论文检索

Deep Learning · Everything Else

Ruihan Yang, yikai zhang, Aili Chen, Xintao Wang, Jiangjie Chen, Siyu Yuan, Deqing Yang, Yanghua Xiao

Large language models (LLMs) have enabled agents to perform complex reasoning and decision-making through free-form language interactions. However, in open-ended language action environments (e.g., negotiation or question-asking games), the action space can be formulated as a joint distribution over tokens, resulting in an extremely large and combinatorial action space. Sampling actions in such a space can lead to extreme reward sparsity, which brings large reward variance, hindering effective reinforcement learning (RL). To address this, we propose **ARIA**, a method that **A**ggregates **R**ewards in **I**ntention space to enable efficient and effective language **A**gents training. ARIA aims to project natural language actions from the high-dimensional joint token distribution space into a low-dimensional intention space, where semantically similar actions are clustered and assigned shared rewards. This intention-aware reward aggregation reduces reward variance by densifying reward signals, fostering efficient and effective policy optimization. Extensive experiments demonstrate that ARIA not only significantly reduces gradient variance, but also delivers substantial performance gains of average 9.95% across four downstream tasks (e.g., negotiation and text-based games), consistently outperforming strong offline and online RL baselines.

Deep Learning · Everything Else

Lei Wang, Jieming Bian, Letian Zhang, Jie Xu

Large Language Models (LLMs) have demonstrated impressive capabilities across various tasks, but fine-tuning them for domain-specific applications often requires substantial domain-specific data that may be distributed across multiple organizations. Federated Learning (FL) offers a privacy-preserving solution, but faces challenges with computational constraints when applied to LLMs. Low-Rank Adaptation (LoRA) has emerged as a parameter-efficient fine-tuning approach, though a single LoRA module often struggles with heterogeneous data across diverse domains. This paper addresses two critical challenges in federated LoRA fine-tuning: 1. determining the optimal number and allocation of LoRA experts across heterogeneous clients, and 2. enabling clients to selectively utilize these experts based on their specific data characteristics. We propose FedLEASE (Federated adaptive LoRA Expert Allocation and SElection), a novel framework that adaptively clusters clients based on representation similarity to allocate and train domain-specific LoRA experts. It also introduces an adaptive top-$M$ Mixture-of-Experts mechanism that allows each client to select the optimal number of utilized experts. Our extensive experiments on diverse benchmark datasets demonstrate that FedLEASE significantly outperforms existing federated fine-tuning approaches in heterogeneous client settings while maintaining communication efficiency.

Reinforcement Learning · Everything Else

Josiah Hanna, Nicholas Corrado

Recent work on large language models has demonstrated the use of model-free reinforcement learning (RL) to train reasoning-like capabilities. The emergence of "thinking" through model-free RL is interesting as thinking actions neither produce reward nor change the external world state to one where the agent is more likely to get reward. This paper seeks to build a domain-independent understanding of when model-free RL will lead to such "thinking" as a strategy for reward maximization. To build this understanding, we first introduce a theoretical model which we call a thought Markov decision process (MDP). Thought MDPs minimally extend the classical MDP model to include an abstract notion of thought state and thought action. Using the thought MDP model, we prove the importance of policy initialization in determining whether or not thinking emerges and show formally that thought actions are equivalent to the agent choosing to perform a step of policy improvement before continuing to act. We then show that open-source LLMs satisfy the conditions that our theory predicts are necessary for model-free RL to produce thinking-like behavior. Finally, we hypothesize sufficient conditions that would enable thinking to be learned outside of language generation and introduce a toy domain where a combination of multi-task pre-training and designated thought actions enable more data-efficient RL compared to non-thinking agents.

General Machine Learning · Representation Learning

Arman Behnam, Binghui Wang

Causal representation learning in the anti-causal setting—labels cause features rather than the reverse—presents unique challenges requiring specialized approaches. We propose Anti-Causal Invariant Abstractions (ACIA), a novel measure-theoretic framework for anti-causal representation learning. ACIA employs a two-level design: low-level representations capture how labels generate observations, while high-level representations learn stable causal patterns across environment-specific variations. ACIA addresses key limitations of existing approaches by: (1) accommodating prefect and imperfect interventions through interventional kernels, (2) eliminating dependency on explicit causal structures, (3) handling high-dimensional data effectively, and (4) providing theoretical guarantees for out-of-distribution generalization. Experiments on synthetic and real-world medical datasets demonstrate that ACIA consistently outperforms state-of-the-art methods in both accuracy and invariance metrics. Furthermore, our theoretical results establish tight bounds on performance gaps between training and unseen environments, confirming the efficacy of our approach for robust anti-causal learning. {{Code is available at \url{https://github.com/ArmanBehnam/ACIA}}}.

Tiezheng Zhang, Yitong Li, Yu-Cheng Chou, Jieneng Chen, Alan Yuille, Chen Wei, Junfei Xiao

Building state-of-the-art Vision-Language Models (VLMs) with strong captioning capabilities typically necessitates training on billions of high-quality image-text pairs, requiring millions of GPU hours. This paper introduces the Vision-Language-Vision **(VLV)** auto-encoder framework, which strategically leverages key pretrained components: a vision encoder, the decoder of a Text-to-Image (T2I) diffusion model, and subsequently, a Large Language Model (LLM). Specifically, we establish an information bottleneck by regularizing the language representation space, achieved through freezing the pretrained T2I diffusion decoder. Our VLV pipeline effectively distills knowledge from the text-conditioned diffusion model using continuous embeddings, demonstrating comprehensive semantic understanding via high-quality reconstructions. Furthermore, by fine-tuning a pretrained LLM to decode the intermediate language representations into detailed descriptions, we construct a state-of-the-art (SoTA) captioner comparable to leading models like GPT-4o and Gemini 2.0 Flash. Our method demonstrates exceptional cost-efficiency and significantly reduces data requirements; by primarily utilizing single-modal images for training and maximizing the utility of existing pretrained models (image encoder, T2I diffusion model, and LLM), it circumvents the need for massive paired image-text datasets, keeping the total training expenditure under $1,000 USD.

Optimization · Neural Network Optimization

Huan Li, Yiming Dong, Zhouchen Lin

As the default optimizer for training large language models, AdamW has achieved remarkable success in deep learning. However, its convergence behavior is not theoretically well-understood. This paper establishes the convergence rate $\frac{1}{K}\sum_{k=1}^K E[||\nabla f(x^k)||_1]\leq O(\frac{\sqrt{d}C}{K^{1/4}})$ for AdamW measured by $\ell_1$ norm, where K represents the iteration number, d denotes the model dimension, and C matches the constant in the optimal convergence rate of SGD. Theoretically, we have $E[||\nabla f(x)||_1]\geq\sqrt{\frac{2d}{\pi}}E[||\nabla f(x)||_2]$ when each element of $\nabla f(x)$ is generated from Gaussian distribution $\mathcal N(0,1)$. Empirically, our experimental results on real-world deep learning tasks reveal $||\nabla f(x)||_1=\varTheta(\sqrt{d})||\nabla f(x)||_2$. Both support that our convergence rate can be considered to be analogous to the optimal convergence rate of SGD.

General Machine Learning · Representation Learning

Hugues Van Assel, Mark Ibrahim, Tommaso Biancalani, Aviv Regev, Randall Balestriero

Reconstruction and joint-embedding have emerged as two leading paradigms in Self‑Supervised Learning (SSL). Reconstruction methods focus on recovering the original sample from a different view in input space. On the other hand, joint-embedding methods align the representations of different views in latent space. Both approaches offer compelling advantages, yet practitioners lack clear guidelines for choosing between them. In this work, we unveil the core mechanisms that distinguish each paradigm. By leveraging closed-form solutions for both approaches, we precisely characterize how the view generation process, e.g. data augmentation, impacts the learned representations. We then demonstrate that, unlike supervised learning, both SSL paradigms require a minimal alignment between augmentations and irrelevant features to achieve asymptotic optimality with increasing sample size. Our findings indicate that in scenarios where these irrelevant features have a large magnitude, joint-embedding methods are preferable because they impose a strictly weaker alignment condition compared to reconstruction-based methods. These results not only clarify the trade-offs between the two paradigms but also substantiate the empirical success of joint-embedding approaches on real-world challenging datasets.

General Machine Learning · Everything Else

Guanghao Li, Wenhao Jiang, Mingfeng Chen, Yan Li, Hao Yu, Shuting Dong, Tao Ren, Ming Tang, Chun Yuan

Chain-of-Thought (CoT) prompting improves the reasoning performance of large language models (LLMs) by encouraging step-by-step thinking. However, CoT-based methods depend on intermediate reasoning steps, which limits scalability and generalization. Recent work explores recursive reasoning, where LLMs reuse internal layers across iterations to refine latent representations without explicit CoT supervision. While promising, these approaches often require costly pretraining and lack a principled framework for how reasoning should evolve across iterations. We address this gap by introducing **Flow Chain-of-Thought (Flow CoT)**, a reasoning paradigm that models recursive inference as a progressive trajectory of latent cognitive states. Flow CoT frames each iteration as a distinct cognitive stage—deepening reasoning across iterations without relying on manual supervision. To realize this, we propose **SCOUT** (*Stepwise Cognitive Optimization Using Teachers*), a lightweight fine-tuning framework that enables Flow CoT-style reasoning without the need for pretraining. SCOUT uses progressive distillation to align each iteration with a teacher of appropriate capacity, and a cross-attention-based retrospective module that integrates outputs from previous iterations while preserving the model’s original computation flow. Experiments across eight reasoning benchmarks show that SCOUT consistently improves both accuracy and explanation quality, achieving up to 1.8\% gains under fine-tuning. Qualitative analyses further reveal that SCOUT enables progressively deeper reasoning across iterations—refining both belief formation and explanation granularity. These results not only validate the effectiveness of SCOUT, but also demonstrate the practical viability of Flow CoT as a scalable framework for enhancing reasoning in LLMs.

Computer Vision · Image and Video Generation

Nate Gillman, Charles Herrmann, Michael Freeman, Daksh Aggarwal, Evan Luo, Deqing Sun, Chen Sun

Recent advances in video generation models have sparked interest in world models capable of simulating realistic environments. While navigation has been well-explored, physically meaningful interactions that mimic real-world forces remain largely understudied. In this work, we investigate using physical forces as a control signal for video generation and propose force prompts which enable users to interact with images through both localized point forces, such as poking a plant, and global wind force fields, such as wind blowing on fabric. We demonstrate that these force prompts can enable videos to respond realistically to physical control signals by leveraging the physical prior in the original pretrained model, without using any 3D asset or physics simulator at inference. The primary challenge of force prompting is the difficulty in obtaining high quality paired force-video training data, both in the real world due to the difficulty of obtaining force signals, and in synthetic data due to limitations in the visual quality and domain diversity of physics simulators. Our key finding is that video generation models can *generalize* remarkably well when adapted to follow physical force conditioning from videos synthesized by Blender, even with limited demonstrations of few objects (e.g., flying flags, rolling balls, etc.). Our method can generate videos which simulate forces across diverse geometries, settings, and materials. We also try to understand the source of this generalization and perform ablations on the training data that reveal two key elements: visual diversity and the use of specific text keywords during training. Our approach is trained on only around 15k training examples for a single day on four A100 GPUs, and outperforms existing methods on force adherence and physics realism, bringing world models closer to real-world physics interactions. All datasets, code, and model weights will be open-sourced. Video examples can be found at https://sites.google.com/view/force-prompting-neurips2025

Optimization · Large Scale, Parallel and Distributed

Xinhao Luo, Zihan Liu, Yangjie Zhou, Shihan Fang, Ziyu Huang, Yu Feng, Chen Zhang, Shixuan Sun, Zhenzhe Zheng, Jingwen Leng 等

Large language model (LLM) decoding suffers from high latency due to fragmented execution across operators and heavy reliance on off-chip memory for data exchange and reduction. This execution model limits opportunities for fusion and incurs significant memory traffic and kernel launch overhead. While modern architectures such as NVIDIA Hopper provide distributed shared memory and low-latency intra-cluster interconnects, they expose only low-level data movement instructions, lacking structured abstractions for collective on-chip communication. To bridge this software-hardware gap, we introduce two cluster-level communication primitives, ClusterReduce and ClusterGather, which abstract common communication patterns and enable structured, high-speed data exchange and reduction between thread blocks within a cluster, allowing intermediate results to be on-chip without involving off-chip memory. Building on these abstractions, we design ClusterFusion, an execution framework that schedules communication and computation jointly to expand operator fusion scope by composing decoding stages such as QKV Projection, Attention, and Output Projection into a single fused kernels. Evaluations on H100 GPUs show that ClusterFusion outperforms state-of-the-art inference frameworks by $1.61\times$ on average in end-to-end latency across different models and configurations.

Social Aspects · Accountability, Transparency and Interpretability

Xinpeng Wang, Mingyang Wang, Yihong Liu, Hinrich Schuetze, Barbara Plank

Refusal mechanisms in large language models (LLMs) are essential for ensuring safety. Recent research has revealed that refusal behavior can be mediated by a single direction in activation space, enabling targeted interventions to bypass refusals. While this is primarily demonstrated in an English-centric context, appropriate refusal behavior is important for any language, but poorly understood. In this paper, we investigate the refusal behavior in LLMs across 14 languages using \textit{PolyRefuse}, a multilingual safety dataset created by translating malicious and benign English prompts into these languages. We uncover the surprising cross-lingual universality of the refusal direction: a vector extracted from English can bypass refusals in other languages with near-perfect effectiveness, without any additional fine-tuning. Even more remarkably, refusal directions derived from any safety-aligned language transfer seamlessly to others. We attribute this transferability to the parallelism of refusal vectors across languages in the embedding space and identify the underlying mechanism behind cross-lingual jailbreaks. These findings provide actionable insights for building more robust multilingual safety defenses and pave the way for a deeper mechanistic understanding of cross-lingual vulnerabilities in LLMs.

Theory · Domain Adaptation and Transfer Learning

Konstantinos Dafnis, Dimitris Metaxas

Vision–language models (VLMs) excel at zero-shot inference but often degrade under test-time domain shifts. For this reason, episodic test-time adaptation strategies have recently emerged as powerful techniques for adapting VLMs to a single unlabeled image. However, existing adaptation strategies, such as test-time prompt tuning, typically require backpropagating through large encoder weights or altering core model components. In this work, we introduce \textbf{S}pectrum-Aware \textbf{T}est-Time \textbf{S}teering (\textbf{STS}), a \textit{lightweight adaptation framework} that extracts a spectral subspace from the textual embeddings to define principal semantic directions, and learns to steer latent representations in a spectrum-aware manner by adapting a small number of per-sample shift parameters to minimize entropy across augmented views. STS operates entirely at inference in the latent space, without backpropagation through or modification of the frozen encoders. Building on standard evaluation protocols, our comprehensive experiments demonstrate that STS largely surpasses or compares favorably against state-of-the-art test-time adaptation methods, while introducing only a handful of additional parameters and achieving inference speeds up to 8× faster with a 12× smaller memory footprint than conventional test-time prompt tuning. The code is available at \url{https://github.com/kdafnis/STS}.

Applications · Time Series

Yinghui Liu, Hao Miao, Guojiang Shen, Yan Zhao, Xiangjie Kong, Ivan Lee

Out-of-town trip recommendation aims to generate a sequence of Points of Interest (POIs) for users traveling from their hometowns to previously unvisited regions based on personalized itineraries, e.g., origin, destination, and trip duration. Modeling the complex user preferences--which often exhibit a two-fold nature of static and dynamic interests--is critical for effective recommendations. However, the sparsity of out-of-town check-in data presents significant challenges in capturing such user preferences. Meanwhile, existing methods often conflate the static and dynamic preferences, resulting in suboptimal performance. In this paper, we for the first time systematically study the problem of out-of-town trip recommendation. A novel framework SPOT-Trip is proposed to explicitly learns the dual static-dynamic user preferences. Specifically, to handle scarce data, we construct a POI attribute knowledge graph to enrich the semantic modeling of users’ hometown and out-of-town check-ins, enabling the static preference modeling through attribute relation-aware aggregation. Then, we employ neural ordinary differential equations (ODEs) to capture the continuous evolution of latent dynamic user preferences and innovatively combine a temporal point process to describe the instantaneous probability of each preference behavior. Further, a static-dynamic fusion module is proposed to merge the learned static and dynamic user preferences. Extensive experiments on real data offer insight into the effectiveness of the proposed solutions, showing that SPOT-Trip achieves performance improvement by up to 17.01%.

General Machine Learning · Representation Learning

Guanming Zhang, David Heeger, Stefano Martiniani

Contrastive self-supervised learning based on point-wise comparisons has been widely studied for vision tasks. In the visual cortex of the brain, neuronal responses to distinct stimulus classes are organized into geometric structures known as neural manifolds. Accurate classification of stimuli can be achieved by effectively separating these manifolds, akin to solving a packing problem. We introduce Contrastive Learning As Manifold Packing (CLAMP), a self-supervised framework that recasts representation learning as a manifold packing problem. CLAMP introduces a loss function inspired by the potential energy of short-range repulsive particle systems, such as those encountered in the physics of simple liquids and jammed packings. In this framework, each class consists of sub-manifolds embedding multiple augmented views of a single image. The sizes and positions of the sub-manifolds are dynamically optimized by following the gradient of a packing loss. This approach yields interpretable dynamics in the embedding space that parallel jamming physics, and introduces geometrically meaningful hyperparameters within the loss function. Under the standard linear evaluation protocol, which freezes the backbone and trains only a linear classifier, CLAMP achieves competitive performance with state-of-the-art self-supervised models. Furthermore, our analysis reveals that neural manifolds corresponding to different categories emerge naturally and are effectively separated in the learned representation space, highlighting the potential of CLAMP to bridge insights from physics, neural science, and machine learning.

Theory · Reinforcement Learning and Planning

Junyu guo, Zhi Zheng, Donghao Ying, Ming Jin, Shangding Gu, Costas J Spanos, Javad Lavaei

Constrained reinforcement learning (RL) seeks high-performance policies under safety constraints. We focus on an offline setting where the agent learns from a fixed dataset—a common requirement in realistic tasks to prevent unsafe exploration. To address this, we propose Diffusion-Regularized Constrained Offline Reinforcement Learning (DRCORL), which first uses a diffusion model to capture the behavioral policy from offline data and then extracts a simplified policy to enable efficient inference. We further apply gradient manipulation for safety adaptation, balancing the reward objective and constraint satisfaction. This approach leverages high-quality offline data while incorporating safety requirements. Empirical results show that DRCORL achieves reliable safety performance, fast inference, and strong reward outcomes across robot learning tasks. Compared to existing safe offline RL methods, it consistently meets cost limits and performs well with the same hyperparameters, indicating practical applicability in real-world scenarios. We open-source our implementation at https://github.com/JamesJunyuGuo/DRCORL.

Computer Vision · Classification and Understanding

Tianhong Zhou, xu yin, Yingtao Zhu, Chuxi Xiao, Haiyang Bian, Lei Wei, Xuegong Zhang

Vision–language models (VLMs) exhibit strong zero-shot generalization on natural images and show early promise in interpretable medical image analysis. However, existing benchmarks do not systematically evaluate whether these models truly reason like human clinicians or merely imitate superficial patterns. To address this gap, we propose DrVD-Bench, the first multimodal benchmark for clinical visual reasoning. DrVD-Bench consists of three modules: Visual Evidence Comprehension, Reasoning Trajectory Assessment, and Report Generation Evaluation, comprising a total of 7,789 image–question pairs. Our benchmark covers 20 task types, 17 diagnostic categories, and five imaging modalities—CT, MRI, ultrasound, radiography, and pathology. DrVD-Bench is explicitly structured to reflect the clinical reasoning workflow from modality recognition to lesion identification and diagnosis. We benchmark 19 VLMs, including general-purpose and medical-specific, open-source and proprietary models, and observe that performance drops sharply as reasoning complexity increases. While some models begin to exhibit traces of human-like reasoning, they often still rely on shortcut correlations rather than grounded visual understanding. DrVD-Bench offers a rigorous and structured evaluation framework to guide the development of clinically trustworthy VLMs.

Deep Learning · Theory

Yunwei Ren, Jason Lee

The information exponent (Ben Arous et al. [2021]) and its extensions --- which are equivalent to the lowest degree in the Hermite expansion of the link function (after a potential label transform) for Gaussian single-index models --- have played an important role in predicting the sample complexity of online stochastic gradient descent (SGD) in various learning tasks. In this work, we demonstrate that, for multi-index models, focusing solely on the lowest degree can miss key structural details of the model and result in suboptimal rates. Specifically, we consider the task of learning target functions of form $f_*(x) = \sum_{k=1}^{P} \phi(v_k^* \cdot x)$, where $P \le d$, the ground-truth directions $\\{ v_k^* \\}_{k=1}^P$ are orthonormal, and the information exponent of $\phi$ is $L$. Based on the theory of information exponent, when $L = 2$, only the relevant subspace (not the exact directions) can be recovered due to the rotational invariance of the second-order terms, and when $L > 2$, recovering the directions using online SGD require $\tilde{O}(P d^{L-1})$ samples. In this work, we show that by considering both second- and higher-order terms, we can first learn the relevant space using the second-order terms, and then the exact directions using the higher-order terms, and the overall sample and complexity of online SGD is $\tilde{O}( d P^{L-1} )$.

Applications

Wenjun Wu, Lingling Zhang, Bo Zhao, Muye Huang, QianYing Wang, Jun Liu

The task of geometry problem solving has been a long-standing focus in the automated mathematics community and draws growing attention due to its complexity for both symbolic and neural models. Although prior studies have explored various effective approaches for enhancing problem solving performances, two fundamental challenges remain unaddressed, which are essential to the application in practical scenarios. First, the multi-step reasoning gap between the initial geometric conditions and ultimate problem goal leads to a great search space for solution exploration. Second, obtaining multiple interpretable and shorter solutions remains an open problem. In this work, we introduce the Causal-Reasoning Geometry Problem Solver to overcome these challenges. Specifically, the Causal Graph Reasoning theory is proposed to perform symbolic reasoning before problem solving. Several causal graphs are constructed according to predefined rule base, where each graph is composed of primitive nodes, causal edges and prerequisite edges. By applying causal graph deduction from initial conditions, the reachability status of nodes are iteratively conveyed by causal edges until reaching the target nodes, representing feasible causal deduction paths. In this way, the search space of solutions is compressed from the beginning, the end and intermediate reasoning paths, while ensuring the interpretability and variety of solutions. To achieve this, we further propose Forward Matrix Deduction which transforms the causal graphs into matrices and vectors, and applies matrix operations to update the status value of reachable nodes in iterations. Finally, multiple solutions can be generated by tracing back from the target nodes after validation. Experiments demonstrate the effectiveness of our method to obtain multiple shorter and interpretable solutions. Code is available after acceptance.

Applications

Leqi Zheng, Chaokun Wang, Zixin Song, Cheng Wu, Shannan Yan, Jiajun Zhang, Ziyang Liu

Traditional recommender systems have relied heavily on positive feedback for learning user preferences, while the abundance of negative feedback in real-world scenarios remains underutilized. To address this limitation, recent years have witnessed increasing attention on leveraging negative feedback in recommender systems to enhance recommendation performance. However, existing methods face three major challenges: limited model compatibility, ineffective information exchange, and computational inefficiency. To overcome these challenges, we propose a model-agnostic Signed Dual-Channel Graph Contrastive Learning (SDCGCL) framework that can be seamlessly integrated with existing graph contrastive learning methods. The framework features three key components: (1) a Dual-Channel Graph Embedding that separately processes positive and negative graphs, (2) a Cross-Channel Distribution Calibration mechanism to maintain structural consistency, and (3) an Adaptive Prediction Strategy that effectively combines signals from both channels. Building upon this framework, we further propose a Dual-channel Feedback Fusion (DualFuse) model and develop a two-stage optimization strategy to ensure efficient training. Extensive experiments on four public datasets demonstrate that our approach consistently outperforms state-of-the-art baselines by substantial margins while exhibiting minimal computational complexity.

Deep Learning · Robustness

Wenquan Lu, Jiaqi Zhang, Hugues Van Assel, Randall Balestriero

Self-Supervised Learning (SSL) has become a powerful solution to extract rich representations from unlabeled data. Yet, SSL research is mostly focused on clean, curated and high-quality datasets. As a result, applying SSL on noisy data remains a challenge, despite being crucial to applications such as astrophysics, medical imaging, geophysics or finance. In this work, we present a fully self-supervised framework that enables noise-robust representation learning without requiring a denoiser at inference or downstream fine-tuning. Our method first trains an SSL denoiser on noisy data, then uses it to construct a denoised-to-noisy data curriculum (i.e., training first on denoised, then noisy samples) for pretraining a SSL backbone (e.g., DINOv2), combined with a teacher-guided regularization that anchors noisy embeddings to their denoised counterparts. This process encourages the model to internalize noise robustness. Notably, the denoiser can be discarded after pretraining, simplifying deployment. On ImageNet-1k with ViT-B under extreme Gaussian noise ($\sigma=255$, SNR = 0.72 dB), our method improves linear probing accuracy by 4.8\% over DINOv2, demonstrating that denoiser-free robustness can emerge from noise-aware pretraining. The code is available at https://github.com/wenquanlu/noisy_dinov2.