arXiv subjects

Wanhua Li

Publications and source records attributed to Wanhua Li.

At least 19 recordsLinked to original sources

arXiv2026

PixelBench: Can LLMs Understand and Generate Images Pixel by Pixel?

Large language models (LLMs) can generate visual assets through text-based formats such as SVG and CAD, but these symbolic representations describe images using high-level primitives rather than explicit raster values. We ask whether LLMs can instead generate pixel images directly through their existing vocabulary, without a learned image tokenizer or decoder. We introduce PixelBench, a benchmark for text-to-pixel (T2P) and image-to-pixel (I2P) generation across multiple resolutions, together with a complementary pixel-understanding evaluation. Models generate an RGB color palette followed by a grid of indices specifying the palette color of each output pixel. In T2P, the strongest models produce recognizable images directly from text. In I2P, the strongest models outperform palette-matched downsampling and specialized pixelization algorithms at low resolutions. Pixel-understanding performance correlates strongly with generation performance. A resolution-scaling study shows that pixel-generation performance often improves initially before declining at high resolutions. Finally, we introduce Pixel Reconstruction Tuning (PRT), which augments visual instruction tuning with caption-palette-grid trajectories. We post-train Molmo2-4B from its pretraining checkpoint by mixing general instruction data with our curated PRT data, improving performance across 16 visual-centric benchmarks, indicating that explicit pixel reconstruction can provide useful auxiliary supervision.

cs.CV↗

arXiv2026

Benchmarking Vision-Language Models on Synapse and Neuron Proofreading in Connectomics

We benchmarked vision-language models (VLMs) on the decisions annotators take when inspecting electron microscopy images in connectomics: synapse proofreading (presence and polarity) and neuron proofreading (split errors and merge errors). For synapse proofreading, we evaluated 19 open and 2 closed models across various architectures and sizes under zero-shot, four-shot in-context learning and LoRA settings, against specialist models, on datasets constructed by us using public resources. For neuron proofreading, we evaluated 3 open and 2 closed models on the ConnectomeBench2 dataset, with cross-species transfer from fly and mouse to human and zebrafish. Most models were at chance zero-shot; a few examples helped mainly the closed and largest open ones. LoRA on a few thousand labels brought open models level with specialist models. When evaluated on unseen species, the best adapted VLMs outperformed specialist models in identifying merge errors. The project will be publicly available upon acceptance.

cs.CV↗

arXiv2026

Refuse the Harm, Complete the Task via Graph-of-Rules Planning for LLM-Enabled Robots

Large Language Models (LLMs) are increasingly used in embodied AI for robotic manipulation and planning, but their deployment in physical systems exposes vulnerabilities such as jailbreak attacks that induce unsafe behavior. Existing inference-time defenses are mainly evaluated on short instructions containing a single hazard, whereas realistic tasks may interleave safe and unsafe sub-tasks, conceal hazardous steps, or combine individually safe actions into unsafe outcomes. Such compound instructions make it harder for a single safety decision to distinguish hazardous sub-tasks from legitimate ones, risking either overlooking the hazard or rejecting the whole instruction. We introduce safe task completion, a task that requires an LLM-enabled robot to refuse the harmful part of an instruction while completing its legitimate part. To address this problem, we propose RuleGraph, a multi-agent Graph-of-Rules planning framework that distributes safety enforcement across multiple reasoning and verification stages. Five cooperating agents are used to ground general safety principles into world-specific rules, assess the instruction's sub-tasks in context, and derive a safe goal from the legitimate ones. Planning toward this goal is then organized as a graph of rule-specific nodes, where higher-level nodes inherit the union of their parents' rules, and candidate plans are progressively verified and aggregated until a single plan satisfies every relevant rule. In this way, RuleGraph reduces reliance on a single centralized safety decision. We evaluate RuleGraph on pure-attack and compound instructions across two benchmarks in two environments using Attack Success Rate (ASR) and Task Success Rate (TSR). RuleGraph reduces ASR to 0% on both benchmarks while achieving the highest TSR among all methods, reaching 98.57% on the RoboGuard benchmark and 84.54% on our long-horizon compound instructions.

cs.RO↗

arXiv2026

Who Improves the Robot? Toward Physical Recursive Self-Improvement

Robots are increasingly autonomous in task execution, yet their improvement remains largely human-driven. This study investigates Physical Recursive Self-Improvement (Physical RSI), in which physical experience guides lasting updates to both a robot's capabilities and its improvement procedure, making subsequent improvement more effective across tasks. Physical RSI faces challenges from noisy observations, continuous dynamics, costly trials, and irreversible consequences. We present an L0-L5 framework that differentiates capability-level from meta-level updates and within-task from across-task gains, reserving L5 for independently verified recursive gain. We study existing methods around harness adaptation and model training, alongside safety and control mechanisms that govern execution and successive updates. Our analysis reveals that, despite substantial progress in self-improvement, the surveyed literature does not yet demonstrate Physical RSI as defined by our framework. We identify research priorities in reliable feedback, causal attribution, safety preservation, and benchmarking, outlining a path toward robots that autonomously improve how they learn while ensuring safety. The GitHub repository for Physical RSI is available at https://github.com/SAIL-Research-Lab/Physical-RSI.

cs.AI↗

arXiv2026

RoboFoundry: System-as-Policy Evolution for Self-Learning Embodied Agents

A foundation model should not act in isolation as an embodied agent. Yet, existing methods often optimize individual components of the agent stack, such as memory, context, skills, or action interfaces, rather than treating the supporting system itself as a unified policy. Moreover, interaction alone does not yield self-improvement unless execution experience is converted into persistent, validated system changes. We therefore propose RoboFoundry, the first embodied agentic framework that formulates this process as Self-Evolving System-as-Policy. RoboFoundry diagnoses capability gaps in decision-making and memory management, converts execution traces into validated task-specific system updates, and promotes recurring improvements to the general system. Evolution operates over two complementary surfaces: a context system that manages active internal context and persistent file-system memory, and a hierarchical skill system that organizes atomic skills, reusable compositions, and failure-conditioned recovery. A shared semantic interface separates embodiment-invariant decisions from embodiment-specific execution, allowing evolved system capabilities to transfer across heterogeneous robots. On EmbodiedBench, RoboFoundry achieves state-of-the-art performance, notably improving GPT-5.5 by 27.8%. It also brings Qwen3.7-Plus to near parity with GPT-5.5 (70.3% vs. 72.7%), showing consistent gains from system-as-policy evolution across foundation models. For long-horizon memory, RoboFoundry outperforms all baselines on RoboMemArena by at least 39.0%, even against methods assisted by external foundation models. On LIBERO-PRO, it further outperforms Cap-Agent0 by 243.8%-679.7% across all perturbation types. In real-world deployments, RoboFoundry demonstrates zero-shot transfer and online evolution across robots and tasks, highlighting its potential for fully autonomous embodied agents.

cs.RO↗

arXiv2026

Rewarding Reasoning, Not Answers: Fixing and Bounding Test-Time Reinforcement Learning on Medical QA

Test-time reinforcement learning adapts a model on its own unlabeled test set using majority-vote pseudo-labels and has shown strong results in mathematics. We show that this recipe collapses on medical multiple-choice QA: accuracy stagnates while output diversity rapidly declines. Through a controlled experiment that keeps the questions, model, and optimizer fixed while changing only the answer space, we trace this failure to answer-space structure rather than domain difficulty. In small answer spaces, incorrect rollouts often collide on the same wrong pseudo-label and reinforce it; in large answer spaces, they disperse and receive little reward. This diagnosis motivates PROSE, Process Reward Guided Self-Training, which rewards reasoning quality instead of answer agreement. PROSE scores each reasoning step with a medical process reward model, assigns the trajectory reward as the minimum score across steps, and enforces answer-format constraints. Without labels, PROSE substantially improves a general Llama model, surpassing purpose-built medical models and matching much larger systems. Because the process signal is internalized into the policy, the adapted model requires no reward model at inference and transfers its gains to unseen datasets. We further show that the minimum aggregation is essential: mean aggregation can be exploited, saturating the proxy reward while degrading accuracy.

cs.CL↗

arXiv2026

MAC-Splat: Multi-Attribute Consistency for High-Fidelity Sparse-View Reconstruction

Reconstructing high-fidelity 3D scenes from sparse-views remains a central problem in generalizable neural rendering. Existing generalizable 3D Gaussian Splatting (3DGS) methods often exhibit geometric artifacts in sparse-view settings, since supervision based solely on 2D photometric losses cannot resolve depth and correspondence ambiguities. To address this issue, we propose MAC-Splat, a training framework built around direct 3D consistency supervision. MAC-Splat builds on the MASt3R geometric backbone and a frozen DINOv3 encoder to obtain semantically informed 2D correspondences, which serve as geometric anchors for 3D supervision. Using these anchors, we define the Multi-Attribute Consistency (MAC) loss. This objective jointly regularizes the 3D attributes of matched Gaussians, including their position, shape, and appearance, by enforcing agreement in a common world coordinate frame. The formulation is robust to outliers and respects the geometry of covariance matrices, which leads to stable training under sparse-view conditions. Experiments on ScanNet++ show that MAC-Splat outperforms strong baselines, with particularly large gains under different overlap regimes. In particular, it improves average PSNR over Splatt3R by more than 4.5 dB, reduces LPIPS, and maintains performance as the camera pose gap increases. These results indicate that a direct, multi-attribute 3D consistency objective, when combined with high-quality correspondences, is effective for addressing the ill-posed sparse-view reconstruction problem.

cs.CV↗

arXiv2026

DETRAM: End-to-end DEtection, Tracking and Recovery of HumAn Meshes

In the task of human mesh recovery (HMR), multi-person scenes are particularly difficult to handle due to the many entities that appear and occlusions between them over time. In particular for video inputs, there is a need to track each entity reliably and consistently. Existing methods rely on pretrained human detection modules, increasing their runtime and limiting the number of tracked entities. We present DETRAM, a unified framework for multi-person HMR and tracking that simultaneously detects, reconstructs, and tracks humans across time, both automatically and via user prompts. DETRAM uses a single transformer decoder with an identity-consistent set of learnable query embeddings that persist across frames: detection queries discover new people, tracking queries maintain pose and shape for existing individuals, and prompt queries follow user-specified identities. Our approach achieves state-of-the-art tracking results on PoseTrack21, 3DPW, BEDLAM, and MuPoTS-3D, and competitive reconstruction accuracy on BEDLAM and 3DPW, while uniquely supporting prompt-based tracking of individuals in multi-person scenes. To our knowledge, this is the first method to unify promptability and multi-person HMR with tracking in an end-to-end trainable framework, enabling user-directed human analysis in videos.

cs.CV↗

arXiv2026

LangFlash: Feed-forward 3D Language Gaussian Splatting from Sparse Unposed Images

We present LangFlash, a feed-forward framework for 3D Language Gaussian Splatting that reconstructs 3D scenes parameterized by Gaussian primitives enriched with language-aligned semantic features from sparse unposed multi-view images. Unlike optimization-based 3D methods, LangFlash directly predicts the geometry and semantics in a single forward pass, enabling low-latency 3D reconstruction and language-consistent scene understanding. To support large-scale training, we enriched the RealEstate10k dataset with coherent and dense semantic information for 3D semantic supervision. Furthermore, we propose a sparse semantic encoding scheme that combines a global semantic dictionary with locally varying per-primitive weights, preserving high-level linguistic information, while reducing representation complexity. Experimental results show that LangFlash achieves superior novel view synthesis and semantic consistency compared with previous methods. This study establishes a new paradigm for pose-free, language-grounded 3D scene reconstruction, advancing generalizable 3D vision and multimodal scene understanding. Demo is available at https://liylo.github.io/langflash.github.io/.

cs.CV↗

arXiv2026

RiGS: Rigid-aware 4D Gaussian Splatting from a Single Monocular Video

Reconstructing dynamic 3D scenes from monocular videos is a fundamental yet highly challenging task, as real-world motions often involve both long-term smooth transformations and short-term complex deformations. Existing methods either struggle to maintain temporal consistency or fail to capture high-frequency dynamics due to limited motion modeling capacity. In this work, we present Rigid-aware 4D Gaussian Splatting (RiGS), which simultaneously captures motions across multiple temporal scales. Specifically, RiGS introduces three types of Gaussian primitives: static, rigid, and transient, which represent static backgrounds, long-term low-frequency motions, and short-term high-frequency dynamics, respectively. An object-wise dynamic mask is proposed to aggregate long-range spatiotemporal motion information and guide the decomposition of static and dynamic regions. To jointly model motion across scales, rigid Gaussians are allowed to transition into transient Gaussians based on their temporal duration, and both are optimized under scene flow guidance, providing dense 3D motion supervision. Extensive experiments demonstrate that RiGS achieves state-of-the-art performance on novel view synthesis benchmarks. Code is available at \hyperlink{https://github.com/ladvu/RiGS}{https://github.com/ladvu/RiGS}.

cs.CV↗

arXiv2026

Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria

Aligning multimodal generative models with human preferences demands reward signals that respect the compositional, multi-dimensional structure of human judgment. Prevailing RLHF approaches reduce this structure to scalar or pairwise labels, collapsing nuanced preferences into opaque parametric proxies and exposing vulnerabilities to reward hacking. While recent Rubrics-as-Reward (RaR) methods attempt to recover this structure through explicit criteria, generating rubrics that are simultaneously reliable, scalable, and data-efficient remains an open problem. We introduce Auto-Rubric as Reward (ARR), a framework that reframes reward modeling from implicit weight optimization to explicit, criteria-based decomposition. Before any pairwise comparison, ARR externalizes a VLM's internalized preference knowledge as prompt-specific rubrics, translating holistic intent into independently verifiable quality dimensions. This conversion of implicit preference structure into inspectable, interpretable constraints substantially suppresses evaluation biases including positional bias, enabling both zero-shot deployment and few-shot conditioning on minimal supervision. To extend these gains into generative training, we propose Rubric Policy Optimization (RPO), which distills ARR's structured multi-dimensional evaluation into a robust binary reward, replacing opaque scalar regression with rubric-conditioned preference decisions that stabilize policy gradients. On text-to-image generation and image editing benchmarks, ARR-RPO outperforms pairwise reward models and VLM judges, demonstrating that explicitly externalizing implicit preference knowledge into structured rubrics achieves more reliable, data-efficient multimodal alignment, revealing that the bottleneck is the absence of a factorized interface, not a deficit of knowledge.

cs.AI↗

arXiv2026

RoboTAG: End-to-end Robot Configuration Estimation via Topological Alignment Graph

Estimating robot pose from a monocular RGB image is a challenge in robotics and computer vision. Existing methods typically build networks on top of 2D visual backbones and depend heavily on labeled data for training, which is often scarce in real-world scenarios, causing a sim-to-real gap. Moreover, these approaches reduce the 3D-based problem to 2D domain, neglecting the 3D priors. To address these, we propose Robot Topological Alignment Graph (RoboTAG), which incorporates a 3D branch to inject 3D priors while enabling co-evolution of the 2D and 3D representations, alleviating the reliance on labels. Specifically, the RoboTAG consists of a 3D branch and a 2D branch, where nodes represent the states of the camera and robot system, and edges capture the dependencies between these variables or denote alignments between them. Closed loops are then defined in the graph, on which a consistency supervision across branches can be applied. Experimental results demonstrate that our method is effective across robot types, suggesting new possibilities of alleviating the data bottleneck in robotics.

cs.RO↗

arXiv2026

MAPLE: Elevating Medical Reasoning from Statistical Consensus to Process-Led Alignment

Recent advances in medical large language models have explored Test-Time Reinforcement Learning (TTRL) to enhance reasoning. However, standard TTRL often relies on majority voting (MV) as a heuristic supervision signal, which can be unreliable in complex medical scenarios where the most frequent reasoning path is not necessarily the clinically correct one. In this work, we propose a novel and unified training paradigm that integrates medical process reward models with TTRL to bridge the gap between test-time scaling (TTS) and parametric model optimization. Specifically, we advance the TTRL framework by replacing the conventional MV with a fine-grained, expert-aligned supervision paradigm using Med-RPM. This integration ensures that reinforcement learning is guided by medical correctness rather than mere consensus, effectively distilling search-based intelligence into the model's parametric memory. Extensive evaluations on four different benchmarks have demonstrated that our developed method consistently and significantly outperforms current TTRL and standalone PRM selection. Our findings establish that transitioning from stochastic heuristics to structured, step-wise rewards is essential for developing reliable and scalable medical AI systems

cs.LG↗

arXiv2025

LangSplatV2: High-dimensional 3D Language Gaussian Splatting with 450+ FPS

In this paper, we introduce LangSplatV2, which achieves high-dimensional feature splatting at 476.2 FPS and 3D open-vocabulary text querying at 384.6 FPS for high-resolution images, providing a 42 $\times$ speedup and a 47 $\times$ boost over LangSplat respectively, along with improved query accuracy. LangSplat employs Gaussian Splatting to embed 2D CLIP language features into 3D, significantly enhancing speed and learning a precise 3D language field with SAM semantics. Such advancements in 3D language fields are crucial for applications that require language interaction within complex scenes. However, LangSplat does not yet achieve real-time inference performance (8.2 FPS), even with advanced A100 GPUs, severely limiting its broader application. In this paper, we first conduct a detailed time analysis of LangSplat, identifying the heavyweight decoder as the primary speed bottleneck. Our solution, LangSplatV2 assumes that each Gaussian acts as a sparse code within a global dictionary, leading to the learning of a 3D sparse coefficient field that entirely eliminates the need for a heavyweight decoder. By leveraging this sparsity, we further propose an efficient sparse coefficient splatting method with CUDA optimization, rendering high-dimensional feature maps at high quality while incurring only the time cost of splatting an ultra-low-dimensional feature. Our experimental results demonstrate that LangSplatV2 not only achieves better or competitive query accuracy but is also significantly faster. Codes and demos are available at our project page: https://langsplat-v2.github.io.

cs.CV↗

arXiv2025

REPARO: Compositional 3D Assets Generation with Differentiable 3D Layout Alignment

Traditional image-to-3D models often struggle with scenes containing multiple objects due to biases and occlusion complexities. To address this challenge, we present REPARO, a novel approach for compositional 3D asset generation from single images. REPARO employs a two-step process: first, it extracts individual objects from the scene and reconstructs their 3D meshes using off-the-shelf image-to-3D models; then, it optimizes the layout of these meshes through differentiable rendering techniques, ensuring coherent scene composition. By integrating optimal transport-based long-range appearance loss term and high-level semantic loss term in the differentiable rendering, REPARO can effectively recover the layout of 3D assets. The proposed method can significantly enhance object independence, detail accuracy, and overall scene coherence. Extensive evaluation of multi-object scenes demonstrates that our REPARO offers a comprehensive approach to address the complexities of multi-object 3D scene generation from single images.

cs.CV↗

arXiv2025

Momentum-GS: Momentum Gaussian Self-Distillation for High-Quality Large Scene Reconstruction

3D Gaussian Splatting has demonstrated notable success in large-scale scene reconstruction, but challenges persist due to high training memory consumption and storage overhead. Hybrid representations that integrate implicit and explicit features offer a way to mitigate these limitations. However, when applied in parallelized block-wise training, two critical issues arise since reconstruction accuracy deteriorates due to reduced data diversity when training each block independently, and parallel training restricts the number of divided blocks to the available number of GPUs. To address these issues, we propose Momentum-GS, a novel approach that leverages momentum-based self-distillation to promote consistency and accuracy across the blocks while decoupling the number of blocks from the physical GPU count. Our method maintains a teacher Gaussian decoder updated with momentum, ensuring a stable reference during training. This teacher provides each block with global guidance in a self-distillation manner, promoting spatial consistency in reconstruction. To further ensure consistency across the blocks, we incorporate block weighting, dynamically adjusting each block's weight according to its reconstruction accuracy. Extensive experiments on large-scale scenes show that our method consistently outperforms existing techniques, achieving a 12.8% improvement in LPIPS over CityGaussian with much fewer divided blocks and establishing a new state of the art. Project page: https://jixuan-fan.github.io/Momentum-GS_Page/

cs.CV↗

arXiv2025

Audio-Plane: Audio Factorization Plane Gaussian Splatting for Real-Time Talking Head Synthesis

Talking head synthesis has emerged as a prominent research topic in computer graphics and multimedia, yet most existing methods often struggle to strike a balance between generation quality and computational efficiency, particularly under real-time constraints. In this paper, we propose a novel framework that integrates Gaussian Splatting with a structured Audio Factorization Plane (Audio-Plane) to enable high-quality, audio-synchronized, and real-time talking head generation. For modeling a dynamic talking head, a 4D volume representation, which consists of three axes in 3D space and one temporal axis aligned with audio progression, is typically required. However, directly storing and processing a dense 4D grid is impractical due to the high memory and computation cost, and lack of scalability for longer durations. We address this challenge by decomposing the 4D volume representation into a set of audio-independent spatial planes and audio-dependent planes, forming a compact and interpretable representation for talking head modeling that we refer to as the Audio-Plane. This factorized design allows for efficient and fine-grained audio-aware spatial encoding, and significantly enhances the model's ability to capture complex lip dynamics driven by speech signals. To further improve region-specific motion modeling, we introduce an audio-guided saliency splatting mechanism based on region-aware modulation, which adaptively emphasizes highly dynamic regions such as the mouth area. This allows the model to focus its learning capacity on where it matters most for accurate speech-driven animation. Extensive experiments on both the self-driven and the cross-driven settings demonstrate that our method achieves state-of-the-art visual quality, precise audio-lip synchronization, and real-time performance, outperforming prior approaches across both 2D- and 3D-based paradigms.

cs.GR↗

arXiv2025

CTRL-GS: Cascaded Temporal Residue Learning for 4D Gaussian Splatting

Recently, Gaussian Splatting methods have emerged as a desirable substitute for prior Radiance Field methods for novel-view synthesis of scenes captured with multi-view images or videos. In this work, we propose a novel extension to 4D Gaussian Splatting for dynamic scenes. Drawing on ideas from residual learning, we hierarchically decompose the dynamic scene into a "video-segment-frame" structure, with segments dynamically adjusted by optical flow. Then, instead of directly predicting the time-dependent signals, we model the signal as the sum of video-constant values, segment-constant values, and frame-specific residuals, as inspired by the success of residual learning. This approach allows more flexible models that adapt to highly variable scenes. We demonstrate state-of-the-art visual quality and real-time rendering on several established datasets, with the greatest improvements on complex scenes with large movements, occlusions, and fine details, where current methods degrade most.

cs.CV↗