S2GS: Streaming Semantic Gaussian Splatting for Online Scene Understanding and Reconstruction

Existing offline feed-forward methods for joint scene understanding and reconstruction on long image streams often repeatedly perform global computation over an ever-growing set of past observations, causing runtime and GPU memory to increase rapidly with sequence length and limiting scalability. We propose Streaming Semantic Gaussian Splatting (S2GS), a strictly causal, incremental 3D Gaussian semantic field framework: it does not leverage future frames and continuously updates scene geometry, appearance, and instance-level semantics without reprocessing historical frames, enabling scalable online joint reconstruction and understanding. S2GS adopts a geometry-semantic decoupled dual-backbone design: the geometry branch performs causal modeling to drive incremental Gaussian updates, while the semantic branch leverages a 2D foundation vision model and a query-driven decoder to predict segmentation masks and identity embeddings, further stabilized by query-level contrastive alignment and lightweight online association with an instance memory. Experiments show that S2GS matches or outperforms strong offline baselines on joint reconstruction-and-understanding benchmarks, while significantly improving long-horizon scalability: it processes 1,000+ frames with much slower growth in runtime and GPU memory, whereas offline global-processing baselines typically run out of memory at around 80 frames under the same setting.

S2D: Sparse to Dense Lifting for 3D Reconstruction with Minimal Inputs

Explicit 3D representations have already become an essential medium for 3D simulation and understanding. However, the most commonly used point cloud and 3D Gaussian Splatting (3DGS) each suffer from non-photorealistic rendering and significant degradation under sparse inputs. In this paper, we introduce Sparse to Dense lifting (S2D), a novel pipeline that bridges the two representations and achieves high-quality 3DGS reconstruction with minimal inputs. Specifically, the S2D lifting is two-fold. We first present an efficient one-step diffusion model that lifts sparse point cloud for high-fidelity image artifact fixing. Meanwhile, to reconstruct 3D consistent scenes, we also design a corresponding reconstruction strategy with random sample drop and weighted gradient for robust model fitting from sparse input views to dense novel views. Extensive experiments show that S2D achieves the best consistency in generating novel view guidance and first-tier sparse view reconstruction quality under different input sparsity. By reconstructing stable scenes with the least possible captures among existing methods, S2D enables minimal input requirements for 3DGS applications.

Zero-Shot Robotic Manipulation via 3D Gaussian Splatting-Enhanced Multimodal Retrieval-Augmented Generation

Existing end-to-end approaches of robotic manipulation often lack generalization to unseen objects or tasks due to limited data and poor interpretability. While recent Multimodal Large Language Models (MLLMs) demonstrate strong commonsense reasoning, they struggle with geometric and spatial understanding required for pose prediction. In this paper, we propose RobMRAG, a 3D Gaussian Splatting-Enhanced Multimodal Retrieval-Augmented Generation (MRAG) framework for zero-shot robotic manipulation. Specifically, we construct a multi-source manipulation knowledge base containing object contact frames, task completion frames, and pose parameters. During inference, a Hierarchical Multimodal Retrieval module first employs a three-priority hybrid retrieval strategy to find task-relevant object prototypes, then selects the geometrically closest reference example based on pixel-level similarity and Instance Matching Distance (IMD). We further introduce a 3D-Aware Pose Refinement module based on 3D Gaussian Splatting into the MRAG framework, which aligns the pose of the reference object to the target object in 3D space. The aligned results are reprojected onto the image plane and used as input to the MLLM to enhance the generation of the final pose parameters. Extensive experiments show that on a test set containing 30 categories of household objects, our method improves the success rate by 7.76% compared to the best-performing zero-shot baseline under the same setting, and by 6.54% compared to the state-of-the-art supervised baseline. Our results validate that RobMRAG effectively bridges the gap between high-level semantic reasoning and low-level geometric execution, enabling robotic systems that generalize to unseen objects while remaining inherently interpretable.

Decoupled Reasoning with Implicit Fact Tokens (DRIFT): A Dual-Model Framework for Efficient Long-Context Inference

The integration of extensive, dynamic knowledge into Large Language Models (LLMs) remains a significant challenge due to the inherent entanglement of factual data and reasoning patterns. Existing solutions, ranging from non-parametric Retrieval-Augmented Generation (RAG) to parametric knowledge editing, are often constrained in practice by finite context windows, retriever noise, or the risk of catastrophic forgetting. In this paper, we propose DRIFT, a novel dual-model architecture designed to explicitly decouple knowledge extraction from the reasoning process. Unlike static prompt compression, DRIFT employs a lightweight knowledge model to dynamically compress document chunks into implicit fact tokens conditioned on the query. These dense representations are projected into the reasoning model's embedding space, replacing raw, redundant text while maintaining inference accuracy. Extensive experiments show that DRIFT significantly improves performance on long-context tasks, outperforming strong baselines among comparably sized models. Our approach provides a scalable and efficient paradigm for extending the effective context window and reasoning capabilities of LLMs. Our code is available at https://github.com/Lancelot-Xie/DRIFT.

FOCUS: DLLMs Know How to Tame Their Compute Bound

Diffusion Large Language Models (DLLMs) offer a compelling alternative to Auto-Regressive models, but their deployment is constrained by high decoding cost. In this work, we identify a key inefficiency in DLLM decoding: while computation is parallelized over token blocks, only a small subset of tokens is decodable at each diffusion step, causing most compute to be wasted on non-decodable tokens. We further observe a strong correlation between attention-derived token importance and token-wise decoding probability. Based on this insight, we propose FOCUS -- an inference system designed for DLLMs. By dynamically focusing computation on decodable tokens and evicting non-decodable ones on-the-fly, FOCUS increases the effective batch size, alleviating compute limitations and enabling scalable throughput. Empirical evaluations demonstrate that FOCUS achieves up to 3.52$\times$ throughput improvement over the production-grade engine LMDeploy, while preserving or improving generation quality across multiple benchmarks. The FOCUS system is publicly available on GitHub: https://github.com/sands-lab/FOCUS.

Explore with Long-term Memory: A Benchmark and Multimodal LLM-based Reinforcement Learning Framework for Embodied Exploration

An ideal embodied agent should possess lifelong learning capabilities to handle long-horizon and complex tasks, enabling continuous operation in general environments. This not only requires the agent to accurately accomplish given tasks but also to leverage long-term episodic memory to optimize decision-making. However, existing mainstream one-shot embodied tasks primarily focus on task completion results, neglecting the crucial process of exploration and memory utilization. To address this, we propose Long-term Memory Embodied Exploration (LMEE), which aims to unify the agent's exploratory cognition and decision-making behaviors to promote lifelong learning.We further construct a corresponding dataset and benchmark, LMEE-Bench, incorporating multi-goal navigation and memory-based question answering to comprehensively evaluate both the process and outcome of embodied exploration. To enhance the agent's memory recall and proactive exploration capabilities, we propose MemoryExplorer, a novel method that fine-tunes a multimodal large language model through reinforcement learning to encourage active memory querying. By incorporating a multi-task reward function that includes action prediction, frontier selection, and question answering, our model achieves proactive exploration. Extensive experiments against state-of-the-art embodied exploration models demonstrate that our approach achieves significant advantages in long-horizon embodied tasks.

EscherVerse: An Open World Benchmark and Dataset for Teleo-Spatial Intelligence with Physical-Dynamic and Intent-Driven Understanding

The ability to reason about spatial dynamics is a cornerstone of intelligence, yet current research overlooks the human intent behind spatial changes. To address these limitations, we introduce Teleo-Spatial Intelligence (TSI), a new paradigm that unifies two critical pillars: Physical-Dynamic Reasoning--understanding the physical principles of object interactions--and Intent-Driven Reasoning--inferring the human goals behind these actions. To catalyze research in TSI, we present EscherVerse, consisting of a large-scale, open-world benchmark (Escher-Bench), a dataset (Escher-35k), and models (Escher series). Derived from real-world videos, EscherVerse moves beyond constrained settings to explicitly evaluate an agent's ability to reason about object permanence, state transitions, and trajectory prediction in dynamic, human-centric scenarios. Crucially, it is the first benchmark to systematically assess Intent-Driven Reasoning, challenging models to connect physical events to their underlying human purposes. Our work, including a novel data curation pipeline, provides a foundational resource to advance spatial intelligence from passive scene description toward a holistic, purpose-driven understanding of the world.

FLEG: Feed-Forward Language Embedded Gaussian Splatting from Any Views

We present FLEG, a feed-forward network that reconstructs language-embedded 3D Gaussians from any views. Previous straightforward solutions combine feed-forward reconstruction with Gaussian heads but suffer from fixed input views and insufficient 3D training data. In contrast, we propose a 3D-annotation-free training framework for 2D-to-3D lifting from arbitrary uncalibrated and unposed multi-view images. Since the framework does not require 3D annotations, we can leverage large-scale video data with easily obtained 2D instance information to enrich semantic embedding. We also propose an instance-guided contrastive learning to align 2D semantics with the 3D representations. In addition, to mitigate the high memory and computational cost of dense views, we further propose a geometry-semantic hierarchical sparsification strategy. Our FLEG efficiently reconstructs language-embedded 3D Gaussian representation in a feed-forward manner from arbitrary sparse or dense views, jointly producing accurate geometry, high-fidelity appearance, and language-aligned semantics. Extensive experiments show that it outperforms existing methods on various related tasks. Project page: https://fangzhou2000.github.io/projects/fleg.

KOSS: Kalman-Optimal Selective State Spaces for Long-Term Sequence Modeling

Recent selective state space models (SSMs), such as Mamba and Mamba-2, have demonstrated strong performance in sequence modeling owing to input-dependent selection mechanisms. However, these mechanisms lack theoretical grounding and cannot support context-aware selection from latent state dynamics. To address these limitations, we propose KOSS, a Kalman-optimal Selective State Space model that formulates selection as latent state uncertainty minimization. Derived from estimation theory, KOSS adopts a continuous-time latent update driven by a Kalman gain that dynamically modulates information propagation based on content and context, enabling a closed-loop, context-aware selectivity mechanism. To ensure stable computation and near-linear scalability, KOSS employs global spectral differentiation for frequency-domain derivative estimation, along with a segment-wise scan for hardware-efficient processing. On a selective copying task with distractors, KOSS achieves over 79\% accuracy while baselines drop below 20\%, demonstrating robust context-aware selection. Furthermore, across nine long-term forecasting benchmarks, KOSS reduces MSE by 2.92--36.23\% and consistently outperforms state-of-the-art models in both accuracy and stability. To assess real-world applicability, a case study on secondary surveillance radar (SSR) tracking confirms KOSS's robustness under irregular intervals and noisy conditions and demonstrates its effectiveness in real-world applications. Finally, supplementary experiments verify Kalman gain convergence and the frequency response of spectral differentiation, providing theoretical support for the proposed closed-loop design.

CaRF: Enhancing Multi-View Consistency in Referring 3D Gaussian Splatting Segmentation

Referring 3D Gaussian Splatting Segmentation (R3DGS) aims to interpret free-form language expressions and localize the corresponding 3D regions in Gaussian fields. While recent advances have introduced cross-modal alignment between language and 3D geometry, existing pipelines still struggle with cross-view consistency due to their reliance on 2D rendered pseudo supervision and view specific feature learning. In this work, we present Camera Aware Referring Field (CaRF), a fully differentiable framework that operates directly in the 3D Gaussian space and achieves multi view consistency. Specifically, CaRF introduces Gaussian Field Camera Encoding (GFCE), which incorporates camera geometry into Gaussian text interactions to explicitly model view dependent variations and enhance geometric reasoning. Building on this, In Training Paired View Supervision (ITPVS) is proposed to align per Gaussian logits across calibrated views during training, effectively mitigating single view overfitting and exposing inter view discrepancies for optimization. Extensive experiments on three representative benchmarks demonstrate that CaRF achieves average improvements of 16.8%, 4.3%, and 2.0% in mIoU over state of the art methods on the Ref LERF, LERF OVS, and 3D OVS datasets, respectively. Moreover, this work promotes more reliable and view consistent 3D scene understanding, with potential benefits for embodied AI, AR/VR interaction, and autonomous perception.