arXiv subjects

Letian Li

Publications and source records attributed to Letian Li.

17 recordsLinked to original sources

arXiv2026

TetherCache: Stabilizing Autoregressive Long-Form Video Generation with Gated Recall and Trusted Alignment

Autoregressive video diffusion models provide a natural formulation for streaming and variable-length video generation by conditioning newly generated frames on previously generated content. However, extending these models to minute-level generation remains challenging: the limited KV-cache budget prevents the model from retaining the full history, while repeatedly conditioning on self-generated frames induces a context distribution shift that accumulates over time, leading to visual artifacts, quality degradation, and temporal drift. In this paper, we propose TetherCache, a training-free and plug-and-play cache management strategy for drift-resistant long video generation. TetherCache organizes the cache into sink, memory, and recent regions, and introduces two complementary mechanisms. First, GRAB (Gated Recall with Attention-Diversity Balancing) selects long-range memory frames using a gated score that combines attention-based relevance with temporal diversity, preserving informative yet diverse historical context under a fixed cache budget. Second, TAME (Trusted Alignment via Memory Editing) lightly edits newly recalled memory tokens by aligning their statistics to a trusted context distribution, reducing the pollution caused by drifted historical features. Built on Self-Forcing, TetherCache consistently improves long-video generation quality on VBench-Long across 30s, 60s, and 240s settings. In particular, for 240s generation, it substantially improves overall and semantic scores while reducing quality drift from 7.84 to 1.33, demonstrating its effectiveness for stable long-horizon autoregressive video diffusion.

cs.CV↗

arXiv2026

Learning to Assist: Collaborative VLAs for Implicit Human-Robot Collaboration

Human-robot collaboration (HRC) combines the complementary strengths of humans and robots to improve task efficiency. However, many existing collaborative systems rely on hand-engineered pipelines, limiting their scalability and flexibility for new tasks. In this work, we show that models trained end-to-end with imitation learning, specifically vision-language-action (VLA) models, can support collaborative manipulation, and characterize the key factors affecting their real-world performance. We evaluate two state-of-the-art models and identify a failure mode of action-chunking policies in implicit HRC, where demonstration action leakage (i.e., action chunks crossing latent task transitions) can cause premature assistive behavior. We find that this issue increases with longer execution horizons and occurs in real-world collaborative VLA systems, such as when a robot attempts to hand over a tool before the person is ready. We propose an inference-time steering method to mitigate these erroneous assistive actions while preserving policy performance. Finally, through a 16-participant user study on a long-horizon collaborative assembly task, we show that steering enables a longer execution horizon while mitigating premature assistance, leading to faster collaboration and fewer failures compared to a shorter-horizon policy.

cs.RO↗

arXiv2026

HDSL: A Hierarchical Domain-Specific Language for Structured 3D Indoor Scene Generation and Localized Editing with LLM Agents

Text-driven indoor scene generation and editing require an intermediate representation that language models can both produce and revise. Existing LLM-based systems often rely on scene graphs or global constraint lists, which are compact but underspecify local geometry and make instruction-based edits difficult to localize. We frame this problem as structured program generation and local program repair, and propose Hierarchical Descriptive Scene Language (HDSL), an XML/CSS-style domain-specific language for structured 3D indoor scenes. HDSL represents rooms, regions, objects, and support surfaces as a tree with local coordinates, making complex scenes easier to plan recursively and easier to retrieve for editing. Our pipeline uses LLM agents to generate HDSL subtrees with bounded verification, grounds non-virtual nodes through multimodal asset retrieval, and applies force-directed layout optimization to repair boundary and collision errors. For editing, Hierarchical Retrieval-Augmented Generation retrieves the relevant subtree, asks the LLM to rewrite only that local context, and merges the result back through a deterministic three-way merge. In our reproduced benchmark, HDSL improves average object coverage, text-scene alignment, and generation time over full text-to-scene baselines while remaining competitive with recent layout-only reproductions on geometry metrics; for editing, HRAG reduces token use by $5.22\times$ and runtime by $6.19\times$, produces valid DSL for all eight paired edits, and better preserves unrelated scene objects.

cs.CV↗

arXiv2026

cMoLLM at Scale: Horizontal Scaling Laws for Mixture-of-LLMs

Scaling large language models (LLMs) has driven their success, yet dense Transformers couple capacity and computation: every parameter is activated for every token, making training and inference costs grow linearly with model size-a critical bottleneck as models approach trillion-parameter regimes. We aim to scale capacity through MoE-style mixture throughout the LLM pipeline rather than only the FFN. Prior pipeline-level approaches include ParaScale, which introduces virtual tokens and parallel streams but incurs substantial overhead and suffers from homogenized routing and gradient collapse, and AltUp, which uses an auxiliary prediction branch but offers limited adaptivity and slow convergence. We establish that MoE-style mixture layers can be reformulated as variable-kernel dynamic convolutions, where each expert corresponds to a $1{\times}1$ convolutional kernel and routing implements input-conditioned kernel aggregation. Building on this equivalence, we introduce cMoLLM: a convolutionally gated mixture-of-LLMs that routes over end-to-end streams through fully differentiable dynamic convolution. In GPT-2-style models trained on FineWeb, cMoLLM improves language modeling perplexity and downstream GLUE and SQuAD accuracy under matched compute, with better stream utilization, more stable optimization, and favorable scaling compared to ParaScale- and AltUp-style baselines.

cs.AI↗

arXiv2026

Position: Multi-Agent Systems Should Prioritize Concurrency Control

LLM-based multi-agent systems (MAS) promise scalable collaboration, yet adding agents often reduces reliability. This position paper argues that many MAS failures are fundamentally concurrency control problems: agents concurrently read and write shared state, and long LLM inference windows amplify the risk of stale reads, lost updates, and inconsistent outcomes. Failure modes commonly attributed to coordination or communication breakdowns can be mapped directly onto classical concurrency anomalies. We contend that MAS frameworks should address these failures through explicit concurrency control mechanisms: conflict detection, isolation guarantees, and structured access to shared resources. Concurrency control should be a first-class design concern, not an afterthought.

cs.AI↗

arXiv2026

Towards Self-Robust LLMs: Intrinsic Prompt Noise Resistance via CoIPO

Large language models (LLMs) have demonstrated remarkable and steadily improving performance across a wide range of tasks. However, LLM performance may be highly sensitive to prompt variations especially in scenarios with limited openness or strict output formatting requirements, indicating insufficient robustness. In real-world applications, user prompts provided to LLMs often contain imperfections, which may undermine the quality of the model's responses. To address this issue, previous work has primarily focused on preprocessing prompts, employing external tools or even LLMs to refine prompt formulations in advance. However, these approaches overlook the intrinsic robustness of LLMs, and their reliance on external components introduces additional computational overhead and uncertainty. In this work, we propose a Contrastive Learning-based Inverse Direct Preference Optimization (CoIPO) method that minimizes the discrepancy between the label-aligned logits produced by the model under a clean prompt and its noisy counterpart, and conduct a detailed analysis using mutual information theory. We augment the FLAN dataset by constructing paired prompts, each consisting of a clean prompt and its corresponding noisy version for training. Additionally, to evaluate the effectiveness, we develop NoisyPromptBench, a benchmark enhanced and derived from the existing PromptBench. Experimental results conducted on NoisyPromptBench demonstrate that our proposed method achieves a significant improvement in average accuracy over the current state-of-the-art approaches. The source code of CoIPO, pair-wise FLAN datasets, and NoisyPromptBench have already been released on https://github.com/vegetable-yx/CoIPO.

cs.CL↗

arXiv2026

CG-MLLM: Captioning and Generating 3D content via Multi-modal Large Language Models

Large Language Models(LLMs) have revolutionized text generation and multimodal perception,but their capabilities in 3D content generation remain underexplored. Existing methods compromise by producing either low-resolution meshes or coarse structural proxies, failing to capture finegrained geometry natively. In this paper, we propose CG-MLLM, a novel Multi-modal Large Language Model (MLLM) capable of 3D captioning and high-resolution 3D generation in a single framework. Leveraging the Mixture-ofTransformer architecture, CG-MLLM decouples disparate modeling needs, where the Token-level Autoregressive (TokenAR) Transformer handles token-level content, and the Block-level Autoregressive (BlockAR) Transformer handles blocklevel content. By integrating a pre-trained visionlanguage backbone with a specialized 3D VAE latent space, CG-MLLM facilitates long-context interactions between standard tokens and spatial blocks within a single integrated architecture. Experimental results show that CG-MLLM significantly outperforms existing MLLMs in generating high-fidelity 3D objects, effectively bringing high-resolution 3D content creation into the mainstream LLM paradigm. Beyond generation, we further observe that learning to produce 3D content transfers back to perception, strengthening the model's image-based 3D understanding.

cs.CV↗

arXiv2026

Biphasic Meniscus Coating for Scalable and Material Efficient Quantum Dot Films

Colloidal quantum dots (cQDs) have emerged as a cornerstone of next-generation optoelectronics, offering unparalleled spectral tunability and solution-processability. However, the transition from laboratory-scale devices to sustainable industrial manufacturing is fundamentally hindered by spin-coating workflows, which are intrinsically wasteful and restricted to planar geometries. These limitations are particularly acute for high-performance cQDs containing regulated elements such as lead, cadmium, or mercury, where poor material utilization exacerbates both environmental burden and cost. Here we report a biphasic dip-coating strategy that redefines the material efficiency of nanocrystal film fabrication. By utilizing an immiscible underlayer to displace ~88% of the active reservoir volume, we demonstrate a deposition geometry that decouples material consumption from total precursor volume. Infrared PbS photodetectors fabricated via this approach maintain their performance against spin-coated benchmarks while reducing ink consumption by up to 20-fold. Our technoeconomic analysis reveals that this biphasic architecture achieves cost parity at film thicknesses an order of magnitude lower than conventional monophasic dip-coating. Our results establish a low-waste framework for solution-processed materials, providing a viable pathway for the resource-efficient manufacturing of optoelectronic devices.

cond-mat.mtrl-sci↗

arXiv2025

IGen: Scalable Data Generation for Robot Learning from Open-World Images

The rise of generalist robotic policies has created an exponential demand for large-scale training data. However, on-robot data collection is labor-intensive and often limited to specific environments. In contrast, open-world images capture a vast diversity of real-world scenes that naturally align with robotic manipulation tasks, offering a promising avenue for low-cost, large-scale robot data acquisition. Despite this potential, the lack of associated robot actions hinders the practical use of open-world images for robot learning, leaving this rich visual resource largely unexploited. To bridge this gap, we propose IGen, a framework that scalably generates realistic visual observations and executable actions from open-world images. IGen first converts unstructured 2D pixels into structured 3D scene representations suitable for scene understanding and manipulation. It then leverages the reasoning capabilities of vision-language models to transform scene-specific task instructions into high-level plans and generate low-level actions as SE(3) end-effector pose sequences. From these poses, it synthesizes dynamic scene evolution and renders temporally coherent visual observations. Experiments validate the high quality of visuomotor data generated by IGen, and show that policies trained solely on IGen-synthesized data achieve performance comparable to those trained on real-world data. This highlights the potential of IGen to support scalable data generation from open-world images for generalist robotic policy training.

cs.RO↗

arXiv2025

RailS: Load Balancing for All-to-All Communication in Distributed Mixture-of-Experts Training

Training Mixture-of-Experts (MoE) models introduces sparse and highly imbalanced all-to-all communication that dominates iteration time. Conventional load-balancing methods fail to exploit the deterministic topology of Rail architectures, leaving multi-NIC bandwidth underutilized. We present RailS, a distributed load-balancing framework that minimizes all-to-all completion time in MoE training. RailS leverages the Rail topology's symmetry to prove that uniform sending ensures uniform receiving, transforming global coordination into local scheduling. Each node independently executes a Longest Processing Time First (LPT) spraying scheduler to proactively balance traffic using local information. RailS activates N parallel rails for fine-grained, topology-aware multipath transmission. Across synthetic and real-world MoE workloads, RailS improves bus bandwidth by 20%--78% and reduces completion time by 17%--78%. For Mixtral workloads, it shortens iteration time by 18%--40% and achieves near-optimal load balance, fully exploiting architectural parallelism in distributed training.

cs.DC↗

arXiv2025

Row-wise Fusion Regularization: An Interpretable Personalized Federated Learning Framework in Large-Scale Scenarios

We study personalized federated learning for multivariate responses where client models are heterogeneous yet share variable-level structure. Existing entry-wise penalties ignore cross-response dependence, while matrix-wise fusion over-couples clients. We propose a Sparse Row-wise Fusion (SROF) regularizer that clusters row vectors across clients and induces within-row sparsity, and we develop RowFed, a communication-efficient federated algorithm that embeds SROF into a linearized ADMM framework with privacy-preserving partial participation. Theoretically, we establish an oracle property for SROF-achieving correct variable-level group recovery with asymptotic normality-and prove convergence of RowFed to a stationary solution. Under random client participation, the iterate gap contracts at a rate that improves with participation probability. Empirically, simulations in heterogeneous regimes show that RowFed consistently lowers estimation and prediction error and strengthens variable-level cluster recovery over NonFed, FedAvg, and a personalized matrix-fusion baseline. A real-data study further corroborates these gains while preserving interpretability. Together, our results position row-wise fusion as an effective and transparent paradigm for large-scale personalized federated multivariate learning, bridging the gap between entry-wise and matrix-wise formulations.

stat.ML↗

arXiv2025

SD-GS: Structured Deformable 3D Gaussians for Efficient Dynamic Scene Reconstruction

Current 4D Gaussian frameworks for dynamic scene reconstruction deliver impressive visual fidelity and rendering speed, however, the inherent trade-off between storage costs and the ability to characterize complex physical motions significantly limits the practical application of these methods. To tackle these problems, we propose SD-GS, a compact and efficient dynamic Gaussian splatting framework for complex dynamic scene reconstruction, featuring two key contributions. First, we introduce a deformable anchor grid, a hierarchical and memory-efficient scene representation where each anchor point derives multiple 3D Gaussians in its local spatiotemporal region and serves as the geometric backbone of the 3D scene. Second, to enhance modeling capability for complex motions, we present a deformation-aware densification strategy that adaptively grows anchors in under-reconstructed high-dynamic regions while reducing redundancy in static areas, achieving superior visual quality with fewer anchors. Experimental results demonstrate that, compared to state-of-the-art methods, SD-GS achieves an average of 60\% reduction in model size and an average of 100\% improvement in FPS, significantly enhancing computational efficiency while maintaining or even surpassing visual quality.

cs.GR↗

arXiv2025

BuildingBlock: A Hybrid Approach for Structured Building Generation

Three-dimensional building generation is vital for applications in gaming, virtual reality, and digital twins, yet current methods face challenges in producing diverse, structured, and hierarchically coherent buildings. We propose BuildingBlock, a hybrid approach that integrates generative models, procedural content generation (PCG), and large language models (LLMs) to address these limitations. Specifically, our method introduces a two-phase pipeline: the Layout Generation Phase (LGP) and the Building Construction Phase (BCP). LGP reframes box-based layout generation as a point-cloud generation task, utilizing a newly constructed architectural dataset and a Transformer-based diffusion model to create globally consistent layouts. With LLMs, these layouts are extended into rule-based hierarchical designs, seamlessly incorporating component styles and spatial structures. The BCP leverages these layouts to guide PCG, enabling local-customizable, high-quality structured building generation. Experimental results demonstrate BuildingBlock's effectiveness in generating diverse and hierarchically structured buildings, achieving state-of-the-art results on multiple benchmarks, and paving the way for scalable and intuitive architectural workflows.

cs.GR↗

arXiv2025

Anisotropic Field Suppression of Morin Transition Temperature in Epitaxially Grown Hematite Thin Films

We have demonstrated the existence of the Morin transition in epitaxially grown hematite thin films exceeding a critical thickness. The Morin transition temperature can be suppressed by magnetic fields applied both parallel and perpendicular to the Dzyaloshinskii-Moriya (DM) vector, exhibiting a distinct anisotropic behavior that is consistent with bulk hematite crystals. Detailed analysis explains the anisotropic behavior and provides a method for determining the DM strength, which remains nearly constant across the sample thickness over four orders of magnitude. Our findings obtained with transport measurements offer a valuable approach for studying antiferromagnetic spin configurations in thin films and nanodevices.

cond-mat.mtrl-sci↗

arXiv2024

Guidelines for Correlative Imaging and Analysis of Reactive Lithium Metal Battery Materials

To unlock the full potential of lithium metal batteries, a deep understanding of lithium metal reactivity and its solid electrolyte interphase is essential. Correlative imaging, combining focused ion beam and electron microscopy offers a powerful approach for multi-scale characterization. However, the extreme reactivity of lithium metal and its SEI presents challenges in investigating deposition and stripping mechanisms. In this work, we systematically evaluated the storage stability of lithium metal in glovebox before and after electrochemical deposition. We then assessed different FIB ion sources for their impact on lithium metal lamella preparation for transmission electron microscopy. Furthermore, we examined cryogenic-TEM transfer methods, optimizing for minimal contamination during sample handling. Contrary to prior assumptions, we demonstrate that high resolution imaging of pure lithium metal at room temperature is achievable using inert gas transfer with an electron dose rate exceeding 1000 e/A2/s, without significant detectable damage. In contrast, SEI components, such as Li2CO3 and LiF display much greater sensitivity to electron beams, requiring cryogenic conditions and precise dose control for nano/atomic scale imaging. We quantified electron dose limits for these SEI components to track their structural evolution under irradiation. Based on these findings, we propose a robust protocol for lithium metal sample handling - from storage to atomic-level characterization - minimizing damage and contamination. This work paves the way for more accurate and reproducible studies, accelerating the development of next-generation lithium metal batteries by ensuing the preservation of native material properties during analysis.

cond-mat.mtrl-sci↗

arXiv2024

Grain Selection Growth of Soft Metal in Electrochemical Processes

Soft metals like lithium and sodium play a critical role in battery technology owing to their high energy density. Texture formation by grain selection growth of soft metals during electrochemical processes is a crucial factor affecting power and safety. Developing a framework to understand and control grain growth is a multifaceted challenge. Here, a general thermodynamic theory and phase-field model are formulated to study grain selection growth of soft metals. Our study focuses on the interplay between surface energy and atomic mobility-related intrinsic strain energy in grain selection growth. Differences in grain selection growth arise from the anisotropy in surface energy and diffusion barrier of soft metal atoms. Our findings highlight the kinetic limitations of solid-state Li metal batteries, which originate from load stress-induced surface energy anisotropy. These insights lead to the development of an amorphous LixSi1-x (0.50<x<0.79) seed layer, improving the critical current density at room temperature for anode-free Li solid-state batteries through the control of grain selection growth.

cond-mat.mtrl-sci↗

arXiv2024

EEG-EMG FAConformer: Frequency Aware Conv-Transformer for the fusion of EEG and EMG

Motor pattern recognition paradigms are the main forms of Brain-Computer Interfaces(BCI) aimed at motor function rehabilitation and are the most easily promoted applications. In recent years, many researchers have suggested encouraging patients to perform real motor control execution simultaneously in MI-based BCI rehabilitation training systems. Electromyography (EMG) signals are the most direct physiological signals that can assess the execution of movements. Multimodal signal fusion is practically significant for decoding motor patterns. Therefore, we introduce a multimodal motion pattern recognition algorithm for EEG and EMG signals: EEG-EMG FAConformer, a method with several attention modules correlated with temporal and frequency information for motor pattern recognition. We especially devise a frequency band attention module to encode EEG information accurately and efficiently. What's more, modules like Multi-Scale Fusion Module, Independent Channel-Specific Convolution Module(ICSCM), and Fuse Module which can effectively eliminate irrelevant information in EEG and EMG signals and fully exploit hidden dynamics are developed and show great effects. Extensive experiments show that EEG-EMG FAConformer surpasses existing methods on Jeong2020 dataset, showcasing outstanding performance, high robustness and impressive stability.

eess.SP↗