language-decoder based mitigation approaches often regulate visual or textual attention independently, overlooking their interaction as two key causal factors. To address this, we propose Owl (Bi-mOdal attention reWeighting for Layer-wise hallucination mitigation), a causally-grounded framework that models hallucination process via a structural causal graph, treating decomposed visual and textual attentions as mediators. We introduce VTACR (Visual-to-Textual Attention Contribution Ratio), a novel metric that quantifies the modality contribution imbalance during decoding. Our analysis reveals that hallucinations frequently occur in low-VTACR scenarios, where textual priors dominate and visual grounding is weakened. To mitigate this, we design a fine-grained attention intervention mechanism that dynamically adjusts token- and layer-wise attention guided by VTACR signals. Finally, we propose a dual-path contrastive decoding strategy: one path emphasizes visually grounded predictions, while the other amplifies hallucinated ones -- letting visual truth shine and hallucination collapse. Experimental results on the POPE and CHAIR benchmarks show that Owl achieves significant hallucination reduction, setting a new SOTA in faithfulness while preserving vision-language understanding capability. Our code is available at https://github.com/CikZ2023/OWL

Jitao Bai, Simiao Zhang, Zhonghao Chen

Focus on Large Language Model based agents should involve more than "human-centered" alignment or application. We argue that more attention should be paid to the agent itself and discuss the potential of establishing tailored social sciences for agents.

Jin Lee, Zhonghao Chen, Xuhang He et al.

In large-scale LLM pre-training systems with 100k+ GPUs, failures become the norm rather than the exception, and restart costs can dominate wall-clock training time. However, existing fault-tolerance mechanisms are largely unprepared for this restart-dominant regime. To address this challenge, we propose SPARe - Stacked Parallelism with Adaptive Reordering - a fault-tolerance framework that masks node failures during gradient synchronization by stacking redundant data shards across parallelism groups and adaptively reordering execution. SPARe achieves availability comparable to traditional replication while maintaining near-constant computation overhead of only 2~3x, even under high redundancy where traditional replication would require linearly inflating overhead. We derive closed-form expressions for endurable failure count and computation overhead, validate them via SimGrid-based discrete-event simulation, and jointly optimize redundancy and checkpointing to minimize time-to-train. At extreme scale with up to 600k GPUs, SPARe reduces time-to-train by 40~50% compared to traditional replication.

Hongmin Gao, Zhonghao Chen, Chenming Li

Recently, convolutional neural network (CNN) techniques have gained popularity as a tool for hyperspectral image classification (HSIC). To improve the feature extraction efficiency of HSIC under the condition of limited samples, the current methods generally use deep models with plenty of layers. However, deep network models are prone to overfitting and gradient vanishing problems when samples are limited. In addition, the spatial resolution decreases severely with deeper depth, which is very detrimental to spatial edge feature extraction. Therefore, this letter proposes a shallow model for HSIC, which is called depthwise over-parameterized convolutional neural network (DOCNN). To ensure the effective extraction of the shallow model, the depthwise over-parameterized convolution (DO-Conv) kernel is introduced to extract the discriminative features. The depthwise over-parameterized Convolution kernel is composed of a standard convolution kernel and a depthwise convolution kernel, which can extract the spatial feature of the different channels individually and fuse the spatial features of the whole channels simultaneously. Moreover, to further reduce the loss of spatial edge features due to the convolution operation, a dense residual connection (DRC) structure is proposed to apply to the feature extraction part of the whole network. Experimental results obtained from three benchmark data sets show that the proposed method outperforms other state-of-the-art methods in terms of classification accuracy and computational efficiency.