Tag
The paper shows that diffusion training acceleration methods like REPA and dispersive regularization primarily reorganize the AdaLN-Zero residual gate α, and proposes a teacher-free radial objective that broadens gate participation to improve FID on ImageNet 256×256 with SiT-B/2.
CausalAlign 论文提出一种因果对齐框架,通过将扩散视频生成模型的内部表征与 3D 基础模型在 delta-token 空间中对齐,改善长时程相机自运动的一致性,在几何一致性上提升超过 20%,并提升了具身操作任务中世界动作模型的成功率。
提出一种通过蒸馏最优传输在仅有100-500个配对样本下对齐两个独立预训练单模态编码器的方法,在Flickr30k、MS-COCO和ChEBI-20上以相同锚点预算超越现有方法。
该论文发现全模态大模型(Omni-LLMs)存在'知道却说不出'(knows-but-cannot-say)的现象,即能感知非文本模态信息却难以在输出中表达。作者提出Token级模态对齐(TMA)指标和OmniCatalyst方法,通过自生成文本补充模态输入,在七个Omni-LLM上平均提升7.29%性能并平衡模态偏好。
论文提出 ViRA 框架,研究视觉基础模型(VFM)表征何时能提升端到端自动驾驶性能,发现 VFM 目标选择与辅助感知监督显著影响规划增益,并推出 ViRA-Diffusion 在 NAVSIM v2 上达到 92.3 EPDMS。
ProAR transforms autoregressive video generation into a goal-oriented reasoning process by anchoring generation to predicted goal frames via an asymmetric attention mask and adding future representation self-alignment, achieving strong benchmark gains while training on only 25% of the steps of standard AR baselines.
PixelDense reframes dense prediction tasks as representation alignment targets for pixel-space diffusion, routing semantic (DINOv2, SAM2) and geometric (DepthAnythingv2, Metric3Dv2) teachers through separate projection streams with an orthogonality penalty. The method improves GenEval, DPG-Bench, and HPSv2.1 scores on PixelGen and DeCo, achieves up to 53.1% PQ gain in partial-noise reconstruction, and trains 1.23x faster than baseline.
ELF-REG scales continuous diffusion language models to reasoning tasks by using representation alignment and entanglement, achieving improved performance on benchmarks like GSM8K and HumanEval compared to other diffusion models.
This survey paper defines model fusion and organizes research into three levels—parameter, representation, and behavior—for integrating capabilities in large language models.
DCRA introduces a diffusion-conditioned representation alignment framework to enhance robustness in time-series learning, particularly for clinical signals like EEG and ECG, by aligning representations across noise levels.
The paper explores K/V-cache interventions for persona control in decoder-only language models, finding dissociations between representation alignment and behavioral expression, with mid-layer interventions being effective for persona expression while preserving lexical diversity.
FlashRender is a few-step generative rendering framework that accelerates video synthesis by aligning representations and using mean-flow objectives, achieving comparable quality to multi-step methods with significantly reduced sampling cost.
ReFlowSET is a latent flow-matching framework for high-fidelity SAR-to-EO image translation, using codec selection and conditional DiT with vision model alignment to achieve state-of-the-art performance.
A systematic empirical study showing that concept directions extracted from one language model can steer other independently trained models when sufficient scale (≥1.7B parameters) is reached, providing functional evidence for the Platonic Representation Hypothesis and highlighting scale thresholds for cross-model interpretability tools.
This paper introduces CGTime, a 4B-parameter computation-grounded time-series-language model that decouples perception from description using deterministic statistics and LLM verbalization, outperforming larger general-purpose models on multivariate understanding tasks.
This paper investigates the mechanisms behind self-alignment methods in diffusion transformers, revealing that performance improvements from methods like Self-Flow primarily come from data augmentation along the noise dimension rather than token interactions between noise levels. The authors introduce Attention Separation to demonstrate this and propose an effective design combining self-representation alignment with dual-timestep augmentation.
GEAR proposes a method to jointly train a vector-quantized tokenizer and autoregressive generator end-to-end via representation alignment, achieving up to 10x faster convergence on ImageNet gFID compared to strong baselines.
HeRA aligns individual attention heads in Multimodal Large Language Models (MLLMs) to preserve local neighborhood relationships across modalities, improving vision-centric task performance and reducing visual hallucinations.
DO-ALL is a plug-and-play framework that uses dataset distillation to generate synthetic anchors that summarize source data, enabling stable long-term continual test-time adaptation without retaining original source data.
This paper presents the first systematic study of multilingual instruction following in Vision-Language-Action (VLA) models, revealing significant performance degradation when models trained on English are evaluated on other languages. The authors propose Multilingual Principal Component Alignment (MPCA) to reduce the multilingual performance gap.