论文检索

Reinforcement Learning · Everything Else

Junhong Shen, Hao Bai, Lunjun Zhang, Yifei Zhou, Amrith Setlur, Peter Tong, Diego Caples, Nan Jiang, Tong Zhang, Ameet Talwalkar 等

Test-time scaling in agentic tasks often relies on generating long reasoning traces ("think" more) before acting, but this does not allow agents to acquire new information from the environment or adapt behavior over time. In this work, we propose scaling test-time interaction, an untapped dimension for test-time scaling that increases the agent's interaction horizon to enable rich behaviors such as exploration, backtracking, and dynamic re-planning within a single rollout. To demonstrate the promise of this scaling dimension, we situate our study in the domain of web agents. We first show that even prompting-based interaction scaling can improve task success on web benchmarks non-trivially. Building on this, we introduce TTI, a curriculum-based online reinforcement learning (RL) approach that trains agents by adaptively adjusting their interaction lengths during rollout. Using a Gemma 3 12B model, TTI sets a new state-of-the-art among open-source agents trained on public data on WebVoyager and WebArena. Case studies further reveal that TTI enables agents to balance exploration and exploitation adaptively. Our results establish interaction scaling as a powerful, complementary axis to scaling per-action compute, offering new avenues for training robust and adaptive agents.

Computer Vision · Image and Video Generation

Huaisheng Zhu, Teng Xiao, Shijie Zhou, Zhimeng Guo, Hangfan Zhang, Siyuan Xu, Vasant Honavar

Diffusion models have established themselves as leading techniques for image generation. However, their reliance on an iterative denoising process results in slow sampling speeds, which limits their applicability to interactive and creative applications. An approach to overcoming this limitation involves distilling multistep diffusion models into efficient one-step generators. However, existing distillation methods typically suffer performance degradation or require complex iterative training procedures which increase their complexity and computational cost. In this paper, we propose Contrastive Energy Distillation (CED), a simple yet effective approach to distill multistep diffusion models into effective one-step generators. Our key innovation is the introduction of an unnormalized joint energy-based model (EBM) that represents the generator and an auxiliary score model. CED optimizes a Noise Contrastive Estimation (NCE) objective to efficiently transfers knowledge from a multistep teacher diffusion model without additional modules or iterative training complexity. We further show that CED implicitly optimizes the KL divergence between the distributions modeled by the multistep diffusion model and the one-step generator. We present results of experiments which demonstrate that CED achieves competitive performance with the representative baselines for distilling multistep diffusion models while maintaining excellent memory efficiency.

Reinforcement Learning · Deep RL

Yaorui Shi, Sihang Li, Chang Wu, ZHIYUAN LIU, Junfeng Fang, Hengxing Cai, An Zhang, Xiang Wang

Large language models have demonstrated impressive reasoning capabilities but are inherently limited by their knowledge reservoir. Retrieval-augmented reasoning mitigates this limitation by allowing LLMs to query external resources, but existing methods often retrieve irrelevant or noisy information, hindering accurate reasoning. In this paper, we propose **AutoRefine**, a reinforcement learning post-training framework that adopts a new "search-and-refine-during-think" paradigm. AutoRefine introduces explicit knowledge refinement steps between successive search calls, enabling the model to iteratively filter, distill, and organize evidence before generating an answer. Furthermore, we incorporate tailored retrieval-specific rewards alongside answer correctness rewards using group relative policy optimization. Experiments on single-hop and multi-hop QA benchmarks demonstrate that AutoRefine significantly outperforms existing approaches, particularly in complex, multi-hop reasoning scenarios. Detailed analysis shows that AutoRefine issues frequent, higher-quality searches and synthesizes evidence effectively.

Optimization · Stochastic & Convex Optimization

Junyu Huang, Qilong Feng, Zhen Zhang, Beirong Cui, Jianxin Wang

Local search is a powerful clustering technique that provides high-quality solutions with theoretical guarantees. With distance-based sampling strategies, local search methods can achieve constant approximations for clustering with linear running time in data size. Despite their effectiveness, existing algorithms still face scalability issues as they require scanning the entire dataset for iterative center swaps. This typically leads to an O(ndk) running time, where n is the data size, d is the dimension, k is the number of clusters. To further improve the efficiency of local search algorithms, we propose new methods based on adaptive sampling and bandit strategies. Specifically, adaptive sampling can well approximate the distance-based sampling distribution without maintaining pairwise distances between data points and the centers, enabling fast and accurate sampling in sublinear time after an $\tilde{O}(nd)$ time preprocessing step. The bandit strategy models the best swap pair selection as a bandit problem, where a grouping strategy is proposed for fast identification of the optimal swap pair. With these techniques, our proposed algorithm can achieve constant approximation in expected running time $\tilde{O}(nd + k^4)$ under mild assumptions on optimal clusters and swap pair distributions. Our approach also extends naturally to the k-median objective, achieving constant approximation in expected running time $\tilde{O}(nd + \sqrt{n}k^3)$ without distributional assumptions. Empirical results demonstrate that our algorithm achieves up to 1000× speedup over existing local search methods on datasets with 100 million points, while delivering comparable clustering quality. Compared to coreset-based approaches, it also provides up to 80× speedup and consistently yields better clustering results.

Applications · Neuroscience, Cognitive Science

Christian Internò, Robert Geirhos, Markus Olhofer, Sunny Liu, Barbara Hammer, David Klindt

The rapid advancement of generative AI enables highly realistic synthetic video, posing significant challenges for content authentication and raising urgent concerns about misuse. Existing detection methods often struggle with generalization and capturing subtle temporal inconsistencies. We propose $ReStraV$ ($Re$presentation $Stra$ightening for $V$ideo), a novel approach to distinguish natural from AI-generated videos. Inspired by the ``perceptual straightening'' hypothesis—which suggests real-world video trajectories become more straight in neural representation domain—we analyze deviations from this expected geometric property. Using a pre-trained self-supervised vision transformer (DINOv2), we quantify the temporal curvature and stepwise distance in the model's representation domain. We aggregate statistical and signals descriptors of these measures for each video and train a classifier. Our analysis shows that AI-generated videos exhibit significantly different curvature and distance patterns compared to real videos. A lightweight classifier achieves state-of-the-art detection performance (e.g., $97.17$ % accuracy and $98.63$ % AUROC on the VidProM benchmark, substantially outperforming existing image- and video-based methods. ReStraV is computationally efficient, it is offering a low-cost and effective detection solution. This work provides new insights into using neural representation geometry for AI-generated video detection.

Computer Vision · Image and Video Generation

Alexandros Graikos, Nebojsa Jojic, Dimitris Samaras

Large denoising diffusion models, such as Stable Diffusion, have been trained on billions of image-caption pairs to perform text-conditioned image generation. As a byproduct of this training, these models have acquired general knowledge about image statistics, which can be useful for other inference tasks. However, when confronted with sampling an image under new constraints, e.g. generating the missing parts of an image, using large pre-trained text-to-image diffusion models is inefficient and often unreliable. Previous approaches either utilized backpropagation through the denoiser network, making them significantly slower and more memory-demanding than simple text-to-image generation, or only enforced the constraint locally, failing to capture critical long-range correlations in the sampled image. In this work, we propose an algorithm that enables fast, high-quality generation under arbitrary constraints. We show that in denoising diffusion models, we can employ an approximation to Newton’s optimization method that allows us to speed up inference and avoid the expensive backpropagation operations. Our approach produces results that rival or surpass the state-of-the-art training-free inference methods while requiring a fraction of the time. We demonstrate the effectiveness of our algorithm under both linear (inpainting, super-resolution) and non-linear (style-guided generation) constraints. An implementation is provided at https://github.com/cvlab-stonybrook/fast-constrained-sampling.

General Machine Learning · Transfer, Multitask and Meta-learning

Anke Tang, Enneng Yang, Li Shen, Yong Luo, Han Hu, Lefei Zhang, Bo Du, Dacheng Tao

Deep model merging represents an emerging research direction that combines multiple fine-tuned models to harness their specialized capabilities across different tasks and domains. Current model merging techniques focus on merging all available models simultaneously, with weight interpolation-based methods being the predominant approach. However, these conventional approaches are not well-suited for scenarios where models become available sequentially, and they often suffer from high memory requirements and potential interference between tasks. In this study, we propose a training-free projection-based continual merging method that processes models sequentially through orthogonal projections of weight matrices and adaptive scaling mechanisms. Our method operates by projecting new parameter updates onto subspaces orthogonal to existing merged parameter updates while using an adaptive scaling mechanism to maintain stable parameter distances, enabling efficient sequential integration of task-specific knowledge. Our approach maintains constant memory complexity to the number of models, minimizes interference between tasks through orthogonal projections, and retains the performance of previously merged models through adaptive task vector scaling. Extensive experiments on CLIP-ViT models demonstrate that our method achieves a 5-8% average accuracy improvement while maintaining robust performance in different task orderings. Code is publicly available at https://github.com/tanganke/opcm .

Social Aspects · Fairness, Equity, Justice and Safety

Zixuan Hu, Li Shen, Zhenyi Wang, Yongxian Wei, Dacheng Tao

Harmful fine-tuning poses critical safety risks to fine-tuning-as-a-service for large language models. Existing defense strategies preemptively build robustness via attack simulation but suffer from fundamental limitations: (i) the infeasibility of extending attack simulations beyond bounded threat models due to the inherent difficulty of anticipating unknown attacks, and (ii) limited adaptability to varying attack settings, as simulation fails to capture their variability and complexity. To address these challenges, we propose Bayesian Data Scheduler (BDS), an adaptive tuning-stage defense strategy with no need for attack simulation. BDS formulates harmful fine-tuning defense as a Bayesian inference problem, learning the posterior distribution of each data point's safety attribute, conditioned on the fine-tuning and alignment datasets. The fine-tuning process is then constrained by weighting data with their safety attributes sampled from the posterior, thus mitigating the influence of harmful data. By leveraging the post hoc nature of Bayesian inference, the posterior is conditioned on the fine-tuning dataset, enabling BDS to tailor its defense to the specific dataset, thereby achieving adaptive defense. Furthermore, we introduce a neural scheduler based on amortized Bayesian learning, enabling efficient transfer to new data without retraining. Comprehensive results across diverse attack and defense settings demonstrate the state-of-the-art performance of our approach. Code is available at https://github.com/Egg-Hu/Bayesian-Data-Scheduler.

General Machine Learning · Transfer, Multitask and Meta-learning

Hainan Fang, Yuanbo Wen, Jun Bi, Yihan Wang, Tonghui He, Yanlin Tang, Di Huang, Jiaming Guo, Rui Zhang, Qi Guo 等

Compilers, while essential, are notoriously complex systems that demand prohibitively expensive human expertise to develop and maintain. The recent advancements in Large Language Models (LLMs) offer a compelling new paradigm: Neural Compilation, which could potentially simplify compiler development for new architectures and facilitate the discovery of innovative optimization techniques. However, several critical obstacles impede its practical adoption. Firstly, a significant lack of dedicated benchmarks and robust evaluation methodologies hinders objective assessment and tracking of progress in the field. Secondly, systematically enhancing the reliability and performance of LLM-generated assembly remains a critical challenge. Addressing these challenges, this paper introduces NeuComBack, a novel benchmark dataset specifically designed for IR-to-assembly compilation. Leveraging this dataset, we first define a foundational Neural Compilation workflow and conduct a comprehensive evaluation of the capabilities of recent frontier LLMs on Neural Compilation, establishing new performance baselines. We further propose a self-evolving prompt optimization method that enables LLMs to iteratively evolve their internal prompt strategies by extracting insights from prior self-debugging traces, thereby enhancing their neural compilation capabilities. Experiments demonstrate that our method significantly improves both the functional correctness and the performance of LLM-generated assembly code. Compared to baseline prompts, the functional correctness rates improved from 44% to 64% on x86_64 and from 36% to 58% on aarch64, respectively. More significantly, among the 16 correctly generated x86_64 programs using our method, 14 (87.5%) surpassed clang-O3 performance. These consistent improvements across diverse architectures (x86_64 and aarch64) and program distributions (NeuComBack L1 and L2) validate our method's superiority over conventional approaches and its potential for broader adoption in low-level neural compilation.

Computer Vision · Segmentation

Yijie Zheng, Weijie Wu, Qingyun Li, Xuehui Wang, Xu Zhou, Aiai Ren, Jun Shen, Long Zhao, Guoqing Li, Xue Yang

Language-guided object recognition in remote sensing imagery is crucial for large-scale mapping and automated data annotation. However, existing open-vocabulary and visual grounding methods rely on explicit category cues, limiting their ability to handle complex or implicit queries that require advanced reasoning. To address this issue, we introduce a new suite of tasks, including Instruction-Oriented Object Counting, Detection, and Segmentation (InstructCDS), covering open-vocabulary, open-ended, and open-subclass scenarios. We further present EarthInstruct, the first InstructCDS benchmark for earth observation. It is constructed from two diverse remote sensing datasets with varying spatial resolutions and annotation rules across 20 categories, necessitating models to interpret dataset-specific instructions. Given the scarcity of semantically rich labeled data in remote sensing, we propose InstructSAM, a training-free framework for instruction-driven object recognition. InstructSAM leverages large vision-language models to interpret user instructions and estimate object counts, employs SAM2 for mask proposal, and formulates mask-label assignment as a binary integer programming problem. By integrating semantic similarity with counting constraints, InstructSAM efficiently assigns categories to predicted masks without relying on confidence thresholds. Experiments demonstrate that InstructSAM matches or surpasses specialized baselines across multiple tasks while maintaining near-constant inference time regardless of object count, reducing output tokens by 89\% and overall runtime by over 32\% compared to direct generation approaches. We believe the contributions of the proposed tasks, benchmark, and effective approach will advance future research in developing versatile object recognition systems. The code is available at https://VoyagerXvoyagerx.github.io/InstructSAM.

Deep Learning · Generative Models and Autoencoders

Neta Shaul, Uriel Singer, Itai Gat, Yaron Lipman

Diffusion and flow matching models have significantly advanced media generation, yet their design space is well-explored, somewhat limiting further improvements. Concurrently, autoregressive (AR) models, particularly those generating continuous tokens, have emerged as a promising direction for unifying text and media generation, showing improved performance at scale. This paper introduces Transition Matching (TM), a novel discrete-time, continuous-state generative paradigm that unifies and advances both diffusion/flow models and continuous AR generation. TM decomposes complex generation tasks into simpler Markov transitions, allowing for expressive non-deterministic probability transition kernels and arbitrary non-continuous supervision processes, thereby unlocking new flexible design avenues. We explore these choices through three TM variants: (i) Difference Transition Matching (DTM), which generalizes flow matching to discrete-time by directly learning transition probabilities, yielding state-of-the-art image quality and text adherence. (ii) Autoregressive Transition Matching (ARTM) and (iii) Full History Transition Matching (FHTM) are partially and fully causal models, respectively, that generalize continuous AR methods. They achieve continuous causal AR generation quality comparable to non-causal approaches and potentially enable seamless integration with existing AR text generation techniques. Notably, FHTM is the first fully causal model to match or surpass the performance of flow-based methods on text-to-image task in continuous domains. We demonstrate these contributions through a rigorous large-scale comparison of TM variants and relevant baselines, maintaining a fixed architecture, training data, and hyperparameters.

Applications · Chemistry and Drug Discovery

Peizhi Niu, Yu-Hsiang Wang, Vishal Rana, Chetan Rupakheti, Abhishek Pandey, Olgica Milenkovic

We introduce a new graph diffusion model for small drug molecule generation which simultaneously offers a 10-fold reduction in the number of diffusion steps when compared to existing methods, preservation of small molecule graph motifs via motif compression, and an average 3\% improvement in SMILES validity over the DiGress model across all real-world molecule benchmarking datasets. Furthermore, our approach outperforms the state-of-the-art DeFoG method with respect to motif-conservation by roughly 4\%, as evidenced by high ChEMBL-likeness, QED and newly introduced shingles distance scores. The key ideas behind the approach are to use a combination of deterministic and random subgraph perturbations, so that the node and edge noise schedules are codependent; to modify the loss function of the training process in order to exploit the deterministic component of the schedule; and, to ''compress'' a collection of highly relevant carbon ring and other motif structures into supernodes in a way that allows for simple subsequent integration into the molecular scaffold.

Deep Learning · Everything Else

Ruiyang Zhou, Shuozhe Li, Amy Zhang, Liu Leqi

Recent advances in large language models have been driven by reinforcement learning (RL)-style post-training, which improves reasoning by optimizing model outputs based on reward or preference signals. GRPO-style approaches implement this by using self-generated samples labeled by an outcome-based verifier. However, these methods depend heavily on the model’s initial ability to produce positive samples. They primarily refine what the model already knows (distribution sharpening) rather than enabling the model to solve problems where it initially fails. This limitation is especially problematic in early-stage RL training and on challenging reasoning tasks, where positive samples are unlikely to be generated. To unlock reasoning ability in such settings, the model must explore new reasoning trajectories beyond its current output distribution. Such exploration requires access to sufficiently good positive samples to guide the learning. While expert demonstrations seem like a natural solution, we find that they are often ineffective in RL post-training. Instead, we identify two key properties of effective positive samples: they should (1) be likely under the current policy, and (2) increase the model’s likelihood of predicting the correct answer. Based on these insights, we propose \textbf{Self-Explanation Policy Optimization (ExPO)}—a simple and modular framework that generates such samples by conditioning on the ground-truth answer. ExPO enables efficient exploration and guides the model to produce reasoning trajectories more aligned with its policy than expert-written CoTs, while ensuring higher quality than its own (incorrect) samples. Experiments show that ExPO improves both learning efficiency and final performance on reasoning benchmarks, surpassing expert-demonstration-based methods in challenging settings such as MATH level-5, where the model initially struggles the most.

Deep Learning · Everything Else

Frank Shih, Zhenghao Jiang, Faming Liang

Uncertainty quantification (UQ) in scientific machine learning is increasingly critical as neural networks are widely adopted to tackle complex problems across diverse scientific disciplines. For physics-informed neural networks (PINNs), a prominent model in scientific machine learning, uncertainty is typically quantified using Bayesian or dropout methods. However, both approaches suffer from a fundamental limitation: the prior distribution or dropout rate required to construct honest confidence sets cannot be determined without additional information. In this paper, we propose a novel method within the framework of extended fiducial inference (EFI) to provide rigorous uncertainty quantification for PINNs. The proposed method leverages a narrow-neck hyper-network to learn the parameters of the PINN and quantify their uncertainty based on imputed random errors in the observations. This approach overcomes the limitations of Bayesian and dropout methods, enabling the construction of honest confidence sets based solely on observed data. This advancement represents a significant breakthrough for PINNs, greatly enhancing their reliability, interpretability, and applicability to real-world scientific and engineering challenges. Moreover, it establishes a new theoretical framework for EFI, extending its application to large-scale models, eliminating the need for sparse hyper-networks, and significantly improving the automaticity and robustness of statistical inference.

Social Aspects · Privacy-preserving Statistics and Machine Learning

Haoyuan Liang, Shilei Cao, Li, Zhiyu Ye, Haohuan Fu, Juepeng Zheng

Federated learning (FL) has recently emerged as the primary approach to overcoming data silos, enabling collaborative model training without sharing sensitive or proprietary data. Parallel federated learning (PFL) aggregates models trained independently on each client’s local data, which can lead to suboptimal convergence due to limited data exposure. In contrast, Sequential Federated Learning (SFL) allows models to traverse client datasets sequentially, enhancing data utilization. However, SFL effectiveness is limited in real-world non-IID scenarios characterized by category shift (inconsistent class distributions) and domain shift (distribution discrepancies). These shifts cause two critical issues: update order sensitivity, where model performance varies significantly with the sequence of client updates, and catastrophic forgetting, where the model forgets previously learned features when trained on new client data. We propose SPFL, a novel updating method that can be integrated into existing FL methods, integrating sequential updates with parallel aggregation to enhance data utilization and ease update order sensitivity. At the same time, we give the convergence analysis of SPFL under strong convex, general convex, and non-convex conditions, proving that this update scheme is significantly better than PFL and SFL. Additionally, we introduce the Global-Local Alignment Module to mitigate catastrophic forgetting by aligning the predictions of the global model with those of the local and previous models during training. Our extensive experiments demonstrate that integrating SPFL into existing PFL methods significantly improves performance under category and domain shifts.

Applications · Climate

Mohamad Hakam Shams Eddin, Yikui Zhang, Stefan Kollet, Jürgen Gall

Recent deep learning approaches for river discharge forecasting have improved the accuracy and efficiency in flood forecasting, enabling more reliable early warning systems for risk management. Nevertheless, existing deep learning approaches in hydrology remain largely confined to local-scale applications and do not leverage the inherent spatial connections of bodies of water. Thus, there is a strong need for new deep learning methodologies that are capable of modeling spatio-temporal relations to improve river discharge and flood forecasting for scientific and operational applications. To address this, we present RiverMamba, a novel deep learning model that is pretrained with long-term reanalysis data and that can forecast global river discharge and floods on a $0.05^\circ$ grid up to 7 days lead time, which is of high relevance in early warning. To achieve this, RiverMamba leverages efficient Mamba blocks that enable the model to capture spatio-temporal relations in very large river networks and enhance its forecast capability for longer lead times. The forecast blocks integrate ECMWF HRES meteorological forecasts, while accounting for their inaccuracies through spatio-temporal modeling. Our analysis demonstrates that RiverMamba provides reliable predictions of river discharge across various flood return periods, including extreme floods, and lead times, surpassing both AI- and physics-based models. The source code and datasets are publicly available at the project page https://hakamshams.github.io/RiverMamba.

Applications · Robotics

Raktim Goswami, Prashanth Krishnamurthy, Yann LeCun, Farshad Khorrami

Visual imitation learning enables robotic agents to acquire skills by observing expert demonstration videos. In the one-shot setting, the agent generates a policy after observing a single expert demonstration without additional fine-tuning. Existing approaches typically train and evaluate on the same set of tasks, varying only object configurations, and struggle to generalize to unseen tasks with different semantic or structural requirements. While some recent methods attempt to address this, they exhibit low success rates on hard test tasks that, despite being visually similar to some training tasks, differ in context and require distinct responses. Additionally, most existing methods lack an explicit model of environment dynamics, limiting their ability to reason about future states. To address these limitations, we propose a novel framework for one-shot visual imitation learning via world-model-guided trajectory generation. Given an expert demonstration video and the agent’s initial observation, our method leverages a learned world model to predict a sequence of latent states and actions. This latent trajectory is then decoded into physical waypoints that guide the agent’s execution. Our method is evaluated on two simulated benchmarks and three real-world robotic platforms, where it consistently outperforms prior approaches, with over 30% improvement in some cases.

Computer Vision · Classification and Understanding

Jin Yao, Radowan Mahmud Redoy, Sebastian Elbaum, Matthew Dwyer, Zezhou Cheng

Detecting objects in 3D space from monocular input is crucial for applications ranging from robotics to scene understanding. Despite advanced performance in the indoor and autonomous driving domains, existing monocular 3D detection models struggle with in-the-wild images due to the lack of 3D in-the-wild datasets and the challenges of 3D annotation. We introduce LabelAny3D, an analysis-by-synthesis framework that reconstructs holistic 3D scenes from 2D images to efficiently produce high-quality 3D bounding box annotations. Built on this pipeline, we present COCO3D, a new benchmark for open-vocabulary monocular 3D detection, derived from the MS-COCO dataset and covering a wide range of object categories absent from existing 3D datasets. Experiments show that annotations generated by LabelAny3D improve monocular 3D detection performance across multiple benchmarks, outperforming prior auto-labeling approaches in quality. These results demonstrate the promise of foundation-model-driven annotation for scaling up 3D recognition in realistic, open-world settings.

Theory · Reinforcement Learning and Planning

Sergio Calo, Anders Jonsson, Gergely Neu, Ludovic Schwartz, Javier Segovia-Aguas

Bisimulation metrics are powerful tools for measuring similarities between stochastic processes, and specifically Markov chains. Recent advances have uncovered that bisimulation metrics are, in fact, optimal-transport distances, which has enabled the development of fast algorithms for computing such metrics with provable accuracy and runtime guarantees. However, these recent methods, as well as all previously known methods, assume full knowledge of the transition dynamics. This is often an impractical assumption in most real-world scenarios, where typically only sample trajectories are available. In this work, we propose a stochastic optimization method that addresses this limitation and estimates bisimulation metrics based on sample access, without requiring explicit transition models. Our approach is derived from a new linear programming (LP) formulation of bisimulation metrics, which we solve using a stochastic primal-dual optimization method. We provide theoretical guarantees on the sample complexity of the algorithm and validate its effectiveness through a series of empirical evaluations.

Applications · Robotics

MingMing Yu, Fei Zhu, Wenzhuo Liu, Yirong Yang, Qunbo Wang, wenjun wu, Jing Liu

Embodied agents are expected to perform object navigation in dynamic, open-world environments. However, existing approaches typically rely on static trajectories and a fixed set of object categories during training, overlooking the real-world requirement for continual adaptation to evolving scenarios. To facilitate related studies, we introduce the continual object navigation benchmark, which requires agents to acquire navigation skills for new object categories while avoiding catastrophic forgetting of previously learned knowledge. To tackle this challenge, we propose C-Nav, a continual visual navigation framework that integrates two key innovations: (1) A dual-path anti-forgetting mechanism, which comprises feature distillation that aligns multi-modal inputs into a consistent representation space to ensure representation consistency, and feature replay that retains temporal features within the action decoder to ensure policy consistency. (2) An adaptive sampling strategy that selects diverse and informative experiences, thereby reducing redundancy and minimizing memory overhead. Extensive experiments across multiple model architectures demonstrate that C-Nav consistently outperforms existing approaches, achieving superior performance even compared to baselines with full trajectory retention, while significantly lowering memory requirements. The code will be publicly available at \url{https://bigtree765.github.io/C-Nav-project}.