Towards Holistic Modeling for Video Frame Interpolation with Auto-regressive Diffusion Transformers

Existing video frame interpolation (VFI) methods often adopt a frame-centric approach, processing videos as independent short segments (e.g., triplets), which leads to temporal inconsistencies and motion artifacts. To overcome this, we propose a holistic, video-centric paradigm named \textbf{L}ocal \textbf{D}iffusion \textbf{F}orcing for \textbf{V}ideo \textbf{F}rame \textbf{I}nterpolation (LDF-VFI). Our framework is built upon an auto-regressive diffusion transformer that models the entire video sequence to ensure long-range temporal coherence. To mitigate error accumulation inherent in auto-regressive generation, we introduce a novel skip-concatenate sampling strategy that effectively maintains temporal stability. Furthermore, LDF-VFI incorporates sparse, local attention and tiled VAE encoding, a combination that not only enables efficient processing of long sequences but also allows generalization to arbitrary spatial resolutions (e.g., 4K) at inference without retraining. An enhanced conditional VAE decoder, which leverages multi-scale features from the input video, further improves reconstruction fidelity. Empirically, LDF-VFI achieves state-of-the-art performance on challenging long-sequence benchmarks, demonstrating superior per-frame quality and temporal consistency, especially in scenes with large motion. The source code is available at https://github.com/xypeng9903/LDF-VFI.

TeleWorld: Towards Dynamic Multimodal Synthesis with a 4D World Model

World models aim to endow AI systems with the ability to represent, generate, and interact with dynamic environments in a coherent and temporally consistent manner. While recent video generation models have demonstrated impressive visual quality, they remain limited in real-time interaction, long-horizon consistency, and persistent memory of dynamic scenes, hindering their evolution into practical world models. In this report, we present TeleWorld, a real-time multimodal 4D world modeling framework that unifies video generation, dynamic scene reconstruction, and long-term world memory within a closed-loop system. TeleWorld introduces a novel generation-reconstruction-guidance paradigm, where generated video streams are continuously reconstructed into a dynamic 4D spatio-temporal representation, which in turn guides subsequent generation to maintain spatial, temporal, and physical consistency. To support long-horizon generation with low latency, we employ an autoregressive diffusion-based video model enhanced with Macro-from-Micro Planning (MMPL)--a hierarchical planning method that reduces error accumulation from frame-level to segment-level-alongside efficient Distribution Matching Distillation (DMD), enabling real-time synthesis under practical computational budgets. Our approach achieves seamless integration of dynamic object modeling and static scene representation within a unified 4D framework, advancing world models toward practical, interactive, and computationally accessible systems. Extensive experiments demonstrate that TeleWorld achieves strong performance in both static and dynamic world understanding, long-term consistency, and real-time generation efficiency, positioning it as a practical step toward interactive, memory-enabled world models for multimodal generation and embodied intelligence.

EVICPRESS: Joint KV-Cache Compression and Eviction for Efficient LLM Serving

Reusing KV cache is essential for high efficiency of Large Language Model (LLM) inference systems. With more LLM users, the KV cache footprint can easily exceed GPU memory capacity, so prior work has proposed to either evict KV cache to lower-tier storage devices, or compress KV cache so that more KV cache can be fit in the fast memory. However, prior work misses an important opportunity: jointly optimizing the eviction and compression decisions across all KV caches to minimize average generation latency without hurting quality. We propose EVICPRESS, a KV-cache management system that applies lossy compression and adaptive eviction to KV cache across multiple storage tiers. Specifically, for each KV cache of a context, EVICPRESS considers the effect of compression and eviction of the KV cache on the average generation quality and delay across all contexts as a whole. To achieve this, EVICPRESS proposes a unified utility function that quantifies the effect of quality and delay of the lossy compression or eviction. To this end, EVICPRESS's profiling module periodically updates the utility function scores on all possible eviction-compression configurations for all contexts and places KV caches using a fast heuristic to rearrange KV caches on all storage tiers, with the goal of maximizing the utility function scores on each storage tier. Compared to the baselines that evict KV cache or compress KV cache, EVICPRESS achieves higher KV-cache hit rates on fast devices, i.e., lower delay, while preserving high generation quality by applying conservative compression to contexts that are sensitive to compression errors. Evaluation on 12 datasets and 5 models demonstrates that EVICPRESS achieves up to 2.19x faster time-to-first-token (TTFT) at equivalent generation quality.

3DGabSplat: 3D Gabor Splatting for Frequency-adaptive Radiance Field Rendering

Recent prominence in 3D Gaussian Splatting (3DGS) has enabled real-time rendering while maintaining high-fidelity novel view synthesis. However, 3DGS resorts to the Gaussian function that is low-pass by nature and is restricted in representing high-frequency details in 3D scenes. Moreover, it causes redundant primitives with degraded training and rendering efficiency and excessive memory overhead. To overcome these limitations, we propose 3D Gabor Splatting (3DGabSplat) that leverages a novel 3D Gabor-based primitive with multiple directional 3D frequency responses for radiance field representation supervised by multi-view images. The proposed 3D Gabor-based primitive forms a filter bank incorporating multiple 3D Gabor kernels at different frequencies to enhance flexibility and efficiency in capturing fine 3D details. Furthermore, to achieve novel view rendering, an efficient CUDA-based rasterizer is developed to project the multiple directional 3D frequency components characterized by 3D Gabor-based primitives onto the 2D image plane, and a frequency-adaptive mechanism is presented for adaptive joint optimization of primitives. 3DGabSplat is scalable to be a plug-and-play kernel for seamless integration into existing 3DGS paradigms to enhance both efficiency and quality of novel view synthesis. Extensive experiments demonstrate that 3DGabSplat outperforms 3DGS and its variants using alternative primitives, and achieves state-of-the-art rendering quality across both real-world and synthetic scenes. Remarkably, we achieve up to 1.35 dB PSNR gain over 3DGS with simultaneously reduced number of primitives and memory consumption.

Alchemist: Towards the Design of Efficient Online Continual Learning System

Continual learning has emerged as a promising solution to refine models incrementally by leveraging user feedback, thereby enhancing model performance in applications like code completion, personal assistants, and chat interfaces. In particular, online continual learning - iteratively training the model with small batches of user feedback - has demonstrated notable performance improvements. However, the existing practice of segregating training and serving processes forces the online trainer to recompute the intermediate results already done during serving. Such redundant computations can account for 30%-42% of total training time. In this paper, we propose Alchemist, to the best of our knowledge, the first online continual learning system that efficiently reuses intermediate results computed during serving to reduce redundant computation with minimal impact on the serving latency or capacity. Alchemist introduces two key techniques: (1) minimal activations recording and saving during serving, where activations are recorded and saved only during the prefill phase to minimize overhead; and (2) offloading of serving activations, which dynamically manages GPU memory by freeing activations in the forward order, while reloading them in the backward order during the backward pass. Evaluations with the ShareGPT dataset show that compared with a separate training cluster, Alchemist significantly increases training throughput by up to 1.72x, reduces up to 47% memory usage during training, and supports up to 2x more training tokens - all while maintaining negligible impact on serving latency.

BotEval: Facilitating Interactive Human Evaluation

Following the rapid progress in natural language processing (NLP) models, language models are applied to increasingly more complex interactive tasks such as negotiations and conversation moderations. Having human evaluators directly interact with these NLP models is essential for adequately evaluating the performance on such interactive tasks. We develop BotEval, an easily customizable, open-source, evaluation toolkit that focuses on enabling human-bot interactions as part of the evaluation process, as opposed to human evaluators making judgements for a static input. BotEval balances flexibility for customization and user-friendliness by providing templates for common use cases that span various degrees of complexity and built-in compatibility with popular crowdsourcing platforms. We showcase the numerous useful features of BotEval through a study that evaluates the performance of various chatbots on their effectiveness for conversational moderation and discuss how BotEval differs from other annotation tools.

Data/moment-driven approaches for fast predictive control of collective dynamics

Feedback control synthesis for large-scale particle systems is reviewed in the framework of model predictive control (MPC). The high-dimensional character of collective dynamics hampers the performance of traditional MPC algorithms based on fast online dynamic optimization at every time step. Two alternatives to MPC are proposed. First, the use of supervised learning techniques for the offline approximation of optimal feedback laws is discussed. Then, a procedure based on sequential linearization of the dynamics based on macroscopic quantities of the particle ensemble is reviewed. Both approaches circumvent the online solution of optimal control problems enabling fast, real-time, feedback synthesis for large-scale particle systems. Numerical experiments assess the performance of the proposed algorithms.

Cascade-Zero123: One Image to Highly Consistent 3D with Self-Prompted Nearby Views

Synthesizing multi-view 3D from one single image is a significant and challenging task. For this goal, Zero-1-to-3 methods aim to extend a 2D latent diffusion model to the 3D scope. These approaches generate the target-view image with a single-view source image and the camera pose as condition information. However, the one-to-one manner adopted in Zero-1-to-3 incurs challenges for building geometric and visual consistency across views, especially for complex objects. We propose a cascade generation framework constructed with two Zero-1-to-3 models, named Cascade-Zero123, to tackle this issue, which progressively extracts 3D information from the source image. Specifically, a self-prompting mechanism is designed to generate several nearby views at first. These views are then fed into the second-stage model along with the source image as generation conditions. With self-prompted multiple views as the supplementary information, our Cascade-Zero123 generates more highly consistent novel-view images than Zero-1-to-3. The promotion is significant for various complex and challenging scenes, involving insects, humans, transparent objects, and stacked multiple objects etc. The project page is at https://cascadezero123.github.io/.

Can Language Model Moderators Improve the Health of Online Discourse?

Human moderation of online conversation is essential to maintaining civility and focus in a dialogue, but is challenging to scale and harmful to moderators. The inclusion of sophisticated natural language generation modules as a force multiplier aid moderators is a tantalizing prospect, but adequate evaluation approaches have so far been elusive. In this paper, we establish a systematic definition of conversational moderation effectiveness through a multidisciplinary lens that incorporates insights from social science. We then propose a comprehensive evaluation framework that uses this definition to asses models' moderation capabilities independently of human intervention. With our framework, we conduct the first known study of conversational dialogue models as moderators, finding that appropriately prompted models can provide specific and fair feedback on toxic behavior but struggle to influence users to increase their levels of respect and cooperation.

CacheGen: Fast Context Loading for Language Model Applications

As large language models (LLMs) take on more complex tasks, their inputs incorporate longer contexts to respond to questions that require domain knowledge or user-specific conversational histories. Yet, using long contexts poses a challenge for responsive LLM systems, as nothing can be generated until all the contexts are fetched to and processed by the LLM. Existing systems optimize only the computation delay in context processing (e.g., by caching intermediate key-value features of the text context) but often cause longer network delays in context fetching (e.g., key-value features consume orders of magnitude larger bandwidth than the text context). This paper presents CacheGen to minimize the delays in fetching and processing contexts for LLMs. CacheGen reduces the bandwidth needed for transmitting long contexts' key-value (KV) features through a novel encoder that compresses KV features into more compact bitstream representations. The encoder combines adaptive quantization with a tailored arithmetic coder, taking advantage of the KV features' distributional properties, such as locality across tokens. Furthermore, CacheGen minimizes the total delay in fetching and processing a context by using a controller that determines when to load the context as compressed KV features or raw text and picks the appropriate compression level if loaded as KV features. We test CacheGen on three models of various sizes and three datasets of different context lengths. Compared to recent methods that handle long contexts, CacheGen reduces bandwidth usage by 3.7-4.3x and the total delay in fetching and processing contexts by 2.7-3x while maintaining similar LLM performance on various tasks as loading the text contexts.