arXiv subjects

Yijing Zhang

Publications and source records attributed to Yijing Zhang.

11 recordsLinked to original sources

arXiv2026

World Narrative Model for Highly Controllable Video Generation: A Paradigm Shift from Pixel Sampling to Physical World Orchestration

The fundamental obstacle to industrial grade video generation is the lack of controllability: existing models treat video as a pixel distribution sampling problem, bypassing the explicit, instance level $4D$ $(3D + T)$ physical world. Consequently, content creators cannot specify geometry, motion, camera parameters, or lighting in a deterministic, quantitative way, leading to the infamous ''gacha'' loop that makes professional content creation prohibitively inefficient and expensive. To address this, we introduce the World Narrative Model (WNM), a paradigm that decouples what to render -- the structured physical narrative -- from how to render -- the pixel generation process. WNM replaces end-to-end black-box sampling with orchestrated $4D$ pre-visualization for media generation. Collaborative agents translate sparse multimodal inputs, including text, reference videos, and sketches, into a fully editable world representation with scene geometry, object layouts, character/animal skeleton motion, trajectories, camera motion, and lighting at quantitative, physically meaningful granularity. This representation acts as a deterministic structural blueprint that drives existing video foundation models, either frozen or lightly adapted, to render final footage, turning the base model into a faithful neural shader. Built on this engine, our human-AI platform supports automatic world generation and pre-visualization aligned with professional filmmaking pipelines, while director consoles enable seamless human refinement. Experiments show that WNM greatly reduces probabilistic ``gacha'' calls and produces videos whose layout, motion, and cinematography closely follow creator intent. The framework is open and modular, allowing each component, such as world representation, control agents, and adapters, to be independently improved. Project website: https://glassroom.sjtu.edu.cn/WNM/.

cs.CV↗

arXiv2026

Breakeven complexity: A new perspective on neural partial differential equation solvers

Neural surrogate solvers of partial differential equations (PDEs) promise dramatic speedups over numerical methods, especially in scenarios requiring many solves. However, current accuracy-based evaluations do not fully consider two central issues: (1) neural solvers incur substantial up-front costs for data generation, training, and tuning; and (2) classical solvers can also generate low-fidelity solutions at a sufficiently low simulation cost. To explicitly account for these realities and fully incorporate end-to-end costs, we propose an evaluation framework centered on breakeven complexity, a metric that counts the forward solves before a learned solver is cost-effective relative to an error-equivalent traditional solver. To evaluate this measure, we apply scaling laws to determine how much training budget to allocate to data generation and discuss how to achieve smooth error-matching in diverse settings. We evaluate the breakeven complexity of multiple neural PDE solvers on three PDEs on 2D periodic domains from APEBench and a novel benchmark of flows past multiple obstacles generated by the GPU-native PyFR code. Among other findings, our results suggest that neural PDE solvers become more effective as problems get harder in terms of cost, dimension, rollout, physics regime (e.g. higher Reynolds number), etc.

cs.LG↗

arXiv2026

Children's Intelligence Tests Pose Challenges for MLLMs? KidGym: A 2D Grid-Based Reasoning Benchmark for MLLMs

Multimodal Large Language Models (MLLMs) combine the linguistic strengths of LLMs with the ability to process multimodal data, enbaling them to address a broader range of visual tasks. Because MLLMs aim at more general, human-like competence than language-only models, we take inspiration from the Wechsler Intelligence Scales - an established battery for evaluating children by decomposing intelligence into interpretable, testable abilities. We introduce KidGym, a comprehensive 2D grid-based benchmark for assessing five essential capabilities of MLLMs: Execution, Perception Reasoning, Learning, Memory and Planning. The benchmark comprises 12 unique tasks, each targeting at least one core capability, specifically designed to guage MLLMs' adaptability and developmental potential, mirroring the stages of children's cognitive growth. Additionally, our tasks encompass diverse scenarios and objects with randomly generated layouts, ensuring a more accurate and robust evluation of MLLM capabilities. KidGym is designed to be fully user-customizable and extensible, allowing researchers to create new evaluation scenarios and adjust difficuly levels to accommodate the rapidly growing MLLM community. Through the evaluation of state-of-the-art MLLMs using KidGym, we identified significant insights into model capabilities and revealed several limitations of current models. We release our benchmark at: https://bobo-ye.github.io/KidGym/.

cs.CL↗

arXiv2025

Scaling Up Audio-Synchronized Visual Animation: An Efficient Training Paradigm

Recent advances in audio-synchronized visual animation enable control of video content using audios from specific classes. However, existing methods rely heavily on expensive manual curation of high-quality, class-specific training videos, posing challenges to scaling up to diverse audio-video classes in the open world. In this work, we propose an efficient two-stage training paradigm to scale up audio-synchronized visual animation using abundant but noisy videos. In stage one, we automatically curate large-scale videos for pretraining, allowing the model to learn diverse but imperfect audio-video alignments. In stage two, we finetune the model on manually curated high-quality examples, but only at a small scale, significantly reducing the required human effort. We further enhance synchronization by allowing each frame to access rich audio context via multi-feature conditioning and window attention. To efficiently train the model, we leverage pretrained text-to-video generator and audio encoders, introducing only 1.9\% additional trainable parameters to learn audio-conditioning capability without compromising the generator's prior knowledge. For evaluation, we introduce AVSync48, a benchmark with videos from 48 classes, which is 3$\times$ more diverse than previous benchmarks. Extensive experiments show that our method significantly reduces reliance on manual curation by over 10$\times$, while generalizing to many open classes.

cs.CV↗

arXiv2025

Revenue Optimization in Wireless Video Caching Networks: A Privacy-Preserving Two-Stage Solution

Video caching can significantly improve delivery efficiency and enhance quality of video streaming, which constitutes the majority of wireless communication traffic. Due to limited cache size, caching strategies must be designed to adapt to and dynamic user demand in order to maximize system revenue. The system revenue depends on the benefits of delivering the requested videos and costs for (a) transporting the files to the users and (b) cache replacement. Since the cache content at any point in time impacts the replacement costs in the future, demand predictions over multiple cache placement slots become an important prerequisite for efficient cache planning. Motivated by this, we introduce a novel two-stage privacy-preserving solution for revenue optimization in wireless video caching networks. First, we train a Transformer using privacy-preserving federated learning (FL) to predict multi-slot future demands. Given that prediction results are never entirely accurate, especially for longer horizons, we further combine global content popularity with per-user prediction results to estimate the content demand distribution. Then, in the second stage, we leverage these estimation results to find caching strategies that maximize the long-term system revenue. This latter problem takes on the form of a multi-stage knapsack problem, which we then transform to a integer linear program. Our extensive simulation results demonstrate that (i) our FL solution delivers nearly identical performance to that of the ideal centralized solution and outperforms other existing caching methods, and (ii) our novel revenue optimization approach provides deeper system performance insights than traditional cache hit ratio (CHR)-based optimization approaches.

cs.NI↗

arXiv2025

Revenue Optimization in Video Caching Networks with Privacy-Preserving Demand Predictions

Performance of video streaming, which accounts for most of the traffic in wireless communication, can be significantly improved by caching popular videos at the wireless edge. Determining the cache content that optimizes performance (defined via a revenue function) is thus an important task, and prediction of the future demands based on past history can make this process much more efficient. However, since practical video caching networks involve various parties (e.g., users, isp, and csp) that do not wish to reveal information such as past history to each other, privacy-preserving solutions are required. Motivated by this, we propose a proactive caching method based on users' privacy-preserving multi-slot future demand predictions -- obtained from a trained Transformer -- to optimize revenue. Specifically, we first use a privacy-preserving fl algorithm to train a Transformer to predict multi-slot future demands of the users. However, prediction accuracy is not perfect and decreases the farther into the future the prediction is done. We model the impact of prediction errors invoking the file popularities, based on which we formulate a long-term system revenue optimization to make the cache placement decisions. As the formulated problem is NP-hard, we use a greedy algorithm to efficiently obtain an approximate solution. Simulation results validate that (i) the fl solution achieves results close to the centralized (non-privacy-preserving) solution and (ii) optimization of revenue may provide different solutions than the classical chr criterion.

cs.NI↗

arXiv2025

Personalize Your LLM: Fake it then Align it

Personalizing large language models (LLMs) is essential for delivering tailored interactions that improve user experience. Many existing personalization methods require fine-tuning LLMs for each user, rendering them prohibitively expensive for widespread adoption. Although retrieval-based approaches offer a more compute-efficient alternative, they still depend on large, high-quality datasets that are not consistently available for all users. To address this challenge, we propose CHAMELEON, a scalable and efficient personalization approach that uses (1) self-generated personal preference data and (2) representation editing to enable quick and cost-effective personalization. Our experiments on various tasks, including those from the LaMP personalization benchmark, show that CHAMELEON efficiently adapts models to personal preferences, improving instruction-tuned models and outperforms two personalization baselines by an average of 40% across two model architectures.

cs.LG↗

arXiv2024

Audio-Synchronized Visual Animation

Current visual generation methods can produce high quality videos guided by texts. However, effectively controlling object dynamics remains a challenge. This work explores audio as a cue to generate temporally synchronized image animations. We introduce Audio Synchronized Visual Animation (ASVA), a task animating a static image to demonstrate motion dynamics, temporally guided by audio clips across multiple classes. To this end, we present AVSync15, a dataset curated from VGGSound with videos featuring synchronized audio visual events across 15 categories. We also present a diffusion model, AVSyncD, capable of generating dynamic animations guided by audios. Extensive evaluations validate AVSync15 as a reliable benchmark for synchronized generation and demonstrate our models superior performance. We further explore AVSyncDs potential in a variety of audio synchronized generation tasks, from generating full videos without a base image to controlling object motions with various sounds. We hope our established benchmark can open new avenues for controllable visual generation. More videos on project webpage https://lzhangbj.github.io/projects/asva/asva.html.

cs.CV↗

arXiv2024

Is Free Self-Alignment Possible?

Aligning pretrained language models (LMs) often requires large-scale preference data and substantial computational resources. These costs become even more prohibitive for multi-objective or pluralistic alignment. Is this truly necessary? Can we perform efficient alignment using only internal model capabilities, and without additional training? To answer this question, we propose AlignEZ, a novel approach that leverages (1) self-generated preference data and (2) representation editing to achieve cost-effective, efficient alignment. By operating directly on learned representations, AlignEZ independently targets different behavioral aspects without the overhead of traditional alignment methods. Our experiments reveal that this cost-efficient procedure improves performance across diverse tasks: up to 19.9% on general alignment and 1.9% on challenging mathematical reasoning tasks, even when starting from a strong base model. AlignEZ can also align models to multiple objectives simultaneously, granting fine-grained control over multiple preference axes. Finally, we show that AlignEZ can accelerate more expensive alignment procedures--such as DPO--even under limited availability of ground-truth preference data.

cs.CL↗

arXiv2018

Bus Trajectory-Based Street-Centric Routing for Message Delivery in Urban Vehicular Ad hoc Networks

This paper focuses on the routing algorithm for the communications between vehicles and places in urban VANET. As one of the basic transportation facilities in an urban setting, buses periodically run along their fixed routes and widely cover city streets. The trajectory of bus lines can be seen as a sub map of a city. Based on the characters of bus networks, we propose a bus trajectory-based street-centric routing algorithm (BTSC), which uses bus as main relay to deliver message. In BTSC, we build a routing graph based on the trajectories of bus lines by analyzing the probability of bus appearing on every street. We propose two novel concepts, i.e. the probability of street consistency (PSC) and the probability of path consistency (PPC) which is used as metrics to determine routing paths for message delivery. This aims to choose the best path with higher density of busses and lower probability of transmission direction deviating from the routing path. In order to improve the bus forwarding opportunity, we design a bus-based forwarding strategy with ant colony optimization (FACO) to find a reliable and steady multi-hop link between two relay buses in order to decrease end-to-end delay. BTSC makes the improvements in the selection of routing path and the strategy of message forwarding. Simulation results show that our proposed routing algorithm has a better performance in transmission ratio, transmission delay and adaptability to different networks.

cs.NI↗

arXiv2016

Experimental Study of Nonlinear Resonances and Anti-resonances in a Forced, Ordered Granular Chain

We experimentally study a one-dimensional uncompressed granular chain composed of a finite number of identical spherical beads with Hertzian interactions. The chain is harmonically excited by an amplitude- and frequency-dependent boundary drive at its left end and has a fixed boundary at its right end. Such ordered granular media represent an interesting new class of nonlinear acoustic metamaterials, since they exhibit essentially nonlinear acoustics and have been designated as 'sonic vacua' due to the fact that their corresponding speed of sound (as defined in classical acoustics) is zero.This paves the way for essentially nonlinear and energy-dependent acoustics with no counterparts in linear theory. We experimentally detect time-periodic, strongly nonlinear resonances whereby the particles (beads) of the granular chain respond at integer multiples of the excitation period, and which correspond to local peaks of the maximum transmitted force at the chain's right, fixed end.In between these resonances we detect a local minimum of the maximum transmitted forces corresponding to an anti-resonance in the stationary-state dynamics. The experimental results of this work confirm previous theoretical predictions, and verify the existence of strongly nonlinear resonance responses in a system with a complete absence of any linear spectrum; as such, the experimentally detected nonlinear resonance spectrum is passively tunable with energy and sensitive to dissipative effects such as internal structural damping in the beads, and friction or plasticity effects. The experimental results are verified by direct numerical simulations and by numerical stability analysis.

math.DS↗