Wenjie Lu

Yingsha Xie, Tiansheng Huang, Enneng Yang et al.

Safety alignment incurs safety tax that perturbs a large reasoning model's (LRM) general reasoning ability. Existing datasets used for safety alignment for an LRM are usually constructed by distilling safety reasoning traces and answers from an external LRM or human labeler. However, such reasoning traces and answers exhibit a distributional gap with the target LRM that needs alignment, and we conjecture such distributional gap is the culprit leading to significant degradation of reasoning ability of the target LRM. Driven by this hypothesis, we propose a safety alignment dataset construction method, dubbed DGR. DGR transforms and refines an existing out-of-distributional safety reasoning dataset to be aligned with the target's LLM inner distribution. Experimental results demonstrate that i) DGR effectively mitigates the safety tax while maintaining safety performance across all baselines, i.e., achieving \textbf{+30.2\%} on DirectRefusal and \textbf{+21.2\%} on R1-ACT improvement in average reasoning accuracy compared to Vanilla SFT; ii) the degree of reasoning degradation correlates with the extent of distribution shift, suggesting that bridging this gap is central to preserving capabilities. Furthermore, we find that safety alignment in LRMs may primarily function as a mechanism to activate latent knowledge, as a mere \textbf{10} samples are sufficient for activating effective refusal behaviors. These findings not only emphasize the importance of distributional consistency but also provide insights into the activation mechanism of safety in reasoning models.

Hongze Mi, Yibo Feng, WenJie Lu et al.

Multimodal Large Language Model (MLLM) agents facilitate Graphical User Interface (GUI) automation but struggle with long-horizon, cross-application tasks due to limited context windows. While memory systems provide a viable solution, existing paradigms struggle to adapt to dynamic GUI environments, suffering from a granularity mismatch between high-level intent and low-level execution, and context pollution where the static accumulation of outdated experiences drives agents into hallucination. To address these bottlenecks, we propose the Darwinian Memory System (DMS), a self-evolving architecture that constructs memory as a dynamic ecosystem governed by the law of survival of the fittest. DMS decomposes complex trajectories into independent, reusable units for compositional flexibility, and implements Utility-driven Natural Selection to track survival value, actively pruning suboptimal paths and inhibiting high-risk plans. This evolutionary pressure compels the agent to derive superior strategies. Extensive experiments on real-world multi-app benchmarks validate that DMS boosts general-purpose MLLMs without training costs or architectural overhead, achieving average gains of 18.0% in success rate and 33.9% in execution stability, while reducing task latency, establishing it as an effective self-evolving memory system for GUI tasks.

Yanming Li, Xuelin Zhang, WenJie Lu et al.

Integrating Large Language Models (LLMs) with external tools via multi-agent systems offers a promising new paradigm for decomposing and solving complex problems. However, training these systems remains notoriously difficult due to the credit assignment challenge, as it is often unclear which specific functional agent is responsible for the success or failure of decision trajectories. Existing methods typically rely on sparse or globally broadcast rewards, failing to capture individual contributions and leading to inefficient reinforcement learning. To address these limitations, we introduce the Shapley-based Hierarchical Attribution for Reinforcement Policy (SHARP), a novel framework for optimizing multi-agent reinforcement learning via precise credit attribution. SHARP effectively stabilizes training by normalizing agent-specific advantages across trajectory groups, primarily through a decomposed reward mechanism comprising a global broadcast-accuracy reward, a Shapley-based marginal-credit reward for each agent, and a tool-process reward to improve execution efficiency. Extensive experiments across various real-world benchmarks demonstrate that SHARP significantly outperforms recent state-of-the-art baselines, achieving average match improvements of 23.66% and 14.05% over single-agent and multi-agent approaches, respectively.

Haotian Luo, Huaisong Zhang, Xuelin Zhang et al.

Autonomous agents have recently achieved remarkable progress across diverse domains, yet most evaluations focus on short-horizon, fully observable tasks. In contrast, many critical real-world tasks, such as large-scale software development, commercial investment, and scientific discovery, unfold in long-horizon and partially observable scenarios where success hinges on sustained reasoning, planning, memory management, and tool use. Existing benchmarks rarely capture these long-horizon challenges, leaving a gap in systematic evaluation. To bridge this gap, we introduce \textbf{UltraHorizon} a novel benchmark that measures the foundational capabilities essential for complex real-world challenges. We use exploration as a unifying task across three distinct environments to validate these core competencies. Agents are designed in long-horizon discovery tasks where they must iteratively uncover hidden rules through sustained reasoning, planning, memory and tools management, and interaction with environments. Under the heaviest scale setting, trajectories average \textbf{200k+} tokens and \textbf{400+} tool calls, whereas in standard configurations they still exceed \textbf{35k} tokens and involve more than \textbf{60} tool calls on average. Our extensive experiments reveal that LLM-agents consistently underperform in these settings, whereas human participants achieve higher scores, underscoring a persistent gap in agents' long-horizon abilities. We also observe that simple scaling fails in our task. To better illustrate the failure of agents, we conduct an in-depth analysis of collected trajectories. We identify eight types of errors and attribute them to two primary causes: in-context locking and functional fundamental capability gaps. \href{https://github.com/StarDewXXX/UltraHorizon}{Our code will be available here.}

Lejla Skelic, Yan Xu, Matthew Cox et al.

The role of Large Language Models (LLMs) has not been extensively explored in analog circuit design, which could benefit from a reasoning-based approach that transcends traditional optimization techniques. In particular, despite their growing relevance, there are no benchmarks to assess LLMs' reasoning capability about circuits. Therefore, we created the CIRCUIT dataset consisting of 510 question-answer pairs spanning various levels of analog-circuit-related subjects. The best-performing model on our dataset, GPT-4o, achieves 48.04% accuracy when evaluated on the final numerical answer. To evaluate the robustness of LLMs on our dataset, we introduced a unique feature that enables unit-test-like evaluation by grouping questions into unit tests. In this case, GPT-4o can only pass 27.45% of the unit tests, highlighting that the most advanced LLMs still struggle with understanding circuits, which requires multi-level reasoning, particularly when involving circuit topologies. This circuit-specific benchmark highlights LLMs' limitations, offering valuable insights for advancing their application in analog integrated circuit design.

Hongze Mi, Yibo Feng, Wenjie Lu et al.

Graphical User Interface (GUI) agents aim to automate a wide spectrum of human tasks by emulating user interaction. Despite rapid advancements, current approaches are hindered by several critical challenges: data bottleneck in end-to-end training, high cost of delayed error detection, and risk of contradictory guidance. Inspired by the human cognitive loop of Thinking, Alignment, and Reflection, we present D-Artemis -- a novel deliberative framework in this paper. D-Artemis leverages a fine-grained, app-specific tip retrieval mechanism to inform its decision-making process. It also employs a proactive Pre-execution Alignment stage, where Thought-Action Consistency (TAC) Check module and Action Correction Agent (ACA) work in concert to mitigate the risk of execution failures. A post-execution Status Reflection Agent (SRA) completes the cognitive loop, enabling strategic learning from experience. Crucially, D-Artemis enhances the capabilities of general-purpose Multimodal large language models (MLLMs) for GUI tasks without the need for training on complex trajectory datasets, demonstrating strong generalization. D-Artemis establishes new state-of-the-art (SOTA) results across both major benchmarks, achieving a 75.8% success rate on AndroidWorld and 96.8% on ScreenSpot-V2. Extensive ablation studies further demonstrate the significant contribution of each component to the framework.

Zhiyuan Cai, Huanhui Cao, Wenjie Lu et al.

Multi-Agent Reinforcement Learning (MARL) algorithms show amazing performance in simulation in recent years, but placing MARL in real-world applications may suffer safety problems. MARL with centralized shields was proposed and verified in safety games recently. However, centralized shielding approaches can be infeasible in several real-world multi-agent applications that involve non-cooperative agents or communication delay. Thus, we propose to combine MARL with decentralized Control Barrier Function (CBF) shields based on available local information. We establish a safe MARL framework with decentralized multiple CBFs and develop Multi-Agent Deep Deterministic Policy Gradient (MADDPG) to Multi-Agent Deep Deterministic Policy Gradient with decentralized multiple Control Barrier Functions (MADDPG-CBF). Based on a collision-avoidance problem that includes not only cooperative agents but obstacles, we demonstrate the construction of multiple CBFs with safety guarantees in theory. Experiments are conducted and experiment results verify that the proposed safe MARL framework can guarantee the safety of agents included in MARL.

Tianming Wang, Wenjie Lu, Huan Yu et al.

Underwater robots in shallow waters usually suffer from strong wave forces, which may frequently exceed robot's control constraints. Learning-based controllers are suitable for disturbance rejection control, but the excessive disturbances heavily affect the state transition in Markov Decision Process (MDP) or Partially Observable Markov Decision Process (POMDP). Also, pure learning procedures on targeted system may encounter damaging exploratory actions or unpredictable system variations, and training exclusively on a prior model usually cannot address model mismatch from the targeted system. In this paper, we propose a transfer learning framework that adapts a control policy for excessive disturbance rejection of an underwater robot under dynamics model mismatch. A modular network of learning policies is applied, composed of a Generalized Control Policy (GCP) and an Online Disturbance Identification Model (ODI). GCP is first trained over a wide array of disturbance waveforms. ODI then learns to use past states and actions of the system to predict the disturbance waveforms which are provided as input to GCP (along with the system state). A transfer reinforcement learning algorithm using Transition Mismatch Compensation (TMC) is developed based on the modular architecture, that learns an additional compensatory policy through minimizing mismatch of transitions predicted by the two dynamics models of the source and target tasks. We demonstrated on a pose regulation task in simulation that TMC is able to successfully reject the disturbances and stabilize the robot under an empirical model of the robot system, meanwhile improve sample efficiency.

Wenjie Lu, Dikai Liu

Reinforcement Learning (RL) is limited in practice by its gray-box nature, which is responsible for insufficient trustiness from users, unsatisfied interpretation for human intervention, inadequate analysis for future improvement, etc. This paper seeks to partially characterize the interplay between dynamical environments and the DOB-net. The DOB-net obtained from RL solves a set of Partially Observable Markovian Decision Processes (POMDPs). The transition function of each POMDP is largely determined by the environments, which are excessive external disturbances in this research. This paper proposes an Attention-based Abstraction (A${}^2$) approach to extract a finite-state automaton, referred to as a Key Moore Machine Network (KMMN), to capture the switching mechanisms exhibited by the DOB-net in dealing with multiple such POMDPs. This approach first quantizes the controlled platform by learning continuous-discrete interfaces. Then it extracts the KMMN by finding the key hidden states and transitions that attract sufficient attention from the DOB-net. Within the resultant KMMN, this study found three patterns of cyclic switchings (between key hidden states), showing controls near their saturation are synchronized with unknown disturbances. Interestingly, the found switching mechanism has appeared previously in the design of hybrid control for often-saturated systems. It is further interpreted via an analogy to the discrete-event subsystem in the hybrid control.

Tianming Wang, Wenjie Lu, Zheng Yan et al.

This paper presents an observer-integrated Reinforcement Learning (RL) approach, called Disturbance OBserver Network (DOB-Net), for robots operating in environments where disturbances are unknown and time-varying, and may frequently exceed robot control capabilities. The DOB-Net integrates a disturbance dynamics observer network and a controller network. Originated from conventional DOB mechanisms, the observer is built and enhanced via Recurrent Neural Networks (RNNs), encoding estimation of past values and prediction of future values of unknown disturbances in RNN hidden state. Such encoding allows the controller generate optimal control signals to actively reject disturbances, under the constraints of robot control capabilities. The observer and the controller are jointly learned within policy optimization by advantage actor critic. Numerical simulations on position regulation tasks have demonstrated that the proposed DOB-Net significantly outperforms a conventional feedback controller and classical RL algorithms.