Weilin Luo

Ka Yiu Lee, Yuxuan Huang, Zhiyuan He et al.

Recent agentic search systems have made substantial progress by emphasising deep, multi-step reasoning. However, this focus often overlooks the challenges of wide-scale information synthesis, where agents must aggregate large volumes of heterogeneous evidence across many sources. As a result, most existing large language model agent systems face severe limitations in data-intensive settings, including context saturation, cascading error propagation, and high end-to-end latency. To address these challenges, we present \framework, a hierarchical framework based on principle of near-decomposability, containing a strategic \textit{Host}, multiple \textit{Managers} and parallel \textit{Workers}. By leveraging aggregation and reflection mechanisms at the Manager layer, our framework enforces strict context isolation to prevent saturation and error propagation. Simultaneously, the parallelism in worker layer accelerates the speed of overall task execution, mitigating the significant latency. Our evaluation on two complementary benchmarks demonstrates both efficiency ($ 3-5 \times$ speed-up) and effectiveness, achieving a $8.4\%$ success rate on WideSearch-en and $52.9\%$ accuracy on BrowseComp-zh. The code is released at https://github.com/agent-on-the-fly/InfoSeeker

Rongzhen Ye, Tianqu Zhuang, Hai Wan et al.

We study the problem of learning a single occurrence regular expression with interleaving (SOIRE) from a set of text strings possibly with noise. SOIRE fully supports interleaving and covers a large portion of regular expressions used in practice. Learning SOIREs is challenging because it requires heavy computation and text strings usually contain noise in practice. Most of the previous studies only learn restricted SOIREs and are not robust on noisy data. To tackle these issues, we propose a noise-tolerant differentiable learning approach SOIREDL for SOIRE. We design a neural network to simulate SOIRE matching and theoretically prove that certain assignments of the set of parameters learnt by the neural network, called faithful encodings, are one-to-one corresponding to SOIREs for a bounded size. Based on this correspondence, we interpret the target SOIRE from an assignment of the set of parameters of the neural network by exploring the nearest faithful encodings. Experimental results show that SOIREDL outperforms the state-of-the-art approaches, especially on noisy data.

Wenkui Yang, Chao Jin, Haisu Zhu et al.

Existing red-teaming studies on GUI agents have important limitations. Adversarial perturbations typically require white-box access, which is unavailable for commercial systems, while prompt injection is increasingly mitigated by stronger safety alignment. To study robustness under a more practical threat model, we propose Semantic-level UI Element Injection, a red-teaming setting that overlays safety-aligned and harmless UI elements onto screenshots to misdirect the agent's visual grounding. Our method uses a modular Editor-Overlapper-Victim pipeline and an iterative search procedure that samples multiple candidate edits, keeps the best cumulative overlay, and adapts future prompt strategies based on previous failures. Across five victim models, our optimized attacks improve attack success rate by up to 4.4x over random injection on the strongest victims. Moreover, elements optimized on one source model transfer effectively to other target models, indicating model-agnostic vulnerabilities. After the first successful attack, the victim still clicks the attacker-controlled element in more than 15% of later independent trials, versus below 1% for random injection, showing that the injected element acts as a persistent attractor rather than simple visual clutter.

Yuxuan Huang, Yihang Chen, Zhiyuan He et al.

Agentic web search increasingly faces two distinct demands: deep reasoning over a single target, and structured aggregation across many entities and heterogeneous sources. Current systems struggle on both fronts. Breadth-oriented tasks demand schema-aligned outputs with wide coverage and cross-entity consistency, while depth-oriented tasks require coherent reasoning over long, branching search trajectories. We introduce \textbf{Web2BigTable}, a multi-agent framework for web-to-table search that supports both regimes. Web2BigTable adopts a bi-level architecture in which an upper-level orchestrator decomposes the task into sub-problems and lower-level worker agents solve them in parallel. Through a closed-loop run--verify--reflect process, the framework jointly improves decomposition and execution over time via persistent, human-readable external memory, with self-evolving updates to each single-agent. During execution, workers coordinate through a shared workspace that makes partial findings visible, allowing them to reduce redundant exploration, reconcile conflicting evidence, and adapt to emerging coverage gaps. Web2BigTable sets a new state of the art on WideSearch, reaching an Avg@4 Success Rate of \textbf{38.50} ($7.5\times$ the second best at 5.10), Row F1 of \textbf{63.53} (+25.03 over the second best), and Item F1 of \textbf{80.12} (+14.42 over the second best). It also generalises to depth-oriented search on XBench-DeepSearch, achieving 73.0 accuracy. Code is available at https://github.com/web2bigtable/web2bigtable.

Zhengxu Yu, Yu Fu, Zhiyuan He et al.

Individual agent capabilities have advanced rapidly through modular skills and tool integrations, yet multi-agent systems remain constrained by fixed team structures, tightly coupled coordination logic, and session-bound learning. We argue that this reflects a deeper absence: a principled organisational layer that governs how a workforce of agents is assembled, governed, and improved over time, decoupled from what individual agents know. To fill this gap, we introduce \emph{OneManCompany (OMC)}, a framework that elevates multi-agent systems to the organisational level. OMC encapsulates skills, tools, and runtime configurations into portable agent identities called \emph{Talents}, orchestrated through typed organisational interfaces that abstract over heterogeneous backends. A community-driven \emph{Talent Market} enables on-demand recruitment, allowing the organisation to close capability gaps and reconfigure itself dynamically during execution. Organisational decision-making is operationalised through an \emph{Explore-Execute-Review} ($\text{E}^2$R) tree search, which unifies planning, execution, and evaluation in a single hierarchical loop: tasks are decomposed top-down into accountable units and execution outcomes are aggregated bottom-up to drive systematic review and refinement. This loop provides formal guarantees on termination and deadlock freedom while mirroring the feedback mechanisms of human enterprises. Together, these contributions transform multi-agent systems from static, pre-configured pipelines into self-organising and self-improving AI organisations capable of adapting to open-ended tasks across diverse domains. Empirical evaluation on PRDBench shows that OMC achieves an $84.67\%$ success rate, surpassing the state of the art by $15.48$ percentage points, with cross-domain case studies further demonstrating its generality.

Weilin Luo, Xueyi Liang, Haotian Deng et al.

Automatically synthesizing verifiable code from natural language requirements ensures software correctness and reliability while significantly lowering the barrier to adopting the techniques of formal methods. With the rise of large language models (LLMs), long-standing efforts at autoformalization have gained new momentum. However, existing approaches suffer from severe syntactic and semantic errors due to the scarcity of domain-specific pre-training corpora and often fail to formalize implicit knowledge effectively. In this paper, we propose AutoICE, an LLM-driven evolutionary search for synthesizing verifiable C code. It introduces the diverse individual initialization and the collaborative crossover to enable diverse iterative updates, thereby mitigating error propagation inherent in single-agent iterations. Besides, it employs the self-reflective mutation to facilitate the discovery of implicit knowledge. Evaluation results demonstrate the effectiveness of AutoICE: it successfully verifies $90.36$\% of code, outperforming the state-of-the-art (SOTA) approach. Besides, on a developer-friendly dataset variant, AutoICE achieves a $88.33$\% verification success rate, significantly surpassing the $65$\% success rate of the SOTA approach.

Jingyuan Ma, Damai Dai, Zihang Yuan et al.

Large Language Models (LLMs) have shown remarkable success on a wide range of math and reasoning benchmarks. However, we observe that they often struggle when faced with unreasonable math problems. Instead of recognizing these issues, models frequently proceed as if the problem is well-posed, producing incorrect answers or falling into overthinking and verbose self-correction. To systematically investigate this overlooked vulnerability, we propose the \textbf{Unreasonable Math Problems (UMP)} benchmark, designed to evaluate LLMs' ability to detect and respond to unreasonable math problem statements. Based on extensive experiments covering 19 LLMs, we find that even state-of-the-art general models like GPT-4o achieve only a score of 0.6 on UMP. While reasoning models such as DeepSeek-R1 demonstrate a higher sensitivity to unreasonable inputs, this often comes at the cost of generating overly long and meaningless responses that fail to converge. We further explore prompting and fine-tuning methods, which offer partial improvements but also introduce trade-offs, shedding light on both the potential and limitations of LLMs in this challenging setting.

Xijun Li, Fangzhou Zhu, Hui-Ling Zhen et al.

In an era of digital ubiquity, efficient resource management and decision-making are paramount across numerous industries. To this end, we present a comprehensive study on the integration of machine learning (ML) techniques into Huawei Cloud's OptVerse AI Solver, which aims to mitigate the scarcity of real-world mathematical programming instances, and to surpass the capabilities of traditional optimization techniques. We showcase our methods for generating complex SAT and MILP instances utilizing generative models that mirror multifaceted structures of real-world problem. Furthermore, we introduce a training framework leveraging augmentation policies to maintain solvers' utility in dynamic environments. Besides the data generation and augmentation, our proposed approaches also include novel ML-driven policies for personalized solver strategies, with an emphasis on applications like graph convolutional networks for initial basis selection and reinforcement learning for advanced presolving and cut selection. Additionally, we detail the incorporation of state-of-the-art parameter tuning algorithms which markedly elevate solver performance. Compared with traditional solvers such as Cplex and SCIP, our ML-augmented OptVerse AI Solver demonstrates superior speed and precision across both established benchmarks and real-world scenarios, reinforcing the practical imperative and effectiveness of machine learning techniques in mathematical programming solvers.

Weidong Zhan, Yue Wang, Nan Hu et al. · pku

Currently, long-chain reasoning remains a key challenge for large language models (LLMs) because natural texts lack sufficient explicit reasoning data. However, existing benchmarks suffer from limitations such as narrow coverage, short reasoning paths, or high construction costs. We introduce SCoRE (Scenario-based Commonsense Reasoning Evaluation), a benchmark that synthesizes multi-hop questions from scenario schemas of entities, relations, and logical rules to assess long-chain commonsense reasoning. SCoRE contains 100k bilingual (Chinese-English) multiple-choice questions whose reasoning chains span 2-11 hops and are grouped into various difficulty levels. Each question is accompanied by fine-grained knowledge labels, explicit reasoning chains, and difficulty levels for diagnostic evaluation. Evaluation results on cutting-edge LLMs such as o3-mini and Deepseek R1 shows that even the best model attains only 69.78% accuracy on SCoRE (even only 47.91% on the hard set), with errors often stemming from rare knowledge, logical inconsistency, and over-interpretation of simple questions. SCoRE offers a scalable, extensible framework for evaluating and diagnosing the long-chain commonsense reasoning abilities of LLMs and guiding future advances in model design and training.

Xijun Li, Weilin Luo, Mingxuan Yuan et al.

The Dynamic Pickup and Delivery Problem (DPDP) is aimed at dynamically scheduling vehicles among multiple sites in order to minimize the cost when delivery orders are not known a priori. Although DPDP plays an important role in modern logistics and supply chain management, state-of-the-art DPDP algorithms are still limited on their solution quality and efficiency. In practice, they fail to provide a scalable solution as the numbers of vehicles and sites become large. In this paper, we propose a data-driven approach, Spatial-Temporal Aided Double Deep Graph Network (ST-DDGN), to solve industry-scale DPDP. In our method, the delivery demands are first forecast using spatial-temporal prediction method, which guides the neural network to perceive spatial-temporal distribution of delivery demand when dispatching vehicles. Besides, the relationships of individuals such as vehicles are modelled by establishing a graph-based value function. ST-DDGN incorporates attention-based graph embedding with Double DQN (DDQN). As such, it can make the inference across vehicles more efficiently compared with traditional methods. Our method is entirely data driven and thus adaptive, i.e., the relational representation of adjacent vehicles can be learned and corrected by ST-DDGN from data periodically. We have conducted extensive experiments over real-world data to evaluate our solution. The results show that ST-DDGN reduces 11.27% number of the used vehicles and decreases 13.12% total transportation cost on average over the strong baselines, including the heuristic algorithm deployed in our UAT (User Acceptance Test) environment and a variety of vanilla DRL methods. We are due to fully deploy our solution into our online logistics system and it is estimated that millions of USD logistics cost can be saved per year.

Weilin Luo, Hai Wan, Xiaotong Song et al.

The boundary conditions (BCs) have shown great potential in requirements engineering because a BC captures the particular combination of circumstances, i.e., divergence, in which the goals of the requirement cannot be satisfied as a whole. Existing researches have attempted to automatically identify lots of BCs. Unfortunately, a large number of identified BCs make assessing and resolving divergences expensive. Existing methods adopt a coarse-grained metric, generality, to filter out less general BCs. However, the results still retain a large number of redundant BCs since a general BC potentially captures redundant circumstances that do not lead to a divergence. Furthermore, the likelihood of BC can be misled by redundant BCs resulting in costly repeatedly assessing and resolving divergences. In this paper, we present a fine-grained metric to filter out the redundant BCs. We first introduce the concept of contrasty of BC. Intuitively, if two BCs are contrastive, they capture different divergences. We argue that a set of contrastive BCs should be recommended to engineers, rather than a set of general BCs that potentially only indicates the same divergence. Then we design a post-processing framework (PPAc) to produce a set of contrastive BCs after identifying BCs. Experimental results show that the contrasty metric dramatically reduces the number of BCs recommended to engineers. Results also demonstrate that lots of BCs identified by the state-of-the-art method are redundant in most cases. Besides, to improve efficiency, we propose a joint framework (JAc) to interleave assessing based on the contrasty metric with identifying BCs. The primary intuition behind JAc is that it considers the search bias toward contrastive BCs during identifying BCs, thereby pruning the BCs capturing the same divergence. Experiments confirm the improvements of JAc in identifying contrastive BCs.

Hongzhen Zhong, Hai Wan, Weilin Luo et al.

In goal-oriented requirements engineering, goal conflict identification is of fundamental importance for requirements analysis. The task aims to find the feasible situations which make the goals diverge within the domain, called boundary conditions (BCs). However, the existing approaches for goal conflict identification fail to find sufficient BCs and general BCs which cover more combinations of circumstances. From the BCs found by these existing approaches, we have observed an interesting phenomenon that there are some pairs of BCs are similar in formula structure, which occurs frequently in the experimental cases. In other words, once a BC is found, a new BC may be discovered quickly by slightly changing the former. It inspires us to develop a local search algorithm named LOGION to find BCs, in which the structural similarity is captured by the neighborhood relation of formulae. Based on structural similarity, LOGION can find a lot of BCs in a short time. Moreover, due to the large number of BCs identified, it potentially selects more general BCs from them. By taking experiments on a set of cases, we show that LOGION effectively exploits the structural similarity of BCs. We also compare our algorithm against the two state-of-the-art approaches. The experimental results show that LOGION produces one order of magnitude more BCs than the state-of-the-art approaches and confirm that LOGION finds out more general BCs thanks to a large number of BCs.

Weilin Luo, Hai Wan, Hongzhen Zhong et al.

Prime compilation, i.e., the generation of all prime implicates or implicants (primes for short) of formulae, is a prominent fundamental issue for AI. Recently, the prime compilation for non-clausal formulae has received great attention. The state-of-the-art approaches generate all primes along with a prime cover constructed by prime implicates using dual rail encoding. However, the dual rail encoding potentially expands search space. In addition, constructing a prime cover, which is necessary for their methods, is time-consuming. To address these issues, we propose a novel two-phase method -- CoAPI. The two phases are the key to construct a cover without using dual rail encoding. Specifically, given a non-clausal formula, we first propose a core-guided method to rewrite the non-clausal formula into a cover constructed by over-approximate implicates in the first phase. Then, we generate all the primes based on the cover in the second phase. In order to reduce the size of the cover, we provide a multi-order based shrinking method, with a good tradeoff between the small size and efficiency, to compress the size of cover considerably. The experimental results show that CoAPI outperforms state-of-the-art approaches. Particularly, for generating all prime implicates, CoAPI consumes about one order of magnitude less time.

Peilin Chen, Hai Wan, Shaowei Cai et al.

The Maximum k-plex Problem is an important combinatorial optimization problem with increasingly wide applications. Due to its exponential time complexity, many heuristic methods have been proposed which can return a good-quality solution in a reasonable time. However, most of the heuristic algorithms are memoryless and unable to utilize the experience during the search. Inspired by the multi-armed bandit (MAB) problem in reinforcement learning (RL), we propose a novel perturbation mechanism named BLP, which can learn online to select a good vertex for perturbation when getting stuck in local optima. To our best of knowledge, this is the first attempt to combine local search with RL for the maximum $ k $-plex problem. Besides, we also propose a novel strategy, named Dynamic-threshold Configuration Checking (DTCC), which extends the original Configuration Checking (CC) strategy from two aspects. Based on the BLP and DTCC, we develop a local search algorithm named BDCC and improve it by a hyperheuristic strategy. The experimental result shows that our algorithms dominate on the standard DIMACS and BHOSLIB benchmarks and achieve state-of-the-art performance on massive graphs.