Sicheng Yu

Gangjian Zhang, Jian Shu, Sicheng Yu et al.

This paper addresses the challenge of reconstructing photorealistic and animatable 3D human avatars from monocular videos. While existing methods rely on combining per-subject optimization with generic human priors, they often fail to capture fine-grained details when training frames are limited. To mitigate this data scarcity, we propose TrioMan, a systematic tri-module framework for augmented 3D avatar learning. Our approach comprises three synergistic components. The Generator creates diverse unseen samples by imposing Gaussian perturbations on pose and camera. The Refiner improves the quality of generated data through one-step diffusion guided by texture and geometry cues. The Examiner selects subject-consistent samples using a dual-branch attention-based similarity evaluation. Experiments on the X-Humans and NeuMan benchmarks show that TrioMan outperforms state-of-the-art methods.

Kai Liu, Jungang Li, Yuchong Sun et al.

This paper presents JavisGPT, the first unified multimodal large language model (MLLM) for joint audio-video (JAV) comprehension and generation. JavisGPT has a concise encoder-LLM-decoder architecture, which has a SyncFusion module for spatio-temporal audio-video fusion and synchrony-aware learnable queries to bridge a pretrained JAV-DiT generator. This design enables temporally coherent video-audio understanding and generation from multimodal instructions. We design an effective three-stage training pipeline consisting of multimodal pretraining, audio-video fine-tuning, and large-scale instruction-tuning, to progressively build multimodal comprehension and generation from existing vision-language models. For instruction tuning, we construct JavisInst-Omni, a high-quality instruction dataset with over 200K GPT-4o-curated audio-video-text dialogues that cover diverse and multi-level comprehension and generation scenarios. On JAV comprehension and generation benchmarks, our experiments show that JavisGPT outperforms existing MLLMs, particularly in complex and temporally synchronized settings.

Cunxiao Du, Jing Jiang, Xu Yuanchen et al.

Speculative decoding is a relatively new decoding framework that leverages small and efficient draft models to reduce the latency of LLMs. In this study, we introduce GliDe and CaPE, two low-hassle modifications to vanilla speculative decoding to further improve the decoding speed of a frozen LLM. Specifically, GliDe is a modified draft model architecture that reuses the cached keys and values from the target LLM, while CaPE is a proposal expansion method that uses the draft model's confidence scores to help select additional candidate tokens for verification. Extensive experiments on different benchmarks demonstrate that our proposed GliDe draft model significantly reduces the expected decoding latency. Additional evaluation using walltime reveals that GliDe can accelerate Vicuna models up to 2.17x and further extend the improvement to 2.61x with CaPE. We will release our code, data, and the trained draft models.

Beizhen Zhao, Sicheng Yu, Ziran Yin et al.

While 3D Gaussian Splatting (3DGS) has revolutionized 3D reconstruction, it suffers from significant overhead due to massive redundant primitives. Existing compression methods typically rely on local sampling or fixed pruning thresholds, which often struggle to balance redundancy reduction with high-fidelity rendering. To address this, we propose a novel framework that formulates Gaussian optimization as a global geometric distribution matching problem. Specifically, our approach integrates three components: (1) we introduce a multi-view 3D Gaussian contribution ranking mechanism that filters primitives using geometric consistency instead of local heuristics; (2) we propose a global Optimal Transport (OT)-based aggregation algorithm that merges redundant primitives while preserving the underlying geometry; and (3) we design an OT-based densification operator that maintains the Gaussian's distributional properties for stable optimization. Our approach achieves state-of-the-art rendering quality with only \textbf{10$\%$} primitives and \textbf{10$\times$} accelerated training speeds compared to vanilla 3DGS.

Beizhen Zhao, Sicheng Yu, Guanzhi Ding et al.

3D scene reconstruction under unposed sparse viewpoints is a highly challenging yet practically important problem, especially in outdoor scenes due to complex lighting and scale variation. With extremely limited input views, directly utilizing diffusion model to synthesize pseudo frames will introduce unreasonable geometry, which will harm the final reconstruction quality. To address these issues, we propose a novel framework for sparse-view outdoor reconstruction that achieves high-quality results through bidirectional pseudo frame restoration and scene perception Gaussian management. Specifically, we introduce a bidirectional pseudo frame restoration method that restores missing content by diffusion-based synthesis guided by adjacent frames with a lightweight pseudo-view deblur model and confidence mask inference algorithm. Then we propose a scene perception Gaussian management strategy that optimize Gaussians based on joint depth-density information. These designs significantly enhance reconstruction completeness, suppress floating artifacts and improve overall geometric consistency under extreme view sparsity. Experiments on outdoor benchmarks demonstrate substantial gains over existing methods in both fidelity and stability.

Xingshu Chen, Sicheng Yu, Chong Cheng et al.

This paper investigates the problem of object detection with a focus on improving both the localization accuracy of bounding boxes and explicitly modeling prediction uncertainty. Conventional detectors rely on deterministic bounding box regression, ignoring uncertainty in predictions and limiting model robustness. In this paper, we propose an uncertainty-aware enhancement framework for DETR-based object detectors. We model bounding boxes as multivariate Gaussian distributions and incorporate the Gromov-Wasserstein distance into the loss function to better align the predicted and ground-truth distributions. Building on this, we derive a Bayes Risk formulation to filter high-risk information and improve detection reliability. We also propose a simple algorithm to quantify localization uncertainty via confidence intervals. Experiments on the COCO benchmark show that our method can be effectively integrated into existing DETR variants, enhancing their performance. We further extend our framework to leukocyte detection tasks, achieving state-of-the-art results on the LISC and WBCDD datasets. These results confirm the scalability of our framework across both general and domain-specific detection tasks. Code page: https://github.com/ParadiseforAndaChen/An-Uncertainty-aware-DETR-Enhancement-Framework-for-Object-Detection.

Sicheng Yu, Chong Cheng, Yifan Zhou et al.

3D Gaussian Splatting (3DGS) has become a popular solution in SLAM, as it can produce high-fidelity novel views. However, previous GS-based methods primarily target indoor scenes and rely on RGB-D sensors or pre-trained depth estimation models, hence underperforming in outdoor scenarios. To address this issue, we propose a RGB-only gaussian splatting SLAM method for unbounded outdoor scenes--OpenGS-SLAM. Technically, we first employ a pointmap regression network to generate consistent pointmaps between frames for pose estimation. Compared to commonly used depth maps, pointmaps include spatial relationships and scene geometry across multiple views, enabling robust camera pose estimation. Then, we propose integrating the estimated camera poses with 3DGS rendering as an end-to-end differentiable pipeline. Our method achieves simultaneous optimization of camera poses and 3DGS scene parameters, significantly enhancing system tracking accuracy. Specifically, we also design an adaptive scale mapper for the pointmap regression network, which provides more accurate pointmap mapping to the 3DGS map representation. Our experiments on the Waymo dataset demonstrate that OpenGS-SLAM reduces tracking error to 9.8\% of previous 3DGS methods, and achieves state-of-the-art results in novel view synthesis. Project Page: https://3dagentworld.github.io/opengs-slam/

Chong Cheng, Sicheng Yu, Zijian Wang et al.

3D Gaussian Splatting (3DGS) has become a popular solution in SLAM due to its high-fidelity and real-time novel view synthesis performance. However, some previous 3DGS SLAM methods employ a differentiable rendering pipeline for tracking, lack geometric priors in outdoor scenes. Other approaches introduce separate tracking modules, but they accumulate errors with significant camera movement, leading to scale drift. To address these challenges, we propose a robust RGB-only outdoor 3DGS SLAM method: S3PO-GS. Technically, we establish a self-consistent tracking module anchored in the 3DGS pointmap, which avoids cumulative scale drift and achieves more precise and robust tracking with fewer iterations. Additionally, we design a patch-based pointmap dynamic mapping module, which introduces geometric priors while avoiding scale ambiguity. This significantly enhances tracking accuracy and the quality of scene reconstruction, making it particularly suitable for complex outdoor environments. Our experiments on the Waymo, KITTI, and DL3DV datasets demonstrate that S3PO-GS achieves state-of-the-art results in novel view synthesis and outperforms other 3DGS SLAM methods in tracking accuracy. Project page: https://3dagentworld.github.io/S3PO-GS/.

Xiongwei Wu, Sicheng Yu, Ee-Peng Lim et al.

In the realm of food computing, segmenting ingredients from images poses substantial challenges due to the large intra-class variance among the same ingredients, the emergence of new ingredients, and the high annotation costs associated with large food segmentation datasets. Existing approaches primarily utilize a closed-vocabulary and static text embeddings setting. These methods often fall short in effectively handling the ingredients, particularly new and diverse ones. In response to these limitations, we introduce OVFoodSeg, a framework that adopts an open-vocabulary setting and enhances text embeddings with visual context. By integrating vision-language models (VLMs), our approach enriches text embedding with image-specific information through two innovative modules, eg, an image-to-text learner FoodLearner and an Image-Informed Text Encoder. The training process of OVFoodSeg is divided into two stages: the pre-training of FoodLearner and the subsequent learning phase for segmentation. The pre-training phase equips FoodLearner with the capability to align visual information with corresponding textual representations that are specifically related to food, while the second phase adapts both the FoodLearner and the Image-Informed Text Encoder for the segmentation task. By addressing the deficiencies of previous models, OVFoodSeg demonstrates a significant improvement, achieving an 4.9\% increase in mean Intersection over Union (mIoU) on the FoodSeg103 dataset, setting a new milestone for food image segmentation.

Sicheng Yu, Yuanchen Xu, Cunxiao Du et al.

Humans are accustomed to reading and writing in a forward manner, and this natural bias extends to text understanding in auto-regressive large language models (LLMs). This paper investigates whether LLMs, like humans, struggle with reverse modeling, specifically with reversed text inputs. We found that publicly available pre-trained LLMs cannot understand such inputs. However, LLMs trained from scratch with both forward and reverse texts can understand them equally well during inference across multiple languages. Our case study shows that different-content texts result in different losses if input (to LLMs) in different directions -- some get lower losses for forward while some for reverse. This leads us to a simple and nice solution for data selection based on the loss differences between forward and reverse directions. Using our selected data in continued pretraining can boost LLMs' performance by a large margin across different language understanding benchmarks.

Yuan Zhang, Chun-Kai Fan, Tao Huang et al.

Inspired by text prompts in large language models (LLMs), visual prompts have been explored to enhance the reasoning capabilities of large vision-language models (LVLMs). Current methods design heuristic visual prompts, such as overlaying a text-query-guided attention heatmap on the original input image. However, designing effective prompts manually is challenging and time-consuming, and it often fails to explore the benefits of different visual prompts, leading to sub-optimal performance. To this end, we propose \textbf{AutoV} that learns to automatically select the optimal visual prompt from various candidates based on given textual queries and the input image. To train AutoV, we developed an automatic data collection and labeling pipeline that evaluates various visual prompts with a pre-trained LVLM. We input a set of visual prompts into the LVLM and rank them according to the prediction losses generated by the model. Using the ranking as a supervision signal, we train AutoV to automatically choose the optimal visual prompt from various visual prompts for LVLMs. Experimental results indicate that AutoV enhances the performance of various LVLMs across multiple popular image understanding tasks. For instance, LLaVA-OV with AutoV achieves $\textbf{1.7}\%$ accuracy gain on LLaVA$^{\text{Wild}}$, and AutoV boosts Qwen2.5-VL by $\textbf{1.9}\%$ on MMMU, highlighting its potential as an optimal visual prompting method for LVLMs.

Zhe Ling, Sicheng Yu, Danyu Yang

Online handwritten character recognition leverages stroke order and dynamic features, which generally provide higher accuracy and robustness compared with offline recognition. However, in practical applications, rotational deformations can disrupt the spatial layout of strokes, substantially reducing recognition accuracy. Extracting rotation-invariant features therefore remains a challenging open problem. In this work, we employ the Sliding Window Path Signature (SW-PS) to capture local structural features of characters, and introduce the lightweight Linear Recurrent Units (LRU) as the classifier. The LRU combine the fast incremental processing capability of recurrent neural networks (RNN) with the efficient parallel training of state space models (SSM), while reliably modelling dynamic stroke characteristics. We conducted recognition experiments with random rotation angle up to $\pm 180^{\circ}$ on three subsets of the CASIA-OLHWDB1.1 dataset: digits, English upper letters, and Chinese radicals. The accuracies achieved after ensemble learning were $99.62\%$, $96.67\%$, and $94.33\%$, respectively. Experimental results demonstrate that the proposed SW-PS+LRU framework consistently surpasses competing models in both convergence speed and test accuracy.

Yu Hu, Chong Cheng, Sicheng Yu et al.

Reconstructing dynamic 4D scenes is challenging, as it requires robust disentanglement of dynamic objects from the static background. While 3D foundation models like VGGT provide accurate 3D geometry, their performance drops markedly when moving objects dominate. Existing 4D approaches often rely on external priors, heavy post-optimization, or require fine-tuning on 4D datasets. In this paper, we propose VGGT4D, a training-free framework that extends the 3D foundation model VGGT for robust 4D scene reconstruction. Our approach is motivated by the key finding that VGGT's global attention layers already implicitly encode rich, layer-wise dynamic cues. To obtain masks that decouple static and dynamic elements, we mine and amplify global dynamic cues via gram similarity and aggregate them across a temporal window. To further sharpen mask boundaries, we introduce a refinement strategy driven by projection gradient. We then integrate these precise masks into VGGT's early-stage inference, effectively mitigating motion interference in both pose estimation and geometric reconstruction. Across six datasets, our method achieves superior performance in dynamic object segmentation, camera pose estimation, and dense reconstruction. It also supports single-pass inference on sequences longer than 500 frames.

Wei Han, Geng Zhan, Sicheng Yu et al.

O1/R1 style large reasoning models (LRMs) signal a substantial leap forward over conventional instruction-following LLMs. By applying test-time scaling to generate extended reasoning paths, they establish many SOTAs across a wide range of complex reasoning tasks. However, recent studies show that LRMs are prone to suffer from overthinking -- the tendency to overcomplicate simple problems, leading to excessive strategy switching and long, convoluted reasoning traces that hinder their interpretability. To mitigate this issue, we conduct a systematic investigation into the reasoning efficiency of a broad set of LRMs and uncover a common dilemma: the difficulty in balancing multiple generation objectives such as correctness and brevity. Based on this discovery, we propose a test-time scaling method, EDIT (Efficient Dynamic Inference Trimming), which efficiently guides LRMs to identify the shortest correct reasoning paths at test time. EDIT employs constraint-guided generation while jointly tracking length and answer distributions under varying constraints, allowing it to select responses that strike an optimal balance between conciseness and correctness. Extensive experiments across diverse models and datasets show that EDIT substantially enhance the reasoning efficiency, producing compact yet informative outputs that improve readability and user experience.

Chong Cheng, Zijian Wang, Sicheng Yu et al.

3D Gaussian Splatting (3DGS) has emerged as a core technique for 3D representation. Its effectiveness largely depends on precise camera poses and accurate point cloud initialization, which are often derived from pretrained Multi-View Stereo (MVS) models. However, in unposed reconstruction task from hundreds of outdoor images, existing MVS models may struggle with memory limits and lose accuracy as the number of input images grows. To address this limitation, we propose a novel unposed 3DGS reconstruction framework that integrates pretrained MVS priors with the probabilistic Procrustes mapping strategy. The method partitions input images into subsets, maps submaps into a global space, and jointly optimizes geometry and poses with 3DGS. Technically, we formulate the mapping of tens of millions of point clouds as a probabilistic Procrustes problem and solve a closed-form alignment. By employing probabilistic coupling along with a soft dustbin mechanism to reject uncertain correspondences, our method globally aligns point clouds and poses within minutes across hundreds of images. Moreover, we propose a joint optimization framework for 3DGS and camera poses. It constructs Gaussians from confidence-aware anchor points and integrates 3DGS differentiable rendering with an analytical Jacobian to jointly refine scene and poses, enabling accurate reconstruction and pose estimation. Experiments on Waymo and KITTI datasets show that our method achieves accurate reconstruction from unposed image sequences, setting a new state of the art for unposed 3DGS reconstruction.

Chong Cheng, Yu Hu, Sicheng Yu et al.

3D Gaussian Splatting (3DGS) has demonstrated its potential in reconstructing scenes from unposed images. However, optimization-based 3DGS methods struggle with sparse views due to limited prior knowledge. Meanwhile, feed-forward Gaussian approaches are constrained by input formats, making it challenging to incorporate more input views. To address these challenges, we propose RegGS, a 3D Gaussian registration-based framework for reconstructing unposed sparse views. RegGS aligns local 3D Gaussians generated by a feed-forward network into a globally consistent 3D Gaussian representation. Technically, we implement an entropy-regularized Sinkhorn algorithm to efficiently solve the optimal transport Mixture 2-Wasserstein $(\text{MW}_2)$ distance, which serves as an alignment metric for Gaussian mixture models (GMMs) in $\mathrm{Sim}(3)$ space. Furthermore, we design a joint 3DGS registration module that integrates the $\text{MW}_2$ distance, photometric consistency, and depth geometry. This enables a coarse-to-fine registration process while accurately estimating camera poses and aligning the scene. Experiments on the RE10K and ACID datasets demonstrate that RegGS effectively registers local Gaussians with high fidelity, achieving precise pose estimation and high-quality novel-view synthesis. Project page: https://3dagentworld.github.io/reggs/.

Fengyun Wang, Sicheng Yu, Jiawei Wu et al.

Large vision-language models (LVLMs) have significantly advanced numerous fields. In this work, we explore how to harness their potential to address 3D scene understanding tasks, using 3D question answering (3D-QA) as a representative example. Due to the limited training data in 3D, we do not train LVLMs but infer in a zero-shot manner. Specifically, we sample 2D views from a 3D point cloud and feed them into 2D models to answer a given question. When the 2D model is chosen, e.g., LLAVA-OV, the quality of sampled views matters the most. We propose cdViews, a novel approach to automatically selecting critical and diverse Views for 3D-QA. cdViews consists of two key components: viewSelector prioritizing critical views based on their potential to provide answer-specific information, and viewNMS enhancing diversity by removing redundant views based on spatial overlap. We evaluate cdViews on the widely-used ScanQA and SQA benchmarks, demonstrating that it achieves state-of-the-art performance in 3D-QA while relying solely on 2D models without fine-tuning. These findings support our belief that 2D LVLMs are currently the most effective alternative (of the resource-intensive 3D LVLMs) for addressing 3D tasks.

Xuan Zhang, Cunxiao Du, Sicheng Yu et al.

Due to the auto-regressive nature of current video large language models (Video-LLMs), the inference latency increases as the input sequence length grows, posing challenges for the efficient processing of video sequences that are usually very long. We observe that during decoding, the attention scores of most tokens in Video-LLMs tend to be sparse and concentrated, with only certain tokens requiring comprehensive full attention. Based on this insight, we introduce Sparse-to-Dense (StD), a novel decoding strategy that integrates two distinct modules: one leveraging sparse top-K attention and the other employing dense full attention. These modules collaborate to accelerate Video-LLMs without loss. The fast (sparse) model speculatively decodes multiple tokens, while the slow (dense) model verifies them in parallel. StD is a tuning-free, plug-and-play solution that achieves up to a 1.94$\times$ walltime speedup in video processing. It maintains model performance while enabling a seamless transition from a standard Video-LLM to a sparse Video-LLM with minimal code modifications.

Qiyuan Zhang, Lei Wang, Sicheng Yu et al.

While diverse question answering (QA) datasets have been proposed and contributed significantly to the development of deep learning models for QA tasks, the existing datasets fall short in two aspects. First, we lack QA datasets covering complex questions that involve answers as well as the reasoning processes to get the answers. As a result, the state-of-the-art QA research on numerical reasoning still focuses on simple calculations and does not provide the mathematical expressions or evidences justifying the answers. Second, the QA community has contributed much effort to improving the interpretability of QA models. However, these models fail to explicitly show the reasoning process, such as the evidence order for reasoning and the interactions between different pieces of evidence. To address the above shortcomings, we introduce NOAHQA, a conversational and bilingual QA dataset with questions requiring numerical reasoning with compound mathematical expressions. With NOAHQA, we develop an interpretable reasoning graph as well as the appropriate evaluation metric to measure the answer quality. We evaluate the state-of-the-art QA models trained using existing QA datasets on NOAHQA and show that the best among them can only achieve 55.5 exact match scores, while the human performance is 89.7. We also present a new QA model for generating a reasoning graph where the reasoning graph metric still has a large gap compared with that of humans, e.g., 28 scores.

Sicheng Yu, Yulei Niu, Shuohang Wang et al.

Deep neural network based question answering (QA) models are neither robust nor explainable in many cases. For example, a multiple-choice QA model, tested without any input of question, is surprisingly "capable" to predict the most of correct options. In this paper, we inspect such spurious "capability" of QA models using causal inference. We find the crux is the shortcut correlation, e.g., unrobust word alignment between passage and options learned by the models. We propose a novel approach called Counterfactual Variable Control (CVC) that explicitly mitigates any shortcut correlation and preserves the comprehensive reasoning for robust QA. Specifically, we leverage multi-branch architecture that allows us to disentangle robust and shortcut correlations in the training process of QA. We then conduct two novel CVC inference methods (on trained models) to capture the effect of comprehensive reasoning as the final prediction. For evaluation, we conduct extensive experiments using two BERT backbones on both multi-choice and span-extraction QA benchmarks. The results show that our CVC achieves high robustness against a variety of adversarial attacks in QA while maintaining good interpretation ability.

Sicheng Yu, Hao Zhang, Wei Jing et al.

Multiple-Choice Question Answering (MCQA) is a challenging task in machine reading comprehension. The main challenge in MCQA is to extract "evidence" from the given context that supports the correct answer. In the OpenbookQA dataset, the requirement of extracting "evidence" is particularly important due to the mutual independence of sentences in the context. Existing work tackles this problem by annotated evidence or distant supervision with rules which overly rely on human efforts. To address the challenge, we propose a simple yet effective approach termed evidence filtering to model the relationships between the encoded contexts with respect to different options collectively and to potentially highlight the evidence sentences and filter out unrelated sentences. In addition to the effective reduction of human efforts of our approach compared, through extensive experiments on OpenbookQA, we show that the proposed approach outperforms the models that use the same backbone and more training data; and our parameter analysis also demonstrates the interpretability of our approach.