Zifan Zhang

Md Sharif Hossen, Zifan Zhang, Dara Ron et al.

The open radio access network (O-RAN) RAN intelligent controller (RIC) hosts data-driven xApps and rApps to optimize network performance. However, two challenges hinder ML-driven xApp/rApp development: (i) key performance metric (KPM) data scarcity caused by interface latency, and (ii) network disruption risks when testing and validating AI models directly on live networks. We develop OpenTwin, a digital twin framework built on an open-source O-RAN simulator (ns-O-RAN-flexRIC) and KPM streaming via the O1 interface, deployed within the non-RT RIC. OpenTwin uses a two-step ML approach: an XGBoost model that learns time-varying network behavior to generate simulator configuration parameters, followed by a time-aware recursive least squares (RLS) tuner that continuously corrects KPM deviations between the twin and real-world measurements. A deviation-aware scoring mechanism monitors twin fidelity and automatically triggers resynchronization upon detecting network drift. We demonstrate OpenTwin with an energy-saving xApp that validates control policies in the virtual space before applying reconfigurations to the physical network. Experimental results show that OpenTwin mirrors real-world KPMs with up to 96% accuracy and enables the xApp to significantly reduce energy consumption without disrupting live operations.

Zhao Tian, Zifan Zhang, Tao Xiao et al.

Recent advances in coding agents have shown remarkable progress in software issue resolution. In practice, real-world issues are typically bug fixes or feature requests in which human developers naturally incorporate refactoring as part of the resolution process, resulting in tangled refactoring. Since LLMs are trained on large-scale open-source repositories, coding agents may inherit such behaviors. In this paper, we conduct an empirical study on Multi-SWE-bench, analyzing 3,691 valid patches generated by three agent frameworks with 12 LLMs. We find that coding agents introduce tangled refactorings less frequently (21.43% vs. 36.72%) and with lower intensity (0.66 vs. 1.75) than human developers, although they exhibit a broader diversity of refactoring types. Logistic regression analysis further shows that tangled refactorings are strongly associated with reduced compilability, while exhibiting no significant association with functional correctness. Based on these findings, we propose a refactoring-aware refinement approach that assesses the necessity and safety of tangled refactorings and selectively removes or repairs problematic operations. Our approach improves compilability from 19.34% to 38.33%, and additionally resolves 2.79% previously unresolved issues. Overall, this work presents the first step towards understanding tangled refactoring practices in agentic issue resolution and opens up avenues for future work.

Lingyan Wu, Xiang Zheng, Weiqi Zhai et al.

Process-Level Reward Models (PRMs) are essential for guiding complex reasoning in large language models, yet existing PRM benchmarks cover only general domains such as mathematics, failing to address medical reasoning -- which is uniquely characterized by safety criticality, knowledge intensity, and diverse error patterns. Without a reliable medical PRM evaluation framework, we cannot quantify models' error detection capabilities in clinical reasoning, leaving their safety in real-world healthcare applications unverified. We propose MedPRMBench, the first process-level reward model benchmark for the medical domain. Built through a three-phase pipeline based on Clinical Reasoning Blueprints (CRBs), MedPRMBench systematically generates high-quality evaluation data from seven medical QA sources, covering 14 fine-grained error types across three categories (Simplicity, Soundness, and Sensitivity) with the first 4-level severity grading system to quantify clinical impact. The benchmark comprises 6{,}500 questions with 13{,}000 reasoning chains and 113{,}910 step-level labels, plus 6{,}879 questions for training. Our medical PRM baseline achieves an 87.1\% overall PRMScore -- substantially surpassing all baselines -- and serves as a plug-and-play verifier that improves downstream medical QA accuracy by 3.2--6.7 percentage points. Systematic evaluation spanning proprietary frontier models, open-source reasoning models, and medical-specialized models reveals critical weaknesses in current models' medical reasoning error detection capabilities, providing clear directions for future PRM improvement.

Zhengbo Jiao, Shaobo Wang, Zifan Zhang et al.

Multimodal Large Language Models (MLLMs) have significantly advanced vision-language understanding. However, even state-of-the-art models struggle with geometric reasoning, revealing a critical bottleneck: the extreme scarcity of high-quality image-text pairs. Human annotation is prohibitively expensive, while automated methods fail to ensure fidelity and training effectiveness. Existing approaches either passively adapt to available images or employ inefficient random exploration with filtering, decoupling generation from learning needs. We propose Socratic-Geo, a fully autonomous framework that dynamically couples data synthesis with model learning through multi-agent interaction. The Teacher agent generates parameterized Python scripts with reflective feedback (Reflect for solvability, RePI for visual validity), ensuring image-text pair purity. The Solver agent optimizes reasoning through preference learning, with failure paths guiding Teacher's targeted augmentation. Independently, the Generator learns image generation capabilities on accumulated "image-code-instruction" triplets, distilling programmatic drawing intelligence into visual generation. Starting from only 108 seed problems, Socratic-Solver achieves 49.11 on six benchmarks using one-quarter of baseline data, surpassing strong baselines by 2.43 points. Socratic-Generator achieves 42.4% on GenExam, establishing new state-of-the-art for open-source models, surpassing Seedream-4.0 (39.8%) and approaching Gemini-2.5-Flash-Image (43.1%).

Hairi, Minghong Fang, Zifan Zhang et al.

In this paper, we study a fully-decentralized multi-agent policy evaluation problem, which is an important sub-problem in cooperative multi-agent reinforcement learning, in the presence of up to $f$ faulty agents. In particular, we focus on the so-called Byzantine faulty model with model poisoning setting. In general, policy evaluation is to evaluate the value function of any given policy. In cooperative multi-agent system, the system-wide rewards are usually modeled as the uniform average of rewards from all agents. We investigate the multi-agent policy evaluation problem in the presence of Byzantine agents, particularly in the setting of heterogeneous local rewards. Ideally, the goal of the agents is to evaluate the accumulated system-wide rewards, which are uniform average of rewards of the normal agents for a given policy. It means that all agents agree upon common values (the consensus part) and furthermore, the consensus values are the value functions (the convergence part). However, we prove that this goal is not achievable. Instead, we consider a relaxed version of the problem, where the goal of the agents is to evaluate accumulated system-wide reward, which is an appropriately weighted average reward of the normal agents. We further prove that there is no correct algorithm that can guarantee that the total number of positive weights exceeds $|\mathcal{N}|-f $, where $|\mathcal{N}|$ is the number of normal agents. Towards the end, we propose a Byzantine-tolerant decentralized temporal difference algorithm that can guarantee asymptotic consensus under scalar function approximation. We then empirically test the effective of the proposed algorithm.

Zhengbo Jiao, Shaobo Wang, Zifan Zhang et al.

Advancing complex reasoning in large language models relies on high-quality, verifiable datasets, yet human annotation remains cost-prohibitive and difficult to scale. Current synthesis paradigms often face a recurring trade-off: maintaining structural validity typically restricts problem complexity, while relaxing constraints to increase difficulty frequently leads to inconsistent or unsolvable instances. To address this, we propose Agentic Proposing, a framework that models problem synthesis as a goal-driven sequential decision process where a specialized agent dynamically selects and composes modular reasoning skills. Through an iterative workflow of internal reflection and tool-use, we develop the Agentic-Proposer-4B using Multi-Granularity Policy Optimization (MGPO) to generate high-precision, verifiable training trajectories across mathematics, coding, and science. Empirical results demonstrate that downstream solvers trained on agent-synthesized data significantly outperform leading baselines and exhibit robust cross-domain generalization. Notably, a 30B solver trained on only 11,000 synthesized trajectories achieves a state-of-the-art 91.6% accuracy on AIME25, rivaling frontier-scale proprietary models such as GPT-5 and proving that a small volume of high-quality synthetic signals can effectively substitute for massive human-curated datasets.

Zhengbo Jiao, Hongyu Xian, Qinglong Wang et al.

Large language models (LLMs) struggle with complex, long-horizon reasoning due to instability caused by their frozen policy assumption. Current test-time scaling methods treat execution feedback merely as an external signal for filtering or rewriting trajectories, without internalizing it to improve the underlying reasoning strategy. Inspired by Popper's epistemology of "conjectures and refutations," we argue that intelligence requires real-time evolution of the model's policy through learning from failed attempts. We introduce Policy of Thoughts (PoT), a framework that recasts reasoning as a within-instance online optimization process. PoT first generates diverse candidate solutions via an efficient exploration mechanism, then uses Group Relative Policy Optimization (GRPO) to update a transient LoRA adapter based on execution feedback. This closed-loop design enables dynamic, instance-specific refinement of the model's reasoning priors. Experiments show that PoT dramatically boosts performance: a 4B model achieves 49.71% accuracy on LiveCodeBench, outperforming GPT-4o and DeepSeek-V3 despite being over 50 smaller.

Zhengbo Jiao, Shaobo Wang, Zifan Zhang et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet how visual evidence is integrated during reasoning remains poorly understood. We explore multimodal RLVR through the lens of cross-modal attention connectivity and find that only a small fraction of tokens (approximately 15%) exhibit strong visual-textual coupling. These high-connectivity tokens act as anchors that ground reasoning in the image, while the majority follow linguistic patterns. During RLVR training, credit assignment naturally concentrates on these anchors, sharpening their visual grounding over time. Building on this insight, we propose Anchor-Token Reinforcement Learning (AT-RL), a lightweight framework that selectively reinforces high-connectivity tokens via graph-based clustering of attention topology. Evaluated across the series (3B-32B), AT-RL introduces only 1.2% overhead yet enables the 32B model to surpass the 72B-Instruct baseline on MathVista (80.2), with consistent gains observed across STEM, video and general tasks. Conversely, training solely on low-connectivity tokens causes severe degradation, confirming that effective multimodal RL hinges on precise credit assignment to visual anchors. Our work reveals that reasoning quality is governed not by token quantity but by the fidelity of cross-modal anchoring.

Zifan Zhang, Dianwei Chen, Anjun Gao et al.

Network digital twins (NDTs) are transforming network management by offering precise virtual replicas of physical network systems. However, their reliance on diverse and sensitive data introduces significant challenges related to data management, regulatory compliance, and user privacy. In scenarios where selective data removal is necessary, such as device deactivation, network reconfiguration, or regulatory compliance, traditional approaches often fall short of preserving the integrity of the twin model. To address this gap, we introduce a network digital untwinning framework that enables the targeted removal of deprecated NDT contributions while maintaining model integrity. Our approach comprises two complementary components: Single Request Untwinning (\algO) and Parallel Request Untwinning (\algM) mechanisms. \algO leverages connectivity metrics based on geographical proximity, data distribution, and network-level attributes to identify and remove the target NDT along with its propagating influence. This is achieved through an optimally selected rollback checkpoint augmented with injected Gaussian noise, followed by a precise remapping phase. \algM extends this mechanism to efficiently handle multiple removal requests by clustering NDTs with similar attributes and performing a coordinated rollback and untwinning schedule. We provide theoretical guarantees on model indistinguishability from scratch-built twins, and validate the framework through extensive experiments on real-world traffic data, demonstrating its effectiveness and operational efficiency.

Zifan Zhang, Minghong Fang, Jiayuan Huang et al.

Federated Learning (FL) offers a distributed framework to train a global control model across multiple base stations without compromising the privacy of their local network data. This makes it ideal for applications like wireless traffic prediction (WTP), which plays a crucial role in optimizing network resources, enabling proactive traffic flow management, and enhancing the reliability of downstream communication-aided applications, such as IoT devices, autonomous vehicles, and industrial automation systems. Despite its promise, the security aspects of FL-based distributed wireless systems, particularly in regression-based WTP problems, remain inadequately investigated. In this paper, we introduce a novel fake traffic injection (FTI) attack, designed to undermine the FL-based WTP system by injecting fabricated traffic distributions with minimal knowledge. We further propose a defense mechanism, termed global-local inconsistency detection (GLID), which strategically removes abnormal model parameters that deviate beyond a specific percentile range estimated through statistical methods in each dimension. Extensive experimental evaluations, performed on real-world wireless traffic datasets, demonstrate that both our attack and defense strategies significantly outperform existing baselines.

Dianwei Chen, Zifan Zhang, Yuchen Liu et al.

Autonomous driving systems face significant challenges in handling unpredictable edge-case scenarios, such as adversarial pedestrian movements, dangerous vehicle maneuvers, and sudden environmental changes. Current end-to-end driving models struggle with generalization to these rare events due to limitations in traditional detection and prediction approaches. To address this, we propose INSIGHT (Integration of Semantic and Visual Inputs for Generalized Hazard Tracking), a hierarchical vision-language model (VLM) framework designed to enhance hazard detection and edge-case evaluation. By using multimodal data fusion, our approach integrates semantic and visual representations, enabling precise interpretation of driving scenarios and accurate forecasting of potential dangers. Through supervised fine-tuning of VLMs, we optimize spatial hazard localization using attention-based mechanisms and coordinate regression techniques. Experimental results on the BDD100K dataset demonstrate a substantial improvement in hazard prediction straightforwardness and accuracy over existing models, achieving a notable increase in generalization performance. This advancement enhances the robustness and safety of autonomous driving systems, ensuring improved situational awareness and potential decision-making in complex real-world scenarios.

Zifan Zhang, Mingzhe Chen, Zhaohui Yang et al.

In recent years, the complexity of 5G and beyond wireless networks has escalated, prompting a need for innovative frameworks to facilitate flexible management and efficient deployment. The concept of digital twins (DTs) has emerged as a solution to enable real-time monitoring, predictive configurations, and decision-making processes. While existing works primarily focus on leveraging DTs to optimize wireless networks, a detailed mapping methodology for creating virtual representations of network infrastructure and properties is still lacking. In this context, we introduce VH-Twin, a novel time-series data-driven framework that effectively maps wireless networks into digital reality. VH-Twin distinguishes itself through complementary vertical twinning (V-twinning) and horizontal twinning (H-twinning) stages, followed by a periodic clustering mechanism used to virtualize network regions based on their distinct geological and wireless characteristics. Specifically, V-twinning exploits distributed learning techniques to initialize a global twin model collaboratively from virtualized network clusters. H-twinning, on the other hand, is implemented with an asynchronous mapping scheme that dynamically updates twin models in response to network or environmental changes. Leveraging real-world wireless traffic data within a cellular wireless network, comprehensive experiments are conducted to verify that VH-Twin can effectively construct, deploy, and maintain network DTs. Parametric analysis also offers insights into how to strike a balance between twinning efficiency and model accuracy at scale.

Shaobo Wang, Zhengbo Jiao, Zifan Zhang et al.

Recent breakthroughs in large language models (LLMs) on reasoning tasks rely heavily on massive, high-quality datasets-typically human-annotated and thus difficult to scale. While data synthesis or distillation offers a promising alternative, existing methods struggle with inconsistent data quality and an inability to dynamically adapt to the evolving capabilities of the model, leading to suboptimal training signals. To address these limitations, we introduce Socratic-Zero, a fully autonomous framework that generates high-quality training data from minimal seed examples through the co-evolution of three agents: the Teacher, the Solver, and the Generator. The Solver continuously refines its reasoning by learning from preference feedback on both successful and failed trajectories; the Teacher adaptively crafts increasingly challenging questions based on the Solver's weaknesses; and the Generator distills the Teacher's question-design strategy to enable scalable, high-fidelity curriculum generation. This closed-loop system produces a self-improving curriculum-requiring no pre-existing tasks or labels. Remarkably, starting from only 100 seed questions, our Socratic-Solver-8B achieves an average gain of +20.2 percentage points over prior data synthesis methods across seven mathematical reasoning benchmarks (AMC23, AIME24-25, Olympiad, MATH-500, Minerva, and GSM8K), with consistent gains on both Qwen3 and GLM4 series models. Even more surprisingly, synthetic data from Socratic-Generator-32B enables student LLMs to achieve superior performance compared to other state-of-the-art (SOTA) commercial LLMs on these benchmarks, including Qwen3-235B-A22B, DeepSeek-V3.1-671B, GPT-5, Gemini-2.5-Pro, Grok-4, and Claude-4.1-Opus.

Fnu Hairi, Zifan Zhang, Jia Liu

In actor-critic framework for fully decentralized multi-agent reinforcement learning (MARL), one of the key components is the MARL policy evaluation (PE) problem, where a set of $N$ agents work cooperatively to evaluate the value function of the global states for a given policy through communicating with their neighbors. In MARL-PE, a critical challenge is how to lower the sample and communication complexities, which are defined as the number of training samples and communication rounds needed to converge to some $ε$-stationary point. To lower communication complexity in MARL-PE, a "natural'' idea is to perform multiple local TD-update steps between each consecutive rounds of communication to reduce the communication frequency. However, the validity of the local TD-update approach remains unclear due to the potential "agent-drift'' phenomenon resulting from heterogeneous rewards across agents in general. This leads to an interesting open question: Can the local TD-update approach entail low sample and communication complexities? In this paper, we make the first attempt to answer this fundamental question. We focus on the setting of MARL-PE with average reward, which is motivated by many multi-agent network optimization problems. Our theoretical and experimental results confirm that allowing multiple local TD-update steps is indeed an effective approach in lowering the sample and communication complexities of MARL-PE compared to consensus-based MARL-PE algorithms. Specifically, the local TD-update steps between two consecutive communication rounds can be as large as $\mathcal{O}(1/ε^{1/2}\log{(1/ε)})$ in order to converge to an $ε$-stationary point of MARL-PE. Moreover, we show theoretically that in order to reach the optimal sample complexity, the communication complexity of local TD-update approach is $\mathcal{O}(1/ε^{1/2}\log{(1/ε)})$.

Wenbin Wang, Qiwen Ma, Zifan Zhang et al.

Federated learning allows multiple clients to collaboratively train a global model with the assistance of a server. However, its distributed nature makes it susceptible to poisoning attacks, where malicious clients can compromise the global model by sending harmful local model updates to the server. To unlearn an accurate global model from a poisoned one after identifying malicious clients, federated unlearning has been introduced. Yet, current research on federated unlearning has primarily concentrated on its effectiveness and efficiency, overlooking the security challenges it presents. In this work, we bridge the gap via proposing BadUnlearn, the first poisoning attacks targeting federated unlearning. In BadUnlearn, malicious clients send specifically designed local model updates to the server during the unlearning process, aiming to ensure that the resulting unlearned model remains poisoned. To mitigate these threats, we propose UnlearnGuard, a robust federated unlearning framework that is provably robust against both existing poisoning attacks and our BadUnlearn. The core concept of UnlearnGuard is for the server to estimate the clients' local model updates during the unlearning process and employ a filtering strategy to verify the accuracy of these estimations. Theoretically, we prove that the model unlearned through UnlearnGuard closely resembles one obtained by train-from-scratch. Empirically, we show that BadUnlearn can effectively corrupt existing federated unlearning methods, while UnlearnGuard remains secure against poisoning attacks.

Zifan Zhang, Minghong Fang, Dianwei Chen et al.

Digital network twins (DNTs) are virtual representations of physical networks, designed to enable real-time monitoring, simulation, and optimization of network performance. When integrated with machine learning (ML) techniques, particularly federated learning (FL) and reinforcement learning (RL), DNTs emerge as powerful solutions for managing the complexities of network operations. This article presents a comprehensive analysis of the synergy of DNTs, FL, and RL techniques, showcasing their collective potential to address critical challenges in 6G networks. We highlight key technical challenges that need to be addressed, such as ensuring network reliability, achieving joint data-scenario forecasting, and maintaining security in high-risk environments. Additionally, we propose several pipelines that integrate DNT and ML within coherent frameworks to enhance network optimization and security. Case studies demonstrate the practical applications of our proposed pipelines in edge caching and vehicular networks. In edge caching, the pipeline achieves over 80% cache hit rates while balancing base station loads. In autonomous vehicular system, it ensure a 100% no-collision rate, showcasing its reliability in safety-critical scenarios. By exploring these synergies, we offer insights into the future of intelligent and adaptive network systems that automate decision-making and problem-solving.

Lei Xin, Yuhao Zheng, Ke Cheng et al.

Modeling long sequences of user behaviors has emerged as a critical frontier in generative recommendation. However, existing solutions face a dilemma: linear attention mechanisms achieve efficiency at the cost of retrieval precision due to limited state capacity, while softmax attention suffers from prohibitive computational overhead. To address this challenge, we propose HyTRec, a model featuring a Hybrid Attention architecture that explicitly decouples long-term stable preferences from short-term intent spikes. By assigning massive historical sequences to a linear attention branch and reserving a specialized softmax attention branch for recent interactions, our approach restores precise retrieval capabilities within industrial-scale contexts involving ten thousand interactions. To mitigate the lag in capturing rapid interest drifts within the linear layers, we furthermore design Temporal-Aware Delta Network (TADN) to dynamically upweight fresh behavioral signals while effectively suppressing historical noise. Empirical results on industrial-scale datasets confirm the superiority that our model maintains linear inference speed and outperforms strong baselines, notably delivering over 8% improvement in Hit Rate for users with ultra-long sequences with great efficiency.

Hu Wei, Ze Xu, Boyu Yang et al.

Large language models (LLMs) now perform strongly on many public math suites, yet frontier separation within mathematics increasingly suffers from ceiling effects. We present two complementary benchmarks: SKYLENAGE-ReasoningMATH, a 100-item, structure-aware diagnostic set with per-item metadata on length, numeric density, and symbolic complexity; and SKYLENAGE-MATH, a 150-item contest-style suite spanning four stages from high school to doctoral under a seven-subject taxonomy. We evaluate fifteen contemporary LLM variants under a single setup and analyze subject x model and grade x model performance. On the contest suite, the strongest model reaches 44% while the runner-up reaches 37%; accuracy declines from high school to doctoral, and top systems exhibit a doctoral-to-high-school retention near 79%. On the reasoning set, the best model attains 81% overall, and hardest-slice results reveal clear robustness gaps between leaders and the mid-tier. In summary, we release SKYLENAGE-ReasoningMATH and report aggregate results for SKYLENAGE-MATH; together, SKYLENAGE provides a hard, reasoning-centered and broadly covering math benchmark with calibrated difficulty and rich metadata, serving as a reference benchmark for future evaluations of mathematical reasoning.

Zifan Zhang, Minghong Fang, Mingzhe Chen et al.

The integration of digital twinning technologies is driving next-generation networks toward new capabilities, allowing operators to thoroughly understand network conditions, efficiently analyze valuable radio data, and innovate applications through user-friendly, immersive interfaces. Building on this foundation, network digital twins (NDTs) accurately depict the operational processes and attributes of network infrastructures, facilitating predictive management through real-time analysis and measurement. However, constructing precise NDTs poses challenges, such as integrating diverse data sources, mapping necessary attributes from physical networks, and maintaining scalability for various downstream tasks. Unlike previous works that focused on the creation and mapping of NDTs from scratch, we explore intra- and inter-operations among NDTs within a Unified Twin Transformation (UTT) framework, which uncovers a new computing paradigm for efficient transfer, merging, and splitting of NDTs to create task-oriented twins. By leveraging joint multi-modal and distributed mapping mechanisms, UTT optimizes resource utilization and reduces the cost of creating NDTs, while ensuring twin model consistency. A theoretical analysis of the distributed mapping problem is conducted to establish convergence bounds for this multi-modal gated aggregation process. Evaluations on real-world twin-assisted applications, such as trajectory reconstruction, human localization, and sensory data generation, demonstrate the feasibility and effectiveness of interoperability among NDTs for corresponding task development.

Zifan Zhang, Yuchen Liu, Zhiyuan Peng et al.

Optimizing edge caching is crucial for the advancement of next-generation (nextG) wireless networks, ensuring high-speed and low-latency services for mobile users. Existing data-driven optimization approaches often lack awareness of the distribution of random data variables and focus solely on optimizing cache hit rates, neglecting potential reliability concerns, such as base station overload and unbalanced cache issues. This oversight can result in system crashes and degraded user experience. To bridge this gap, we introduce a novel digital twin-assisted optimization framework, called D-REC, which integrates reinforcement learning (RL) with diverse intervention modules to ensure reliable caching in nextG wireless networks. We first develop a joint vertical and horizontal twinning approach to efficiently create network digital twins, which are then employed by D-REC as RL optimizers and safeguards, providing ample datasets for training and predictive evaluation of our cache replacement policy. By incorporating reliability modules into a constrained Markov decision process, D-REC can adaptively adjust actions, rewards, and states to comply with advantageous constraints, minimizing the risk of network failures. Theoretical analysis demonstrates comparable convergence rates between D-REC and vanilla data-driven methods without compromising caching performance. Extensive experiments validate that D-REC outperforms conventional approaches in cache hit rate and load balancing while effectively enforcing predetermined reliability intervention modules.

Minghong Fang, Zifan Zhang, Hairi et al.

Federated learning (FL) enables multiple clients to collaboratively train machine learning models without revealing their private training data. In conventional FL, the system follows the server-assisted architecture (server-assisted FL), where the training process is coordinated by a central server. However, the server-assisted FL framework suffers from poor scalability due to a communication bottleneck at the server, and trust dependency issues. To address challenges, decentralized federated learning (DFL) architecture has been proposed to allow clients to train models collaboratively in a serverless and peer-to-peer manner. However, due to its fully decentralized nature, DFL is highly vulnerable to poisoning attacks, where malicious clients could manipulate the system by sending carefully-crafted local models to their neighboring clients. To date, only a limited number of Byzantine-robust DFL methods have been proposed, most of which are either communication-inefficient or remain vulnerable to advanced poisoning attacks. In this paper, we propose a new algorithm called BALANCE (Byzantine-robust averaging through local similarity in decentralization) to defend against poisoning attacks in DFL. In BALANCE, each client leverages its own local model as a similarity reference to determine if the received model is malicious or benign. We establish the theoretical convergence guarantee for BALANCE under poisoning attacks in both strongly convex and non-convex settings. Furthermore, the convergence rate of BALANCE under poisoning attacks matches those of the state-of-the-art counterparts in Byzantine-free settings. Extensive experiments also demonstrate that BALANCE outperforms existing DFL methods and effectively defends against poisoning attacks.