Large language model-based multi-agent systems are evolving from fixed interaction topologies toward dynamically orchestrated Agent Swarms. However, existing benchmarks are still largely based on single-agent or general-purpose agent tasks, making it difficult to systematically evaluate key orchestration capabilities. We propose SwarmBench, a benchmark that evaluates model performance from multiple perspectives, including accuracy, efficiency, cost, and process quality. Experimental results show that current models exhibit substantial differences in orchestration capability. These differences are reflected not only in final accuracy, efficiency, and cost, but also in the overall quality of the orchestration process itself. Based on these findings, we further propose SwarmExp, a simple yet effective method based on experience extraction and experience replay, which consistently improves the orchestration performance of large language models.

Large language model (LLM)-based agents solve complex long-horizon tasks through iterative interaction with environments, where effectively learning from experience under sparse or delayed feedback remains a key challenge. Existing approaches typically rely on either non-parametric memory-based experience reusing or parametric reinforcement learning, each with inherent limitations when used in isolation. To this end, we propose **ExpRL**, a unified experience learning framework that bridges non-parametric memory and parametric policy optimization for LLM agents. ExpRL induces abstracted success and failure experiences from on-policy interaction trajectories, stores them in an external memory, and integrates experience acquisition and utilization directly into the reinforcement learning process via an agentic experience acquisition mechanism. This design enables memory and policy to co-evolve and mutually reinforce each other, effectively unifying online policy learning with offline experience reuse. Experiments on ALFWorld and WebShop demonstrate that ExpRL achieves an average performance improvement of **13.0** over GRPO, with improved learning efficiency and stability. We further show that the induced experience memory generalizes across different models and can be directly transferred to stronger LLMs without additional training.

Multimodal web agents can assist humans in operating repetitive GUI tasks, where effective task planning is essential for decomposing complex tasks into executable actions. While small open-source MLLMs are cost-efficient and privacy-preserving compared with commercial large models, they suffer from weak planning and limited cross-website generalization. To address these limitations, we introduce the planning experience exploration and utilization (PEEU) method, which autonomously explores environments to discover experiences and utilizes hindsight experience to synthesize strictly aligned, high-level training data. To quantitatively analyze the generalization behaviors driving this performance, we propose the task decomposition hierarchical analysis framework (TDHAF) to systematically study compositional generalization across three task granularities: low, middle and high levels. Our analysis reveals that mastering low-level atomic skills does not guarantee high-level planning competence, while high-level task training yields stronger OOD generalization. Experiments on real-world benchmarks demonstrate PEEU’s superior effectiveness: our 7B model achieves 30.6% accuracy, outperforming the much larger Qwen2.5-VL-32B model. These demonstrate constructing hindsight high-level tasks and leveraging experiences is crucial for OOD planning abilities of small MLLMs.

Despite the significant progress made by existing retrieval augmented language models (RALMs) in providing trustworthy responses and grounding in reliable sources, they often overlook effective alignment with human preferences. In the alignment process, reward models (RMs) act as a crucial proxy for human values to guide optimization. However, it remains unclear how to evaluate and select a reliable RM for preference alignment in RALMs. To this end, we propose RAG-RewardBench, the first benchmark for evaluating RMs in RAG settings. First, we design four crucial and challenging RAG-specific scenarios to assess RMs, including multi-hop reasoning, fine-grained citation, appropriate abstain, and conflict robustness. Then, we incorporate 18 RAG subsets, six retrievers, and 24 RALMs to increase the diversity of data sources. Finally, we adopt an LLM-as-a-judge approach to improve preference annotation efficiency and effectiveness, exhibiting a strong correlation with human annotations. Based on the RAG-RewardBench, we conduct a comprehensive evaluation of 45 RMs and uncover their limitations in RAG scenarios. Additionally, we also reveal that existing trained RALMs show almost no improvement in preference alignment, highlighting the need for a shift towards preference-aligned training.

With the development of large language models, they are widely used as agents in various fields. A key component of agents is memory, which stores vital information but is susceptible to jailbreak attacks. Existing research mainly focuses on single-agent attacks and shared memory attacks. However, real-world scenarios often involve independent memory. In this paper, we propose the Troublemaker Makes Chaos in Honest Town (TMCHT) task, a large-scale, multi-agent, multi-topology text-based attack evaluation framework. TMCHT involves one attacker agent attempting to mislead an entire society of agents. We identify two major challenges in multi-agent attacks: (1) Non-complete graph structure, (2) Large-scale systems. We attribute these challenges to a phenomenon we term toxicity disappearing. To address these issues, we propose an Adversarial Replication Contagious Jailbreak (ARCJ) method, which optimizes the retrieval suffix to make poisoned samples more easily retrieved and optimizes the replication suffix to make poisoned samples have contagious ability. We demonstrate the superiority of our approach in TMCHT, with 23.51%, 18.95%, and 52.93% improvements in line, star topologies, and 100-agent settings. It reveals potential contagion risks in widely used multi-agent architectures.

As Multimodal Large Language Models (MLLMs) advance, multimodal agents show promise in real-world tasks like web navigation and embodied intelligence. However, due to limitations in a lack of external feedback, these agents struggle with self-correction and generalization. A promising approach is to use reward models as external feedback, but there is no clear on how to select reward models for agents. Thus, there is an urgent need to build a reward bench targeted at agents. To address these challenges, we propose Agent-RewardBench, a benchmark designed to evaluate reward modeling ability in MLLMs. The benchmark is characterized by three key features: (1) Multiple dimensions and real-world agent scenarios evaluation. It covers perception, planning, and safety with 7 scenarios; (2) Step-level reward evaluation. It allows for the assessment of agent capabilities at the individual steps of a task, providing a more granular view of performance during the planning process; and (3) Appropriately difficulty and high-quality. We carefully sample from 10 diverse models, difficulty control to maintain task challenges, and manual verification to ensure the integrity of the data. Experiments demonstrate that even state-of-the-art multimodal models show limited performance, highlighting the need for specialized training in agent reward modeling. Code is available at github.

Planning, as the core module of agents, is crucial in various fields such as embodied agents, web navigation, and tool using. With the development of large language models (LLMs), some researchers treat large language models as intelligent agents to stimulate and evaluate their planning capabilities. However, the planning mechanism is still unclear. In this work, we focus on exploring the look-ahead planning mechanism in large language models from the perspectives of information flow and internal representations. First, we study how planning is done internally by analyzing the multi-layer perception (MLP) and multi-head self-attention (MHSA) components at the last token. We find that the output of MHSA in the middle layers at the last token can directly decode the decision to some extent. Based on this discovery, we further trace the source of MHSA by information flow, and we reveal that MHSA extracts information from spans of the goal states and recent steps. According to information flow, we continue to study what information is encoded within it. Specifically, we explore whether future decisions have been considered in advance in the representation of flow. We demonstrate that the middle and upper layers encode a few short-term future decisions. Overall, our research analyzes the look-ahead planning mechanisms of LLMs, facilitating future research on LLMs performing planning tasks.

As the first step of modern natural language processing, text representation encodes discrete texts as continuous embeddings. Pre-trained language models (PLMs) have demonstrated strong ability in text representation and significantly promoted the development of natural language understanding (NLU). However, existing PLMs represent a text solely by its context, which is not enough to support knowledge-intensive NLU tasks. Knowledge is power, and fusing external knowledge explicitly into PLMs can provide knowledgeable text representations. Since previous knowledge-enhanced methods differ in many aspects, making it difficult for us to reproduce previous methods, implement new methods, and transfer between different methods. It is highly desirable to have a unified paradigm to encompass all kinds of methods in one framework. In this paper, we propose CogKTR, a knowledge-enhanced text representation toolkit for natural language understanding. According to our proposed Unified Knowledge-Enhanced Paradigm (UniKEP), CogKTR consists of four key stages, including knowledge acquisition, knowledge representation, knowledge injection, and knowledge application. CogKTR currently supports easy-to-use knowledge acquisition interfaces, multi-source knowledge embeddings, diverse knowledge-enhanced models, and various knowledge-intensive NLU tasks. Our unified, knowledgeable and modular toolkit is publicly available at GitHub, with an online system and a short instruction video.

In this paper, we propose CogKGE, a knowledge graph embedding (KGE) toolkit, which aims to represent multi-source and heterogeneous knowledge. For multi-source knowledge, unlike existing methods that mainly focus on entity-centric knowledge, CogKGE also supports the representations of event-centric, commonsense and linguistic knowledge. For heterogeneous knowledge, besides structured triple facts, CogKGE leverages additional unstructured information, such as text descriptions, node types and temporal information, to enhance the meaning of embeddings. Designing CogKGE aims to provide a unified programming framework for KGE tasks and a series of knowledge representations for downstream tasks. As a research framework, CogKGE consists of five parts, including core, data, model, knowledge and adapter module. As a knowledge discovery toolkit, CogKGE provides pre-trained embedders to discover new facts, cluster entities and check facts. Furthermore, we construct two benchmark datasets for further research on multi-source heterogeneous KGE tasks: EventKG240K and CogNet360K. We also release an online system to discover knowledge visually. Source code, datasets and pre-trained embeddings are publicly available at GitHub, with a short instruction video.