arXiv subjects

Zhuoyi Yang

Publications and source records attributed to Zhuoyi Yang.

At least 19 recordsLinked to original sources

arXiv2026

Asymmetric Capacity Allocation in Self-Refinement Pipelines

Self-refinement, typically structured as generation, critique, and revision, is a widely adopted paradigm for improving LLM generation and serves as a core mechanism in many LLM agents. While the three stages involve different cognitive demands, most existing approaches conveniently treat the model size as an implementation detail rather than a subject of study, which may lead to a waste of resources. Little work has systematically examined how model size affects each stage or whether effective self-refinement requires equally capable models for generation, critique, and revision. We present the first stage-wise model size study of the self-refinement pipeline on 5 benchmarks from different domains using 6 model sizes of Qwen3 and 4 model sizes of Gemma 3. We conclude that larger generators and refiners generally improve the pipeline, whereas an undersized refiner can even harm performance. Second, performance is highly insensitive to the size of the critic, although including even a small critic consistently outperforms omitting critique altogether. Our findings demonstrate that model capacity should not be allocated uniformly across self-refinement pipelines. Instead, different stages exhibit distinct size scaling characteristics, providing practical guidance for designing more computationally efficient multi-stage language model systems.

cs.LG↗

arXiv2026

SCAIL-2: Unifying Controlled Character Animation with End-to-End In-Context Conditioning

Controlled character animation aims to transfer motion from a driving sequence to a reference character. Prior works heavily rely on intermediate representations, such as pose skeletons for motion and masked backgrounds for environment, inevitably resulting in information loss. In this work, we present SCAIL-2, a framework that adopts an end-to-end driving paradigm by directly concatenating latent visual information to the model's input sequence. We enable end-to-end training through a data synthesis pipeline that produces MotionPair-60K, a curated dataset for several character animation subtasks. We unify the subtasks using decoupled conditions to accommodate different driving patterns, facilitated by In-Context Mask Conditioning and Mode-Specific RoPE, which provide soft guidance beyond textual instructions and visual information. To address synthetic discrepancy in detailed regions, we propose Bias-Aware DPO to construct preference items to mitigate the errors. Extensive experiments demonstrate that our method achieves state-of-the-art performance across various character animation tasks. Code, model weights, and a large subset of the dataset are available at: https://teal024.github.io/SCAIL-2/.

cs.CV↗

arXiv2026

Delving into Latent Spectral Biasing of Video VAEs for Superior Diffusability

Latent diffusion models pair VAEs with diffusion backbones, and the structure of VAE latents strongly influences the difficulty of diffusion training. However, existing video VAEs typically focus on reconstruction fidelity, overlooking latent structure. We present a statistical analysis of video VAE latent spaces and identify two spectral properties essential for diffusion training: a spatio-temporal frequency spectrum biased toward low frequencies, and a channel-wise eigenspectrum dominated by a few modes. To induce these properties, we propose two lightweight, backbone-agnostic regularizers: Local Correlation Regularization and Latent Masked Reconstruction. Experiments show that our Spectral-Structured VAE (SSVAE) achieves a $3\times$ speedup in text-to-video generation convergence and a 10\% gain in video reward, outperforming strong open-source VAEs. The code is available at https://github.com/zai-org/SSVAE.

cs.CV↗

arXiv2026

MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models

In recent years, vision language models (VLMs) have made significant advancements in video understanding. However, a crucial capability - fine-grained motion comprehension - remains under-explored in current benchmarks. To address this gap, we propose MotionBench, a comprehensive evaluation benchmark designed to assess the fine-grained motion comprehension of video understanding models. MotionBench evaluates models' motion-level perception through six primary categories of motion-oriented question types and includes data collected from diverse sources, ensuring a broad representation of real-world video content. Experimental results reveal that existing VLMs perform poorly in understanding fine-grained motions. To enhance VLM's ability to perceive fine-grained motion within a limited sequence length of LLM, we conduct extensive experiments reviewing VLM architectures optimized for video feature compression and propose a novel and efficient Through-Encoder (TE) Fusion method. Experiments show that higher frame rate inputs and TE Fusion yield improvements in motion understanding, yet there is still substantial room for enhancement. Our benchmark aims to guide and motivate the development of more capable video understanding models, emphasizing the importance of fine-grained motion comprehension. Project page: https://motion-bench.github.io .

cs.CV↗

arXiv2026

InstrAct: Towards Action-Centric Understanding in Instructional Videos

Understanding instructional videos requires recognizing fine-grained actions and modeling their temporal relations, which remains challenging for current Video Foundation Models (VFMs). This difficulty stems from noisy web supervision and a pervasive "static bias", where models rely on objects rather than motion cues. To address this, we propose InstrAction, a pretraining framework for instructional videos' action-centric representations. We first introduce a data-driven strategy, which filters noisy captions and generates action-centric hard negatives to disentangle actions from objects during contrastive learning. At the visual feature level, an Action Perceiver extracts motion-relevant tokens from redundant video encodings. Beyond contrastive learning, we introduce two auxiliary objectives: Dynamic Time Warping alignment (DTW-Align) for modeling sequential temporal structure, and Masked Action Modeling (MAM) for strengthening cross-modal grounding. Finally, we introduce the InstrAct Bench to evaluate action-centric understanding, where our method consistently outperforms state-of-the-art VFMs on semantic reasoning, procedural logic, and fine-grained retrieval tasks.

cs.CV↗

arXiv2026

SCAIL: Towards Studio-Grade Character Animation via In-Context Learning of 3D-Consistent Pose Representations

Achieving controllable character animation that meets studio-grade standards remains challenging despite recent progress. Existing approaches can transfer motion from a driving video to a reference image, but often fail to preserve structural fidelity and temporal consistency in wild scenarios involving complex motion and cross-identity animations. In this work, we present \textbf{SCAIL} (a framework toward \textbf{S}tudio-grade \textbf{C}haracter \textbf{A}nimation via \textbf{I}n-context \textbf{L}earning), which is designed to address these challenges from two key innovations. First, we propose a novel 3D pose representation, providing a robust and flexible motion signal. Second, we introduce a full-context pose injection mechanism within a diffusion-transformer, enabling effective spatio-temporal reasoning over full motion sequences. To align with studio-grade requirements, we develop a curated data pipeline ensuring both diversity and quality, and establish a comprehensive benchmark for systematic evaluation. Experiments show that \textbf{SCAIL} achieves state-of-the-art performance and advances character animation toward studio-grade controlling. Code and model are available at \href{https://github.com/zai-org/SCAIL}{zai-org/SCAIL}.

cs.CV↗

arXiv2026

Kaleido: Open-Sourced Multi-Subject Reference Video Generation Model

We present Kaleido, a subject-to-video~(S2V) generation framework, which aims to synthesize subject-consistent videos conditioned on multiple reference images of target subjects. Despite recent progress in S2V generation models, existing approaches remain inadequate at maintaining multi-subject consistency and at handling background disentanglement, often resulting in lower reference fidelity and semantic drift under multi-image conditioning. These shortcomings can be attributed to several factors. Primarily, the training dataset suffers from a lack of diversity and high-quality samples, as well as cross-paired data, i.e., paired samples whose components originate from different instances. In addition, the current mechanism for integrating multiple reference images is suboptimal, potentially resulting in the confusion of multiple subjects. To overcome these limitations, we propose a dedicated data construction pipeline, incorporating low-quality sample filtering and diverse data synthesis, to produce consistency-preserving training data. Moreover, we introduce Reference Rotary Positional Encoding (R-RoPE) to process reference images, enabling stable and precise multi-image integration. Extensive experiments across numerous benchmarks demonstrate that Kaleido significantly outperforms previous methods in consistency, fidelity, and generalization, marking an advance in S2V generation.

cs.CV↗

arXiv2026

Fine-Tuning vs. RAG for Multi-Hop Question Answering with Novel Knowledge

Multi-hop question answering is widely used to evaluate the reasoning capabilities of large language models (LLMs), as it requires integrating multiple pieces of supporting knowledge to arrive at a correct answer. While prior work has explored different mechanisms for providing knowledge to LLMs, such as finetuning and retrieval-augmented generation (RAG), their relative effectiveness for multi-hop question answering remains insufficiently understood, particularly when the required knowledge is temporally novel. In this paper, we systematically compare parametric and non-parametric knowledge injection methods for open-domain multi-hop question answering. We evaluate unsupervised fine-tuning (continual pretraining), supervised fine-tuning, and retrieval-augmented generation across three 7B-parameter open-source LLMs. Experiments are conducted on two benchmarks: QASC, a standard multi-hop science question answering dataset, and a newly constructed dataset of over 10,000 multi-hop questions derived from Wikipedia events in 2024, designed to test knowledge beyond the models' pretraining cutoff. Our results show that unsupervised fine-tuning provides only limited gains over base models, suggesting that continual pretraining alone is insufficient for improving multi-hop reasoning accuracy. In contrast, retrieval-augmented generation yields substantial and consistent improvements, particularly when answering questions that rely on temporally novel information. Supervised fine-tuning achieves the highest overall accuracy across models and datasets. These findings highlight fundamental differences in how knowledge injection mechanisms support multi-hop question answering and underscore the importance of retrieval-based methods when external or compositional knowledge is required.

cs.CL↗

arXiv2026

VisionReward: Fine-Grained Multi-Dimensional Human Preference Learning for Image and Video Generation

Visual generative models have achieved remarkable progress in synthesizing photorealistic images and videos, yet aligning their outputs with human preferences across critical dimensions remains a persistent challenge. Though reinforcement learning from human feedback offers promise for preference alignment, existing reward models for visual generation face limitations, including black-box scoring without interpretability and potentially resultant unexpected biases. We present VisionReward, a general framework for learning human visual preferences in both image and video generation. Specifically, we employ a hierarchical visual assessment framework to capture fine-grained human preferences, and leverages linear weighting to enable interpretable preference learning. Furthermore, we propose a multi-dimensional consistent strategy when using VisionReward as a reward model during preference optimization for visual generation. Experiments show that VisionReward can significantly outperform existing image and video reward models on both machine metrics and human evaluation. Notably, VisionReward surpasses VideoScore by 17.2% in preference prediction accuracy, and text-to-video models with VisionReward achieve a 31.6% higher pairwise win rate compared to the same models using VideoScore. All code and datasets are provided at https://github.com/THUDM/VisionReward.

cs.CV↗

arXiv2025

Test-time Scaling of LLMs: A Survey from A Subproblem Structure Perspective

With this paper, we survey techniques for improving the predictive accuracy of pretrained large language models by allocating additional compute at inference time. In categorizing test-time scaling methods, we place special emphasis on how a problem is decomposed into subproblems and on the topological organization of these subproblems whether sequential, parallel, or tree-structured. This perspective allows us to unify diverse approaches such as Chain-of-Thought, Branch-Solve-Merge, and Tree-of-Thought under a common lens. We further synthesize existing analyses of these techniques, highlighting their respective strengths and weaknesses, and conclude by outlining promising directions for future research

cs.CL↗

arXiv2025

AMAQ: Adaptive Mixed-bit Activation Quantization for Collaborative Parameter Efficient Fine-tuning

Large Language Models (LLMs) are scaling rapidly, creating significant challenges for collaborative server client distributed training, particularly in terms of communication efficiency and computational overheads. To address these challenges, we implement Parameter-efficient Split Learning, which effectively balances efficiency and performance for collaborative training on low-resource devices. To reduce communication overhead in collaborative training, we introduce Adaptive Mixed bit Activation Quantization (AMAQ), a strategy that progressively compresses activations and gradients from high precision (6 to 8 bits) to low precision (3 to 4 bits). AMAQ achieves this by effectively allocating bit budgets across channels based on feature wise and layer wise importance using bit regularization. Under the same bit budgets, AMAQ outperforms fixed-precision approaches, delivering about 2.5% higher generation accuracy and about 1.3% better classification accuracy for models like LLaMA3 8B and Qwen2.5 7B. In addition, it significantly enhances training stability and reducing ultra-low bit representation collapse during the training. Experiments demonstrate that AMAQ integrates effectively into practical multi-machine collaborative training setups, offering superior inference accuracy with only a modest communication overhead for bits adaptation during training. This trade off makes AMAQ a practical and effective solution for collaborative training with minimal communication cost.

cs.LG↗

arXiv2025

Stabilizing Information Flow Entropy: Regularization for Safe and Interpretable Autonomous Driving Perception

Deep perception networks in autonomous driving traditionally rely on data-intensive training regimes and post-hoc anomaly detection, often disregarding fundamental information-theoretic constraints governing stable information processing. We reconceptualize deep neural encoders as hierarchical communication chains that incrementally compress raw sensory inputs into task-relevant latent features. Within this framework, we establish two theoretically justified design principles for robust perception: (D1) smooth variation of mutual information between consecutive layers, and (D2) monotonic decay of latent entropy with network depth. Our analysis shows that, under realistic architectural assumptions, particularly blocks comprising repeated layers of similar capacity, enforcing smooth information flow (D1) naturally encourages entropy decay (D2), thus ensuring stable compression. Guided by these insights, we propose Eloss, a novel entropy-based regularizer designed as a lightweight, plug-and-play training objective. Rather than marginal accuracy improvements, this approach represents a conceptual shift: it unifies information-theoretic stability with standard perception tasks, enabling explicit, principled detection of anomalous sensor inputs through entropy deviations. Experimental validation on large-scale 3D object detection benchmarks (KITTI and nuScenes) demonstrates that incorporating Eloss consistently achieves competitive or improved accuracy while dramatically enhancing sensitivity to anomalies, amplifying distribution-shift signals by up to two orders of magnitude. This stable information-compression perspective not only improves interpretability but also establishes a solid theoretical foundation for safer, more robust autonomous driving perception systems.

cs.LG↗

arXiv2025

VPO: Aligning Text-to-Video Generation Models with Prompt Optimization

Video generation models have achieved remarkable progress in text-to-video tasks. These models are typically trained on text-video pairs with highly detailed and carefully crafted descriptions, while real-world user inputs during inference are often concise, vague, or poorly structured. This gap makes prompt optimization crucial for generating high-quality videos. Current methods often rely on large language models (LLMs) to refine prompts through in-context learning, but suffer from several limitations: they may distort user intent, omit critical details, or introduce safety risks. Moreover, they optimize prompts without considering the impact on the final video quality, which can lead to suboptimal results. To address these issues, we introduce VPO, a principled framework that optimizes prompts based on three core principles: harmlessness, accuracy, and helpfulness. The generated prompts faithfully preserve user intents and, more importantly, enhance the safety and quality of generated videos. To achieve this, VPO employs a two-stage optimization approach. First, we construct and refine a supervised fine-tuning (SFT) dataset based on principles of safety and alignment. Second, we introduce both text-level and video-level feedback to further optimize the SFT model with preference learning. Our extensive experiments demonstrate that VPO significantly improves safety, alignment, and video quality compared to baseline methods. Moreover, VPO shows strong generalization across video generation models. Furthermore, we demonstrate that VPO could outperform and be combined with RLHF methods on video generation models, underscoring the effectiveness of VPO in aligning video generation models. Our code and data are publicly available at https://github.com/thu-coai/VPO.

cs.CV↗

arXiv2025

Entropy Loss: An Interpretability Amplifier of 3D Object Detection Network for Intelligent Driving

With the increasing complexity of the traffic environment, the significance of safety perception in intelligent driving is intensifying. Traditional methods in the field of intelligent driving perception rely on deep learning, which suffers from limited interpretability, often described as a "black box." This paper introduces a novel type of loss function, termed "Entropy Loss," along with an innovative training strategy. Entropy Loss is formulated based on the functionality of feature compression networks within the perception model. Drawing inspiration from communication systems, the information transmission process in a feature compression network is expected to demonstrate steady changes in information volume and a continuous decrease in information entropy. By modeling network layer outputs as continuous random variables, we construct a probabilistic model that quantifies changes in information volume. Entropy Loss is then derived based on these expectations, guiding the update of network parameters to enhance network interpretability. Our experiments indicate that the Entropy Loss training strategy accelerates the training process. Utilizing the same 60 training epochs, the accuracy of 3D object detection models using Entropy Loss on the KITTI test set improved by up to 4.47\% compared to models without Entropy Loss, underscoring the method's efficacy. The implementation code is available at https://github.com/yhbcode000/Eloss-Interpretability.

cs.CV↗

arXiv2025

Concat-ID: Towards Universal Identity-Preserving Video Synthesis

We present Concat-ID, a unified framework for identity-preserving video generation. Concat-ID employs variational autoencoders to extract image features, which are then concatenated with video latents along the sequence dimension. It relies exclusively on inherent 3D self-attention mechanisms to incorporate them, eliminating the need for additional parameters or modules. A novel cross-video pairing strategy and a multi-stage training regimen are introduced to balance identity consistency and facial editability while enhancing video naturalness. Extensive experiments demonstrate Concat-ID's superiority over existing methods in both single and multi-identity generation, as well as its seamless scalability to multi-subject scenarios, including virtual try-on and background-controllable generation. Concat-ID establishes a new benchmark for identity-preserving video synthesis, providing a versatile and scalable solution for a wide range of applications.

cs.CV↗

arXiv2025

CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

We present CogVideoX, a large-scale text-to-video generation model based on diffusion transformer, which can generate 10-second continuous videos aligned with text prompt, with a frame rate of 16 fps and resolution of 768 * 1360 pixels. Previous video generation models often had limited movement and short durations, and is difficult to generate videos with coherent narratives based on text. We propose several designs to address these issues. First, we propose a 3D Variational Autoencoder (VAE) to compress videos along both spatial and temporal dimensions, to improve both compression rate and video fidelity. Second, to improve the text-video alignment, we propose an expert transformer with the expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. Third, by employing a progressive training and multi-resolution frame pack technique, CogVideoX is adept at producing coherent, long-duration, different shape videos characterized by significant motions. In addition, we develop an effective text-video data processing pipeline that includes various data preprocessing strategies and a video captioning method, greatly contributing to the generation quality and semantic alignment. Results show that CogVideoX demonstrates state-of-the-art performance across both multiple machine metrics and human evaluations. The model weight of both 3D Causal VAE, Video caption model and CogVideoX are publicly available at https://github.com/THUDM/CogVideo.

cs.CV↗

arXiv2025

LogLLaMA: Transformer-based log anomaly detection with LLaMA

Log anomaly detection refers to the task that distinguishes the anomalous log messages from normal log messages. Transformer-based large language models (LLMs) are becoming popular for log anomaly detection because of their superb ability to understand complex and long language patterns. In this paper, we propose LogLLaMA, a novel framework that leverages LLaMA2. LogLLaMA is first finetuned on normal log messages from three large-scale datasets to learn their patterns. After finetuning, the model is capable of generating successive log messages given previous log messages. Our generative model is further trained to identify anomalous log messages using reinforcement learning (RL). The experimental results show that LogLLaMA outperforms the state-of-the-art approaches for anomaly detection on BGL, Thunderbird, and HDFS datasets.

cs.LG↗

arXiv2024

CogVLM2: Visual Language Models for Image and Video Understanding

Beginning with VisualGLM and CogVLM, we are continuously exploring VLMs in pursuit of enhanced vision-language fusion, efficient higher-resolution architecture, and broader modalities and applications. Here we propose the CogVLM2 family, a new generation of visual language models for image and video understanding including CogVLM2, CogVLM2-Video and GLM-4V. As an image understanding model, CogVLM2 inherits the visual expert architecture with improved training recipes in both pre-training and post-training stages, supporting input resolution up to $1344 \times 1344$ pixels. As a video understanding model, CogVLM2-Video integrates multi-frame input with timestamps and proposes automated temporal grounding data construction. Notably, CogVLM2 family has achieved state-of-the-art results on benchmarks like MMBench, MM-Vet, TextVQA, MVBench and VCGBench. All models are open-sourced in https://github.com/THUDM/CogVLM2 and https://github.com/THUDM/GLM-4, contributing to the advancement of the field.

cs.CV↗