Recent text-to-video (T2V) diffusion models rely heavily on auxiliary reward signals (e.g., via reward models or DPO) to align generated content with human aesthetics and improve realism. These signals, however, incur substantial computational overhead, require costly human annotations, and often yield limited improvement in fine-grained local details. In this paper, we argue that your data manifold is secretly a reward model. By explicitly modeling the manifold structure of high-quality Supervised Fine-Tuning (SFT) data and encouraging video latents to lie on this manifold, we derive dense, differentiable, and nearly cost-free reward signals that significantly improve video quality, particularly in mitigating low-level distortions. Our modeling builds upon Local Coordinate Coding (LCC), which captures the `skeleton' of the manifold. However, directly applying LCC suffers from mean regression, pulling latents toward the geometric mean and losing high-frequency details. We therefore extend it to Shell Local Coordinate Coding (Shell-LCC), which models the manifold `surface' as an isotropic shell to align with the true high-density region. Experiments demonstrate that our approach imp
In this work, we analyze the ability of NCSN++ U-Net based audio dereverberation models to capture global room characteristics in their intermediate representations. Through an empirical study of both a state-of-the-art diffusion-based model and a discriminative counterpart, we show that deeper layers encode structured room impulse response (RIR)-dependent embeddings. Moreover, the discriminative ability of this implicit room representation correlates with dereverberation performance across objective metrics. Motivated by this observation, we propose a training strategy that explicitly conditions the network on pre-trained RIR embeddings, obtained via self-supervised contrastive learning. Incorporating RIR conditioning improves representation quality, accelerates convergence, and enhances dereverberation performance, while significantly reducing the number of reverse diffusion steps required by the diffusion-based model during inference.
Retrieval augmented generation (RAG) has been widely adopted to help Large Language Models (LLMs) to process tasks involving long documents. However, existing retrieval models are not designed for long document retrieval and fail to address several key challenges of long document retrieval, including context-awareness, causal dependence, and scope of retrieval. In this paper, we proposed AttentionRetriever, a novel long document retrieval model that leverages attention mechanism and entity-based retrieval to build context-aware embeddings for long document and determine the scope of retrieval. With extensive experiments, we found AttentionRetriever is able to outperform existing retrieval models on long document retrieval datasets by a large margin while remaining as efficient as dense retrieval models.
We consider the protein sequence engineering problem, which aims to find protein sequences with high fitness levels, starting from a given wild-type sequence. Directed evolution has been a dominating paradigm in this field which has an iterative process to generate variants and select via experimental feedback. We demonstrate large language models (LLMs), despite being trained on massive texts, are secretly protein sequence optimizers. With a directed evolutionary method, LLM can perform protein engineering through Pareto and experiment-budget constrained optimization, demonstrating success on both synthetic and experimental fitness landscapes.
High-quality instance and panoptic segmentation has traditionally relied on dense instance-level annotations such as masks, boxes, or points, which are costly, inconsistent, and difficult to scale. Unsupervised and weakly-supervised approaches reduce this burden but remain constrained by semantic backbone constraints and human bias, often producing merged or fragmented outputs. We present TRACE (TRAnsforming diffusion Cues to instance Edges), showing that text-to-image diffusion models secretly function as instance edge annotators. TRACE identifies the Instance Emergence Point (IEP) where object boundaries first appear in self-attention maps, extracts boundaries through Attention Boundary Divergence (ABDiv), and distills them into a lightweight one-step edge decoder. This design removes the need for per-image diffusion inversion, achieving 81x faster inference while producing sharper and more connected boundaries. On the COCO benchmark, TRACE improves unsupervised instance segmentation by +5.1 AP, and in tag-supervised panoptic segmentation it outperforms point-supervised baselines by +1.7 PQ without using any instance-level labels. These results reveal that diffusion models encode
Dense retrievers enhance retrieval by encoding queries and documents into continuous vectors, but they often struggle with reasoning-intensive queries. Although Large Language Models (LLMs) can reformulate queries to capture complex reasoning, applying them universally incurs significant computational cost. In this work, we propose Adaptive Query Reasoning (AdaQR), a hybrid query rewriting framework. Within this framework, a Reasoner Router dynamically directs each query to either fast dense reasoning or deep LLM reasoning. The dense reasoning is achieved by the Dense Reasoner, which performs LLM-style reasoning directly in the embedding space, enabling a controllable trade-off between efficiency and accuracy. Experiments on large-scale retrieval benchmarks BRIGHT show that AdaQR reduces reasoning cost by 28% while preserving-or even improving-retrieval performance by 7%.
Inverse Constraint Learning (ICL) is the problem of inferring constraints from safe (i.e., constraint-satisfying) demonstrations. The hope is that these inferred constraints can then be used downstream to search for safe policies for new tasks and, potentially, under different dynamics. Our paper explores the question of what mathematical entity ICL recovers. Somewhat surprisingly, we show that both in theory and in practice, ICL recovers the set of states where failure is inevitable, rather than the set of states where failure has already happened. In the language of safe control, this means we recover a backwards reachable tube (BRT) rather than a failure set. In contrast to the failure set, the BRT depends on the dynamics of the data collection system. We discuss the implications of the dynamics-conditionedness of the recovered constraint on both the sample-efficiency of policy search and the transferability of learned constraints.
Recently, contrastive learning has risen to be a promising paradigm for extracting meaningful data representations. Among various special designs, adding a projection head on top of the encoder during training and removing it for downstream tasks has proven to significantly enhance the performance of contrastive learning. However, despite its empirical success, the underlying mechanism of the projection head remains under-explored. In this paper, we develop an in-depth theoretical understanding of the projection head from the information-theoretic perspective. By establishing the theoretical guarantees on the downstream performance of the features before the projector, we reveal that an effective projector should act as an information bottleneck, filtering out the information irrelevant to the contrastive objective. Based on theoretical insights, we introduce modifications to projectors with training and structural regularizations. Empirically, our methods exhibit consistent improvement in the downstream performance across various real-world datasets, including CIFAR-10, CIFAR-100, and ImageNet-100. We believe our theoretical understanding on the role of the projection head will in
Discrete diffusion language models learn to reconstruct text from randomly masked inputs, yet under mild assumptions their denoiser already implements the exact Bayesian posterior over the original tokens. We prove that the expected denoiser output under the forward corruption distribution recovers the true posterior, and that a simple Monte Carlo estimator converges to this posterior at rate O(1/sqrt(K)) with finite-sample concentration bounds. Building on this insight, we introduce an inference-time ensemble that runs K independent denoising passes and aggregates both posterior means and variances without any extra training. On WikiText-2, our MC-marginal sampler recovers the analytic lambda-DCE zero-shot perplexity (approximately 39) to within a few points at K=128, and its per-token variance shows a strong rank correlation with reconstruction error (Spearman rho = 0.996). This cost-proportional procedure yields calibrated uncertainty estimates and a direct trade-off between compute and posterior fidelity in discrete diffusion LMs.
Autoregressive (AR) transformers have emerged as a powerful paradigm for visual generation, largely due to their scalability, computational efficiency and unified architecture with language and vision. Among them, next scale prediction Visual Autoregressive Generation (VAR) has recently demonstrated remarkable performance, even surpassing diffusion-based models. In this work, we revisit VAR and uncover a theoretical insight: when equipped with a Markovian attention mask, VAR is mathematically equivalent to a discrete diffusion. We term this reinterpretation as Scalable Visual Refinement with Discrete Diffusion (SRDD), establishing a principled bridge between AR transformers and diffusion models. Leveraging this new perspective, we show how one can directly import the advantages of diffusion such as iterative refinement and reduce architectural inefficiencies into VAR, yielding faster convergence, lower inference cost, and improved zero-shot reconstruction. Across multiple datasets, we show that the diffusion based perspective of VAR leads to consistent gains in efficiency and generation.
GRPO has emerged as a prominent reinforcement learning algorithm for post-training LLMs. Unlike critic-based methods, GRPO computes advantages by estimating the \emph{value baselines} from group-level statistics, eliminating the need for a critic network. Consequently, the prevailing view emphasizes the necessity of large group sizes, which are assumed to yield more accurate statistical estimates. In this paper, we propose a different view that the efficacy of GRPO stems from its implicit contrastive objective in the optimization, which helps reduce variance via the control variate method. This makes GRPO structurally related to preference learning methods such as DPO. This perspective motivates 2-GRPO, a minimal group-size variant that constructs contrastive signals with only two rollouts. We provide a rigorous theoretical analysis of 2-GRPO and empirically validate its effectiveness: 2-GRPO retains $97.6\%$ of the performance of 16-GRPO, while requiring only $12.5\%$ of the rollouts and $21\%$ of the training time.
Process reward models (PRMs) allow for fine-grained credit assignment in reinforcement learning (RL), and seemingly contrast with outcome reward models (ORMs), which assign a single reward to an entire trajectory. However, we provide theoretical proof in this work that the Group Relative Policy Optimization (GRPO) RL algorithm equipped with an ORM is in fact equivalent to a PRM-aware RL objective equipped with a non-trivial, Monte-Carlo-based PRM (given mild assumptions). Leveraging the framework of GRPO-as-a-PRM, we identify a flaw in the GRPO objective that interacts with imbalanced process steps and rewards to hinder both exploration and exploitation (under different conditions). We propose a simple modification to the algorithm to mitigate this defect ($λ$-GRPO), and show that LLMs tuned with $λ$-GRPO outperform LLMs tuned with standard GRPO on downstream reasoning tasks\textemdash and reach peak performance more rapidly. These results show that we can leverage the hidden, built-in PRM structure within the vanilla GRPO algorithm to boost model performance without employing an explicit PRM, and with a negligible impact on training time and cost.
Gaussian Splatting has become a popular technique for various 3D Computer Vision tasks, including novel view synthesis, scene reconstruction, and dynamic scene rendering. However, the challenge of natural-looking object insertion, where the object's appearance seamlessly matches the scene, remains unsolved. In this work, we propose a method, dubbed D3DR, for inserting a 3DGS-parametrized object into a 3DGS scene while correcting its lighting, shadows, and other visual artifacts to ensure consistency. We reveal a hidden ability of diffusion models trained on large real-world datasets to implicitly understand correct scene lighting, and leverage it in our pipeline. After inserting the object, we optimize a diffusion-based Delta Denoising Score (DDS)-inspired objective to adjust its 3D Gaussian parameters for proper lighting correction. We introduce a novel diffusion personalization technique that preserves object geometry and texture across diverse lighting conditions, and utilize it to achieve consistent identity matching between original and inserted objects. Finally, we demonstrate the effectiveness of the method by comparing it to existing approaches, achieving 2.0 dB PSNR improv
Recent advancements in image generation have enabled the creation of high-quality images from text conditions. However, when facing multi-modal conditions, such as text combined with reference appearances, existing methods struggle to balance multiple conditions effectively, typically showing a preference for one modality over others. To address this challenge, we introduce EMMA, a novel image generation model accepting multi-modal prompts built upon the state-of-the-art text-to-image (T2I) diffusion model, ELLA. EMMA seamlessly incorporates additional modalities alongside text to guide image generation through an innovative Multi-modal Feature Connector design, which effectively integrates textual and supplementary modal information using a special attention mechanism. By freezing all parameters in the original T2I diffusion model and only adjusting some additional layers, we reveal an interesting finding that the pre-trained T2I diffusion model can secretly accept multi-modal prompts. This interesting property facilitates easy adaptation to different existing frameworks, making EMMA a flexible and effective tool for producing personalized and context-aware images and even videos.
Quantifying uncertainty in neural networks is a highly relevant problem which is essential to many applications. The two predominant paradigms to tackle this task are Bayesian neural networks (BNNs) and deep ensembles. Despite some similarities between these two approaches, they are typically surmised to lack a formal connection and are thus understood as fundamentally different. BNNs are often touted as more principled due to their reliance on the Bayesian paradigm, whereas ensembles are perceived as more ad-hoc; yet, deep ensembles tend to empirically outperform BNNs, with no satisfying explanation as to why this is the case. In this work we bridge this gap by showing that deep ensembles perform exact Bayesian averaging with a posterior obtained with an implicitly learned data-dependent prior. In other words deep ensembles are Bayesian, or more specifically, they implement an empirical Bayes procedure wherein the prior is learned from the data. This perspective offers two main benefits: (i) it theoretically justifies deep ensembles and thus provides an explanation for their strong empirical performance; and (ii) inspection of the learned prior reveals it is given by a mixture of
While large language models (LLMs) excel on generation tasks, their decoder-only architecture often limits their potential as embedding models if no further representation finetuning is applied. Does this contradict their claim of generalists? To answer the question, we take a closer look at Mixture-of-Experts (MoE) LLMs. Our study shows that the expert routers in MoE LLMs can serve as an off-the-shelf embedding model with promising performance on a diverse class of embedding-focused tasks, without requiring any finetuning. Moreover, our extensive analysis shows that the MoE routing weights (RW) is complementary to the hidden state (HS) of LLMs, a widely-used embedding. Compared to HS, we find that RW is more robust to the choice of prompts and focuses on high-level semantics. Motivated by the analysis, we propose MoEE combining RW and HS, which achieves better performance than using either separately. Our exploration of their combination and prompting strategy shed several novel insights, e.g., a weighted sum of RW and HS similarities outperforms the similarity on their concatenation. Our experiments are conducted on 6 embedding tasks with 20 datasets from the Massive Text Embeddi
This paper reveals a novel linear characteristic exclusive to transformer decoders, including models such as GPT, LLaMA, OPT, BLOOM and others. We analyze embedding transformations between sequential layers, uncovering a near-perfect linear relationship (Procrustes similarity score of 0.99). However, linearity decreases when the residual component is removed due to a consistently low output norm of the transformer layer. Our experiments show that removing or linearly approximating some of the most linear blocks of transformers does not affect significantly the loss or model performance. Moreover, in our pretraining experiments on smaller models we introduce a cosine-similarity-based regularization, aimed at reducing layer linearity. This regularization improves performance metrics on benchmarks like Tiny Stories and SuperGLUE and as well successfully decreases the linearity of the models. This study challenges the existing understanding of transformer architectures, suggesting that their operation may be more linear than previously assumed.
Despite large neural networks demonstrating remarkable abilities to complete different tasks, they require excessive memory usage to store the optimization states for training. To alleviate this, the low-rank adaptation (LoRA) is proposed to reduce the optimization states by training fewer parameters. However, LoRA restricts overall weight update matrices to be low-rank, limiting the model performance. In this work, we investigate the dynamics of LoRA and identify that it can be approximated by a random projection. Based on this observation, we propose Flora, which is able to achieve high-rank updates by resampling the projection matrices while enjoying the sublinear space complexity of optimization states. We conduct experiments across different tasks and model architectures to verify the effectiveness of our approach.
As the parameter size of large language models (LLMs) continues to expand, the need for a large memory footprint and high communication bandwidth have become significant bottlenecks for the training and inference of LLMs. To mitigate these bottlenecks, various tensor compression techniques have been proposed to reduce the data size, thereby alleviating memory requirements and communication pressure. Our research found that video codecs, despite being originally designed for compressing videos, show excellent efficiency when compressing various types of tensors. We demonstrate that video codecs can be versatile and general-purpose tensor codecs while achieving the state-of-the-art compression efficiency in various tasks. We further make use of the hardware video encoding and decoding module available on GPUs to create a framework capable of both inference and training with video codecs repurposed as tensor codecs. This greatly reduces the requirement for memory capacity and communication bandwidth, enabling training and inference of large models on consumer-grade GPUs.
Classifier-free guidance is an effective sampling technique in diffusion models that has been widely adopted. The main idea is to extrapolate the model in the direction of text guidance and away from null-text guidance. In this paper, we demonstrate that null-text guidance in diffusion models is secretly a cartoon-style creator, i.e., the generated images can be efficiently transformed into cartoons by simply perturbing the null-text guidance. Specifically, we proposed two disturbance methods, i.e., Rollback disturbance (Back-D) and Image disturbance (Image-D), to construct misalignment between the noisy images used for predicting null-text guidance and text guidance (subsequently referred to as \textbf{null-text noisy image} and \textbf{text noisy image} respectively) in the sampling process. Back-D achieves cartoonization by altering the noise level of null-text noisy image via replacing $x_t$ with $x_{t+Δt}$. Image-D, alternatively, produces high-fidelity, diverse cartoons by defining $x_t$ as a clean input image, which further improves the incorporation of finer image details. Through comprehensive experiments, we delved into the principle of noise disturbing for null-text and