Reinforcement learning (RL) has demonstrated considerable potential for enhancing reasoning in large language models (LLMs). However, existing methods suffer from Gradient Starvation and Policy Degradation when training directly on samples with mixed difficulty. To mitigate this, prior approaches leverage Chain-of-Thought (CoT) data, but the construction of high-quality CoT annotations remains labor-intensive. Alternatively, curriculum learning strategies have been explored but frequently encounter challenges, such as difficulty mismatch, reliance on manual curriculum design, and catastrophic forgetting. To address these issues, we propose AdaCuRL, a Adaptive Curriculum Reinforcement Learning framework that integrates coarse-to-fine difficulty estimation with adaptive curriculum scheduling. This approach dynamically aligns data difficulty with model capability and incorporates a data revisitation mechanism to mitigate catastrophic forgetting. Furthermore, AdaCuRL employs adaptive reference and sparse KL strategies to prevent Policy Degradation. Extensive experiments across diverse reasoning benchmarks demonstrate that AdaCuRL consistently achieves significant performance improvements on both LLMs and MLLMs.

Given a table T in a database and a question Q in natural language, the table question answering (TQA) task aims to return an accurate answer to Q based on the content of T. Recent state-of-the-art solutions leverage large language models (LLMs) to obtain high-quality answers. However, most rely on proprietary, large-scale LLMs with costly API access, posing a significant financial barrier. This paper instead focuses on TQA with smaller, open-weight LLMs that can run on a desktop or laptop. This setting is challenging, as such LLMs typically have weaker capabilities than large proprietary models, leading to substantial performance degradation with existing methods. We observe that a key reason for this degradation is that prior approaches often require the LLM to solve a highly sophisticated task using long, complex prompts, which exceed the capabilities of small open-weight LLMs. Motivated by this observation, we present Orchestra, a multi-agent approach that unlocks the potential of accessible LLMs for high-quality, cost-effective TQA. Orchestra coordinates a group of LLM agents, each responsible for a relatively simple task, through a structured, layered workflow to solve complex TQA problems -- akin to an orchestra. By reducing the prompt complexity faced by each agent, Orchestra significantly improves output reliability. We implement Orchestra on top of AgentScope, an open-source multi-agent framework, and evaluate it on multiple TQA benchmarks using a wide range of open-weight LLMs. Experimental results show that Orchestra achieves strong performance even with small- to medium-sized models. For example, with Qwen2.5-14B, Orchestra reaches 72.1% accuracy on WikiTQ, approaching the best prior result of 75.3% achieved with GPT-4; with larger Qwen, Llama, or DeepSeek models, Orchestra outperforms all prior methods and establishes new state-of-the-art results across all benchmarks.

Image editing in our daily lives often requires models to first understand user’s intention and then proceed with the editing. Despite significant advancements in image editing technology, understanding and executing complex instructions remains a substantial challenge. Existing image editing models either fail to comprehend complex intentions or make errors when dealing with multiple objects. To address these challenges, we present an innovative image editing framework that employs Chain-of-Thought (CoT) reasoning and localizing capabilities of multimodal Large Language Models (LLMs) to assist diffusion models in generating more refined images. We meticulously design a CoT process comprising instruction decomposition, region localization, and detailed description. We train our model to learn the CoT process and the mask of the edited image. By providing diffusion models with generated prompts and generated masks, our model edits images with a superior understanding of instructions. Extensive experiments demonstrate that our model outperforms existing state-of-the-art models in image generation both qualitatively and quantitatively. Notably, our model exhibits an enhanced ability of understanding complex prompts and generating corresponding images.

With the fast-paced development of multimodal large language models (MLLMs), we can now converse with AI systems in natural languages to understand images. However, the reasoning power and world knowledge embedded in the large language models have been much less investigated and exploited for image perception tasks. In this paper, we propose Lenna, a Language enhanced reasoning detection assistant, which utilizes the robust multimodal feature representation of MLLMs, while preserving location information for detection. This is achieved by incorporating an additionaltoken in the MLLM vocabulary that is free of explicit semantic context but serves as a prompt for the detector to identify the corresponding position. To evaluate the reasoning capability of Lenna, we construct a ReasonDet dataset to measure its performance on reasoning-based detection. Remarkably, Lenna demonstrates outstanding performance on ReasonDet and comes with significantly low training costs. It also incurs minimal transferring overhead when extended to other tasks.

Zero-shot domain adaptation (ZSDA) presents substantial challenges due to the lack of images in the target domain. Previous approaches leverage Vision-Language Models (VLMs) to tackle this challenge, exploiting their zero-shot learning capabilities. However, these methods primarily address domain distribution shifts and overlook the misalignment between the detection task and VLMs, which rely on manually crafted prompts. To overcome these limitations, we propose the unified prompt and representation enhancement (UPRE) framework, which jointly optimizes both textual prompts and visual representations. Specifically, our approach introduces a multi-view domain prompt that combines linguistic domain priors with detection-specific knowledge, and a visual representation enhancement module that produces domain style variations. Furthermore, we introduce multi-level enhancement strategies, including relative domain distance and positive-negative separation, which align multi-modal representations at the image level and capture diverse visual representations at the instance level, respectively. Extensive experiments conducted on nine benchmark datasets demonstrate the superior performance of our framework in ZSDA detection scenarios. Code is available at https://github.com/AMAP-ML/UPRE.

Analysis of distributed data typically requires the collaboration of the data owners, as well as privacy protection. This paper focuses on the scenario where the database is vertically partitioned onto the data owners (referred to as vertical federated learning or VFL), e.g., an e-commerce platform and an online payment service collaborate to build a model to predict user behavior. To avoid revealing their private data during model fitting, the data owners commonly participate in a cryptographic protocol such as secure multiparty computation. However, the resulting model may still leak sensitive information under sophisticated data extraction attacks. A rigorous solution to this issue is to compute the model with differential privacy (DP), which provides strong and well-accepted privacy guarantees. Enforcing DP on VFL turns out to be highly challenging, and there does not yet exist an effective solution that avoids reliance on any trusted party. Consequently, practitioners are left with rather basic approaches for ensuring DP, e.g., each data owner perturbs her local data with additive noises, leading to suboptimal model utility. Can we achieve privacy-utility trade-offs for VFL with DP comparable to the centralized setting, without trusting any party? In this paper, we make a significant step towards providing a positive answer to this question. We focus on a subset of the data analysis and machine learning tasks-the class of tasks where the sensitive information to release can be expressed as a polynomial function of the input. Following the distributed DP framework that does not require any trusted party, we propose a generic mechanism to solve this class of problems, called the Skellam Quantization Mechanism (SQM). We formally prove the privacy guarantee of our solution, and show that it is able to match the privacy-utility trade-offs in the centralized setting. We then instantiate SQM on two classical tasks, principal component analysis and logistic regression. Extensive experiments on real-world datasets confirm the strong performance of SQM.

Local differential privacy ( LDP ) is a strong privacy standard that has been adopted by popular software systems, including Chrome, iOS, MacOS, and Windows. The main idea is that each individual perturbs their own data locally, and only submits the resulting noisy version to a data aggregator. Although much effort has been devoted to computing various types of aggregates and building machine learning applications under LDP, research on fundamental perturbation mechanisms has not achieved significant improvement in recent years. Towards a more refined result utility, existing works in the literature mainly focus on improving the worst-case guarantee. However, this approach does not necessarily promise a better average performance given the fact that the data in practice obey a certain distribution, which is not known beforehand. In this paper, we propose the advanced adaptive additive ( AAA ) mechanism, which is a distribution-aware approach that addresses the average utility and tackles the classical mean estimation problem. AAA is carried out in a two-step approach: first, as the global data distribution is not available beforehand, the data aggregator selects a random subset of individuals to compute a (noisy) quantized data descriptor; then, in the second step, the data aggregator collects data from the remaining individuals, which are perturbed in a distribution-aware fashion. The perturbation involved in the latter step is obtained by solving an optimization problem, which is formulated with the data descriptor obtained in the former step and the desired properties of task-determined utilities. We provide rigorous privacy proofs and utility analyses, as well as extensive experiments comparing AAA with state-of-the-art mechanisms. The evaluation results demonstrate that the AAA mechanism consistently outperforms existing solutions with a clear margin in terms of result utility, on a wide range of privacy constraints and real-world and synthetic datasets.

In recent years, the field of audio deepfake detection has witnessed significant advancements. Nonetheless, the majority of solutions have concentrated on high-quality audio, largely overlooking the challenge of low-quality compressed audio in real-world scenarios. Low-quality compressed audio typically suffers from a loss of high-frequency details and time-domain information, which significantly undermines the performance of advanced deepfake detection systems when confronted with such data. In this paper, we introduce a deepfake detection model that employs knowledge distillation across the frequency and time domains. Our approach aims to train a teacher model with high-quality data and a student model with low-quality compressed data. Subsequently, we implement frequency-domain and time-domain distillation to facilitate the student model's learning of high-frequency information and time-domain details from the teacher model. Experimental evaluations on the ASVspoof 2019 LA and ASVspoof 2021 DF datasets illustrate the effectiveness of our methodology. On the ASVspoof 2021 DF dataset, which consists of low-quality compressed audio, we achieved an Equal Error Rate (EER) of 2.82%. To our knowledge, this performance is the best among all deepfake voice detection systems tested on the ASVspoof 2021 DF dataset. Additionally, our method proves to be versatile, showing notable performance on high-quality data with an EER of 0.30% on the ASVspoof 2019 LA dataset, closely approaching state-of-the-art results.

Diffusion models have recently gained significant attention in both academia and industry due to their impressive generative performance in terms of both sampling quality and distribution coverage. Accordingly, proposals are made for sharing pre-trained diffusion models across different organizations, as a way of improving data utilization while enhancing privacy protection by avoiding sharing private data directly. However, the potential risks associated with such an approach have not been comprehensively examined. In this paper, we take an adversarial perspective to investigate the potential privacy and fairness risks associated with the sharing of diffusion models. Specifically, we investigate the circumstances in which one party (the sharer) trains a diffusion model using private data and provides another party (the receiver) black-box access to the pre-trained model for downstream tasks. We demonstrate that the sharer can execute fairness poisoning attacks to undermine the receiver’s downstream models by manipulating the training data distribution of the diffusion model. Meanwhile, the receiver can perform property inference attacks to reveal the distribution of sensitive features in the sharer’s dataset. Our experiments conducted on real-world datasets demonstrate remarkable attack performance on different types of diffusion models, which highlights the critical importance of robust data auditing and privacy protection protocols in pertinent applications.

Aremarkable number of backdoor attack methods have been proposed in the literature on deep neural networks (DNNs). However, it hasn't been sufficiently addressed in the existing methods of achieving true senseless backdoor attacks that are visually invisible and label-consistent. In this paper, we propose a new backdoor attack method where the labels of the backdoor images are perfectly aligned with their content, ensuring label consistency. Additionally, the backdoor trigger is meticulously designed, allowing the attack to evade DNN model checks and human inspection. Our approach employs an auto-encoder (AE) to conduct representation learning of benign images and interferes with salient classification features to increase the dependence of backdoor image classification on backdoor triggers. To ensure visual invisibility, we implement a method inspired by image steganography that embeds trigger patterns into the image using the DNN and enable sample-specific backdoor triggers. We conduct comprehensive experiments on multiple benchmark datasets and network architectures to verify the effectiveness of our proposed method under the metric of attack success rate and invisibility. The results also demonstrate satisfactory performance against a variety of defense methods.

Previous fully-supervised defocus deblurring has made significant progress. However, training such deep models requires abundant paired ground truth, which is expensive and error-prone. This paper makes an attempt to train a defocus deblurring model without using paired ground truth and any other unpaired data. Related reblur-to-deblur schemes generally use physics-based reblur or GAN-based reblur, suffering from the robustness of blur kernel and hallucination generated by GAN. Besides, the domain gap between the realistic blurred image and reblurred image hinders deblurring performance. Addressing these challenges, we propose a weakly-supervised defocus deblurring framework via defocus detection attack. On one hand, we build a focused area detection attack (FADA) to enforce the focused area to reblur, thereby reversing its detection result by a pretrained defocus blur detection network. Moreover, we introduce a blur-aware transfer modulated from the defocused region to help FADA render a robust reblurred region. On the other hand, we implement a defocused region detection attack to guide the realistic blurred region to deblur in the process of training deblurring network with simulated-paired areas. Extensive experiments on three widely-used datasets verify the effectiveness of our framework.

Benefiting from deep learning, defocus blur detection (DBD) has made prominent progress. Existing DBD methods generally study multiscale and multilevel features to improve performance. In this article, from a different perspective, we explore to generate confrontational images to attack DBD network. Based on the observation that defocus area and focus region in an image can provide mutual feature reference to help improve the quality of the confrontational image, we propose a novel mutual-referenced attack framework. Firstly, we design a divide-and-conquer perturbation image generation model, where the focus region attack image and defocus area attack image are generated respectively. Then, we integrate mutual-referenced feature transfer (MRFT) models to improve attack performance. Comprehensive experiments are provided to verify the effectiveness of our method. Moreover, related applications of our study are presented, e.g., sample augmentation to improve DBD and paired sample generation to boost defocus deblurring.

Existing defocus blur detection (DBD) methods generally perform well on a single type of unfocused blur scene (e.g., foreground focus), thereby suffering from the performance degradation for the other types of unfocused blur scenes. In this paper, we present the first exploration on full-scene DBD, and propose a separate-and-combine framework to achieve excellent performance for diverse defocus blur scenes. We firstly structure full-scene DBD dataset (named as DeFBD+) through collecting more types of unfocused blur scenes (e.g., background focus, full focus and full out of focus) with pixel-level annotations. Then, to avoid performance degradation caused by mutual interference from local feature representation and global content perception, we implement a pixel-level DBD network and an image-level DBD classification network to learn these two abilities separately. After that, we propose an isomeric distillation mechanism to combine these two abilities. Extensive experiments show that the proposed approach achieves superior performance compared with state-of-the-art methods.

Most differential privacy mechanisms are applied (i.e., composed) numerous times on sensitive data. We study the design of optimal differential privacy mechanisms in the limit of a large number of compositions. As a consequence of the law of large numbers, in this regime the best privacy mechanism is the one that minimizes the Kullback-Leibler divergence between the conditional output distributions of the mechanism given two different inputs. We formulate an optimization problem to minimize this divergence subject to a cost constraint on the noise. We first prove that additive mechanisms are optimal. Since the optimization problem is infinite dimensional, it cannot be solved directly; nevertheless, we quantize the problem to derive nearoptimal additive mechanisms that we call "cactus mechanisms" due to their shape. We show that our quantization approach can be arbitrarily close to an optimal mechanism. Surprisingly, for quadratic cost, the Gaussian mechanism is strictly suboptimal compared to this cactus mechanism. Finally, we provide numerical results which indicate that cactus mechanisms outperform Gaussian and Laplace mechanisms for a finite number of compositions.The full proofs can be found in the extended version at [1]. This paper is Part I in a pair of papers, where Part II is [2].

The second-order converse bound of multiple access channels is an intriguing problem in information theory. In this work, in the setting of the two-user discrete memoryless multiple access channel (DM-MAC) under the maximal error probability criterion, we investigate the gap between the best achievable rates and the asymptotic capacity region. With “wringing techniques” and meta-converse arguments, we show that gap at blocklength$n$is upper bounded by$O(1/\sqrt{n})$.

Deep neural networks (DNNs) have seen extensive studies on image recognition and classification, image segmentation, and related topics. However, recent studies show that DNNs are vulnerable in defending adversarial examples. The classification network can be deceived by adding a small amount of perturbation to clean samples. There are challenges when researchers want to design a general approach to defend against a wide variety of adversarial examples. To solve this problem, we introduce a defensive method to prevent adversarial examples from generating. Instead of designing a stronger classifier, we built a more robust classification system that can be viewed as a structural black box. After adding a buffer to the classification system, attackers can be efficiently deceived. The real evaluation results of the generated adversarial examples are often contrary to what the attacker thinks. Additionally, we do not assume a specific attack method premise. This incognizance to underlying attacks demonstrates the generalizability of the buffer to potential adversarial attacks. Extensive experiments indicate that the defense method greatly improves the security performance of DNNs.

The edge removal problem studies the loss in network coding rates that results when a network communication edge is removed from a given network. It is known, for example, that in networks restricted to linear coding schemes and networks restricted to Abelian group codes, removing an edge e* with capacity Re. reduces the achievable rate on each source by no more than Re*. In this work, we seek to uncover larger families of encoding functions for which the edge removal statement holds. We take a local perspective: instead of requiring that all network encoding functions satisfy certain restrictions (e.g., linearity), we limit only the function carried on the removed edge e*. Our central results give sufficient conditions on the function carried by edge e* in the code used to achieve a particular rate vector under which we can demonstrate the achievability of a related rate vector once e* is removed.

For a large-scale global optimization (LSGO) problem, divide-and-conquer is usually considered an effective strategy to decompose the problem into smaller subproblems, each of which can then be solved individually. Among these decomposition methods, variable grouping is shown to be promising in recent years. Existing variable grouping methods usually assume the problem to be black-box (i.e., assuming that an analytical model of the objective function is unknown), and they attempt to learn appropriate variable grouping that would allow for a better decomposition of the problem. In such cases, these variable grouping methods do not make a direct use of the formula of the objective function. However, it can be argued that many real-world problems are white-box problems, that is, the formulas of objective functions are often known a priori. These formulas of the objective functions provide rich information which can then be used to design an effective variable group method. In this article, a formula-based grouping strategy (FBG) for white-box problems is first proposed. It groups variables directly via the formula of an objective function which usually consists of a finite number of operations (i.e., four arithmetic operations “[Formula: see text]”, “[Formula: see text]”, “[Formula: see text]”, “[Formula: see text]” and composite operations of basic elementary functions). In FBG, the operations are classified into two classes: one resulting in nonseparable variables, and the other resulting in separable variables. In FBG, variables can be automatically grouped into a suitable number of non-interacting subcomponents, with variables in each subcomponent being interdependent. FBG can easily be applied to any white-box problem and can be integrated into a cooperative coevolution framework. Based on FBG, a novel cooperative coevolution algorithm with formula-based variable grouping (so-called CCF) is proposed in this article for decomposing a large-scale white-box problem into several smaller subproblems and optimizing them respectively. To further enhance the efficiency of CCF, a new local search scheme is designed to improve the solution quality. To verify the efficiency of CCF, experiments are conducted on the standard LSGO benchmark suites of CEC'2008, CEC'2010, CEC'2013, and a real-world problem. Our results suggest that the performance of CCF is very competitive when compared with those of the state-of-the-art LSGO algorithms.

Evolutionary algorithms (EAs) are a kind of efficient and effective algorithms for global optimization problems. However, their efficiency and effectiveness will be greatly reduced for large scale problems. To handle this issue, a variable grouping strategy is first designed, in which the variables with the interaction each other are classified into one group, while the variables without interaction are classified into different groups. Then, evolution can be conducted in these groups separately. In this way, a large scale problem can be decomposed into several small scale problems and this makes the problem solving much easier. Furthermore, an auxiliary function, which can help algorithm to escape from the current local optimal solution and find a better one, is designed and integrated into EA. Based on these, a variable grouping based differential evolution algorithm (briefly, VGDE) using auxiliary function is proposed. At last, the simulations are made on the standard benchmark suite in CEC'2013, and VGDE is compared with several well performed algorithms. The results indicate the proposed algorithm VGDE is more efficient and effective.

For large scale global optimization problems, the efficiency and effectiveness of evolutionary algorithms (EAs) will be much reduced with the dimension increasing. In this paper, a novel evolutionary algorithm is proposed in order to improve the performance of EAs. In the proposed algorithm, on one hand, a variable grouping strategy is introduced. It can group all variables into several subcomponents, while the variables in each subcomponent are non-separable. In this way, a large scale problem can be decomposed into several small scale problems. On the other hand, a filled function with one parameter is integrated into EAs, which can help algorithm to escape from the current local optimal solution and find a better one. The simulations are made on the standard benchmark suite in CEC'2013, and the proposed algorithm is compared with several well performed algorithms. The results indicate the proposed algorithm is more efficient and effective.

In this paper, a novel evolutionary algorithm framework called smoothing and auxiliary functions based cooperative coevolution (Briefly, SACC) for large scale global optimization problems is proposed. In this new algorithm pattern, a smoothing function and an auxiliary function are well integrated with a cooperative coevolution algorithm. In this way, the performance of the cooperative coevolution algorithm may be improved. In SACC, the cooperative coevolution is responsible for the parallel searching in multiple areas simultaneously. Afterwards, an existing smoothing function is used to eliminate all the local optimal solutions no better than the best one obtained until now. Unfortunately, as the above takes place, the smoothing function will lose descent directions, which will weaken the local search. However, a proposed auxiliary function can overcome the drawback, which helps to find a better local optimal solution. A clever strategy on BFGS quasi-Newton method is designed to make the local search more efficient. The simulations on standard benchmark suite in CEC'2013 are made, and the results indicate the proposed algorithm SACC is effective and efficient.

There are many algorithms to solve large sparse linear systems in parallel; however, most of them acquire synchronization and thus are lack of scalability. In this paper, we propose a new distributed numerical algorithm, called Directed Transmission Method (DTM). DTM is a fully asynchronous, scalable and continuous-time iterative algorithm to solve the arbitrarily-large sparse linear system whose coefficient matrix is symmetric-positive-definite (SPD). DTM is able to be freely running on the heterogeneous parallel computer with arbitrary number of processors, which might be manycore microprocessors, clusters, grids, clouds, and the Internet. We proved that DTM is convergent by making use of the final value theorem of Laplacian Transformation. Numerical experiments show that DTM is efficient.