Zhenglin Wan

Xin Yan, Aqiang Wang, Zhenglin Wan et al.

Diffusion large language models (DLLMs) have recently emerged as a promising alternative to autoregressive LLMs by generating text through iterative masked denoising with bidirectional context. However, their large model sizes and iterative denoising process introduce substantial memory and computational overhead, motivating post-training quantization for efficient deployment. In this paper, we identify two key challenges for low-bit DLLM quantization: state-dependent activation disparity and temporal error accumulation. Masked and unmasked tokens exhibit different activation distributions within each denoising step, while quantization errors can accumulate across steps during iterative decoding. To address these challenges, we propose STaR-Quant, a state-time consistent PTQ framework for DLLMs. STaR-Quant introduces State-Guided Activation Transformation (SGAT) to assign masked and unmasked tokens to different activation transformation spaces with a unified static weight-side transformation. It further introduces Temporal Attention Compensation (TAC) to correct the quantized attention representation via a lightweight block-diagonal affine mapping. Experiments on representative DLLMs demonstrate that STaR-Quant consistently improves low-bit weight-activation quantization over strong PTQ baselines, while delivering up to 1.69x speedup and 3.14x memory saving over FP16 deployment.

Tong Bai, Zhenglin Wan, Pengfei Zhou et al.

As LLM agents adopt large skill libraries, selecting the right subset becomes a structural problem rather than a similarity-matching one: skills depend on, conflict with, specialize, or duplicate one another, a structure invisible to both full enumeration and embedding similarity. We present SkillDAG, which models inter-skill relationships as a typed directed graph and exposes it to an LLM agent as an inference-time, agent-callable structural retrieval interface, queried and evolved during execution rather than baked into a fixed retrieval pipeline: each search returns vector matches, typed-edge neighbors, and conflict signals, and a propose-then-commit protocol lets the agent register execution-backed edges so the graph accumulates structure across episodes. On ALFWorld and SkillsBench with MiniMax-M2.7, SkillDAG reaches 67.1% success and 27.3% reward, exceeding the strongest reported Graph-of-Skills baseline by +12.8 and +8.6 points; the advantage ports to gpt-5.2-codex, and intrinsic SkillsBench Ret@K rises from 65.5 to 78.2 under matched queries. These gains trace to isolable mechanisms: candidate ranking that stays robust as the pool grows 10x where a fixed seeding-diffusion pipeline degrades, and set-monotone online edits that enlarge ground-truth recall without evicting prior hits.

Anjun Gao, Zhenglin Wan, Pingfu Chao et al.

The integration of GNSS data into portable devices has led to the generation of vast amounts of trajectory data, which is crucial for applications such as map-matching. To tackle the limitations of rule-based methods, recent works in deep learning for trajectory-related tasks occur. However, existing models remain challenging due to issues such as the difficulty of large-scale data labeling, ineffective modeling of spatial-temporal relationships, and discrepancies between training and test data distributions. To tackle these challenges, we propose HSTGMatch, a novel model designed to enhance map-matching performance. Our approach involves a two-stage process: hierarchical self-supervised learning and spatial-temporal supervised learning. We introduce a hierarchical trajectory representation, leveraging both grid cells and geographic tuples to capture moving patterns effectively. The model constructs an Adaptive Trajectory Adjacency Graph to dynamically capture spatial relationships, optimizing GATs for improved efficiency. Furthermore, we incorporate a Spatial-Temporal Factor to extract relevant features and employ a decay coefficient to address variations in trajectory length. Our extensive experiments demonstrate the model's superior performance, module effectiveness, and robustness, providing a promising solution for overcoming the existing limitations in map-matching applications. The source code of HSTGMatch is publicly available on GitHub at https://github.com/Nerooo-g/HSTGMatch.

Zhenglin Wan, Jingxuan Wu, Xingrui Yu et al.

Learning from demonstrations in embodied control is often cast as behavioral cloning, and recent diffusion or flow-matching policies improve this paradigm by modeling multi-modal expert actions. Yet these methods remain offline supervised learners: the policy is trained only on expert states and receives no corrective signal on the states it actually visits. On-policy distillation (OPD) offers a natural remedy, but standard OPD assumes a strong fixed teacher, which is unavailable in demonstration-only control. We propose \textbf{FA-OPD}, an \emph{adversarial dual on-policy distillation} method in which a Flow Matching (FM) teacher is learned from demonstrations and co-trained with a lightweight MLP student. The teacher provides two complementary signals on student rollouts. The reward channel learns an expert-likeness objective over state-action pairs and drives online exploration through long-horizon policy optimization. The action channel supplies dense local targets at student-visited states, stabilizing exploitation. FA-OPD couples them so that reward distillation enables generalization beyond point-wise demonstrations, while action distillation keeps exploration anchored near expert-like behavior. Across six robot navigation, manipulation, and locomotion benchmarks, FA-OPD beats strong baselines and shows much stronger robustness under noisy or limited demonstrations.

Ming Sun, Rui Wang, Xingrui Yu et al.

Vision-Language-Action models (VLAs) support generalist robotic control by enabling end-to-end decision policies directly from multi-modal inputs. As trained VLAs are increasingly shared and adapted, protecting model ownership becomes essential for secure deployment and responsible open-source usage. In this paper, we present GuardVLA, the first backdoor-based ownership verification framework specifically designed for VLAs. GuardVLA embeds a stealthy and harmless backdoor watermark into the protected model during training by injecting secret messages into embodied visual data. For post-release verification, we propose a swap-and-detect mechanism, in which the trigger projector and an external classifier head are used to activate and detect the embedded backdoor based on prediction probabilities. Extensive experiments across multiple datasets, model architectures, and adaptation settings demonstrate that GuardVLA enables reliable ownership verification while preserving benign task performance. Further results show that the embedded watermark remains detectable under post-release model adaptation.

Cooper Lin, Yanting Zhang, Maohao Ran et al.

E-commerce platforms and payment solution providers face increasingly sophisticated fraud schemes, ranging from identity theft and account takeovers to complex money laundering operations that exploit the speed and anonymity of digital transactions. However, despite their theoretical promise, the application of Large Language Models (LLMs) to fraud detection in real-world financial contexts remains largely unexploited, and their practical effectiveness in handling domain-specific e-commerce transaction data has yet to be empirically validated. To bridge this gap between conventional machine learning limitations and the untapped potential of LLMs in fraud detection, this paper proposes a novel approach that employs Reinforcement Learning (RL) to post-train lightweight language models specifically for fraud detection tasks using only raw transaction data. We utilize the Group Sequence Policy Optimization (GSPO) algorithm combined with a rule-based reward system to fine-tune language models of various sizes on a real-life transaction dataset provided by a Chinese global payment solution company. Through this reinforcement learning framework, the language models are encouraged to explore diverse trust and risk signals embedded within the textual transaction data, including patterns in customer information, shipping details, product descriptions, and order history. Our experimental results demonstrate the effectiveness of this approach, with post-trained language models achieving substantial F1-score improvements on held-out test data. Our findings demonstrate that the observed performance improvements are primarily attributable to the exploration mechanism inherent in reinforcement learning, which allows models to discover novel fraud indicators beyond those captured by traditional engineered features.

Cooper Lin, Maohao Ran, Yanting Zhang et al.

Standard safety alignment optimizes Large Language Models (LLMs) for universal helpfulness and honesty, effectively instilling a rigid "Boy Scout" morality. While robust for general-purpose assistants, this one-size-fits-all ethical framework imposes a "transparency tax" on professional domains requiring strategic ambiguity and information withholding, such as public relations, negotiation, and crisis management. To measure this gap between general safety and professional utility, we introduce Crisis-Bench, a multi-agent Partially Observable Markov Decision Process (POMDP) that evaluates LLMs in high-stakes corporate crises. Spanning 80 diverse storylines across 8 industries, Crisis-Bench tasks an LLM-based Public Relations (PR) Agent with navigating a dynamic 7-day corporate crisis simulation while managing strictly separated Private and Public narrative states to enforce rigorous information asymmetry. Unlike traditional benchmarks that rely on static ground truths, we introduce the Adjudicator-Market Loop: a novel evaluation metric where public sentiment is adjudicated and translated into a simulated stock price, creating a realistic economic incentive structure. Our results expose a critical dichotomy: while some models capitulate to ethical concerns, others demonstrate the capacity for Machiavellian, legitimate strategic withholding in order to stabilize the simulated stock price. Crisis-Bench provides the first quantitative framework for assessing "Reputation Management" capabilities, arguing for a shift from rigid moral absolutism to context-aware professional alignment.

Lang Feng, Fuchao Yang, Feng Chen et al.

Recent advances in large language models (LLMs) enable agentic systems trained with reinforcement learning (RL) over multi-turn interaction trajectories, but practical deployment is bottlenecked by rapidly growing textual histories that inflate token budgets and memory usage. We introduce AgentOCR, a framework that exploits the superior information density of visual tokens by representing the accumulated observation-action history as a compact rendered image. To make multi-turn rollouts scalable, AgentOCR proposes segment optical caching. By decomposing history into hashable segments and maintaining a visual cache, this mechanism eliminates redundant re-rendering. Beyond fixed rendering, AgentOCR introduces agentic self-compression, where the agent actively emits a compression rate and is trained with compression-aware reward to adaptively balance task success and token efficiency. We conduct extensive experiments on challenging agentic benchmarks, ALFWorld and search-based QA. Remarkably, results demonstrate that AgentOCR preserves over 95\% of text-based agent performance while substantially reducing token consumption (>50\%), yielding consistent token and memory efficiency. Our further analysis validates a 20x rendering speedup from segment optical caching and the effective strategic balancing of self-compression.

Jingxuan Wu, Zhenglin Wan, Xingrui Yu et al.

Diffusion language models (Diffusion-LMs) introduce an explicit temporal dimension into text generation, yet how this structure can be leveraged to control generation diversity for exploring multiple valid semantic or reasoning paths remains underexplored. In this paper, we show that Diffusion-LMs, like diffusion models in image generation, exhibit a temporal division of labor: early denoising steps largely determine the global semantic structure, while later steps focus on local lexical refinement. Building on this insight, we propose Time-Annealed Perturbation Sampling (TAPS), a training-free inference strategy that encourages semantic branching early in the diffusion process while progressively reducing perturbations to preserve fluency and instruction adherence. TAPS is compatible with both non-autoregressive and semi-autoregressive Diffusion backbones, demonstrated on LLaDA and TraDo in our paper, and consistently improves output diversity across creative writing and reasoning benchmarks without compromising generation quality.

Anjun Gao, Feng Wang, Zhenglin Wan et al.

Federated learning (FL) enables multiple clients to collaboratively train a global machine learning model via a server without sharing their private training data. In traditional FL, the system follows a synchronous approach, where the server waits for model updates from numerous clients before aggregating them to update the global model. However, synchronous FL is hindered by the straggler problem. To address this, the asynchronous FL architecture allows the server to update the global model immediately upon receiving any client's local model update. Despite its advantages, the decentralized nature of asynchronous FL makes it vulnerable to poisoning attacks. Several defenses tailored for asynchronous FL have been proposed, but these mechanisms remain susceptible to advanced attacks or rely on unrealistic server assumptions. In this paper, we introduce SecureAFL, an innovative framework designed to secure asynchronous FL against poisoning attacks. SecureAFL improves the robustness of asynchronous FL by detecting and discarding anomalous updates while estimating the contributions of missing clients. Additionally, it utilizes Byzantine-robust aggregation techniques, such as coordinate-wise median, to integrate the received and estimated updates. Extensive experiments on various real-world datasets demonstrate the effectiveness of SecureAFL.

Fanqing Meng, Lingxiao Du, Zijian Wu et al.

Language-model agents are increasingly used as persistent coworkers that assist users across multiple working days. During such workflows, the surrounding environment may change independently of the agent: new emails arrive, calendar entries shift, knowledge-base records are updated, and evidence appears across images, scanned PDFs, audio, video, and spreadsheets. Existing benchmarks do not adequately evaluate this setting because they typically run within a single static episode and remain largely text-centric. We introduce \bench{}, a benchmark for coworker agents built around multi-turn multi-day tasks, a stateful sandboxed service environment whose state evolves between turns, and rule-based verification. The current release contains 100 tasks across 13 professional scenarios, executed against five stateful sandboxed services (filesystem, email, calendar, knowledge base, spreadsheet) and scored by 1537 deterministic Python checkers over post-execution service state; no LLM-as-judge is invoked during scoring. We benchmark seven frontier agent systems. The strongest model reaches 75.8 weighted score, but the best strict Task Success is only 20.0\%, indicating that partial progress is common while complete end-to-end workflow completion remains rare. Turn-level analysis shows that performance drops after the first exogenous environment update, highlighting adaptation to changing state as a key open challenge. We release the benchmark, evaluation harness, and construction pipeline to support reproducible coworker-agent evaluation.

Yiheng Zhang, Yiming Wang, Kaiyan Zhao et al.

Proximal Policy Optimization (PPO) dominates deep RL but faces a fundamental dilemma. Its "hard clipping" mechanism discards valuable gradient information from outliers, leading to sample inefficiency. Conversely, removing clipping (as in SPO) exposes optimization to unbounded gradients, causing significant instability and hyperparameter sensitivity. To resolve this, we establish a Unified Trust Region Framework that generalizes existing objectives. Within this framework, we derive Anchored Neighborhood Optimization (ANO) based on a set of design principles. We identify that the failure of standard policy gradients stems from a misapplication of gradient influence on outliers. We propose the Redescending Influence Principle, a paradigm shift from monotonic penalties (SPO) and hard-thresholding (PPO) to dynamic outlier suppression, and prove its necessity for stability in high-variance stochastic optimization. Theoretically, we prove ANO possesses the minimal structural complexity required for robust optimization. Empirically, ANO achieves state-of-the-art performance on MuJoCo benchmarks, significantly outperforming PPO and SPO. Notably, ANO demonstrates superior stability, preventing policy collapse even under aggressive hyperparameters (e.g., learning rates 3x larger than standard) where PPO fails completely.

Chubin Zhang, Zhenglin Wan, Feng Chen et al.

Reinforcement learning (RL) faces a persistent tension: policies that are stable to optimize are often too simple to represent the multimodal action distributions needed for complex control. Gaussian policies provide tractable likelihoods and smooth gradients, but their unimodal form limits expressiveness. Conversely, generative policies based on diffusion or flow matching can model rich multimodal behaviors; however, in online RL, they are frequently unstable due to intractable likelihoods and noisy gradients propagating through deep sampling chains. We address this tension with a key structural principle: decoupling optimization from generation. Building on this insight, we introduce GoRL (Generative Online Reinforcement Learning), a framework that optimizes a tractable latent policy while utilizing a conditional generative decoder to synthesize actions. A two-timescale update schedule enables the latent policy to learn stably while the decoder steadily increases expressiveness, without requiring tractable action likelihoods. Across a range of continuous-control tasks, GoRL consistently outperforms both Gaussian policies and recent generative-policy baselines. Notably, on the HopperStand task, it reaches a normalized return above 870, more than 3 times that of the strongest baseline. These results demonstrate that separating optimization from generation provides a practical path to policies that are both stable and highly expressive.

Maohao Ran, Zhenglin Wan, Cooper Lin et al.

LLM-based agents are increasingly capable of complex task execution, yet current agentic systems remain constrained by text-centric paradigms. Traditional approaches rely on procedural JSON-based function calling, which often struggles with long-horizon tasks due to fragile multi-turn dependencies and context drift. In this paper, we present CaveAgent, a framework that transforms the paradigm from "LLM-as-Text-Generator" to "LLM-as-Runtime-Operator." We introduce a Dual-stream Context Architecture that decouples state management into a lightweight semantic stream for reasoning and a persistent, deterministic Python Runtime stream for execution. In addition to leveraging code generation to efficiently resolve interdependent sub-tasks (e.g., loops, conditionals) in a single step, we introduce \textit{Stateful Runtime Management} in CaveAgent. Distinct from existing code-based approaches that remain text-bound and lack the support for external object injection and retrieval, CaveAgent injects, manipulates, and retrieves complex Python objects (e.g., DataFrames, database connections) that persist across turns. This persistence mechanism acts as a high-fidelity external memory to eliminate context drift, avoid catastrophic forgetting, while ensuring that processed data flows losslessly to downstream applications. Comprehensive evaluations on Tau$^2$-bench, BFCL and various case studies across representative SOTA LLMs demonstrate CaveAgent's superiority. Specifically, our framework achieves a 10.5\% success rate improvement on retail tasks and reduces total token consumption by 28.4\% in multi-turn scenarios. On data-intensive tasks, direct variable storage and retrieval reduces token consumption by 59\%, allowing CaveAgent to handle large-scale data that causes context overflow failures in both JSON-based and Code-based agents.

Xingrui Yu, Zhenglin Wan, David Mark Bossens et al.

Learning diverse and high-performance behaviors from a limited set of demonstrations is a grand challenge. Traditional imitation learning methods usually fail in this task because most of them are designed to learn one specific behavior even with multiple demonstrations. Therefore, novel techniques for \textit{quality diversity imitation learning}, which bridges the quality diversity optimization and imitation learning methods, are needed to solve the above challenge. This work introduces Wasserstein Quality Diversity Imitation Learning (WQDIL), which 1) improves the stability of imitation learning in the quality diversity setting with latent adversarial training based on a Wasserstein Auto-Encoder (WAE), and 2) mitigates a behavior-overfitting issue using a measure-conditioned reward function with a single-step archive exploration bonus. Empirically, our method significantly outperforms state-of-the-art IL methods, achieving near-expert or beyond-expert QD performance on the challenging continuous control tasks derived from MuJoCo environments.

Zhenglin Wan, Jingxuan Wu, Xingrui Yu et al.

Flow Matching (FM) has shown remarkable ability in modeling complex distributions and achieves strong performance in offline imitation learning for cloning expert behaviors. However, despite its behavioral cloning expressiveness, FM-based policies are inherently limited by their lack of environmental interaction and exploration. This leads to poor generalization in unseen scenarios beyond the expert demonstrations, underscoring the necessity of online interaction with environment. Unfortunately, optimizing FM policies via online interaction is challenging and inefficient due to instability in gradient computation and high inference costs. To address these issues, we propose to let a student policy with simple MLP structure explore the environment and be online updated via RL algorithm with a reward model. This reward model is associated with a teacher FM model, containing rich information of expert data distribution. Furthermore, the same teacher FM model is utilized to regularize the student policy's behavior to stabilize policy learning. Due to the student's simple architecture, we avoid the gradient instability of FM policies and enable efficient online exploration, while still leveraging the expressiveness of the teacher FM model. Extensive experiments show that our approach significantly enhances learning efficiency, generalization, and robustness, especially when learning from suboptimal expert data.

Jingxuan Wu, Zhenglin Wan, Xingrui Yu et al.

Flow-based text-to-image models follow deterministic trajectories, forcing users to repeatedly sample to discover diverse modes, which is a costly and inefficient process. We present a training-free, inference-time control mechanism that makes the flow itself diversity-aware. Our method simultaneously encourages lateral spread among trajectories via a feature-space objective and reintroduces uncertainty through a time-scheduled stochastic perturbation. Crucially, this perturbation is projected to be orthogonal to the generation flow, a geometric constraint that allows it to boost variation without degrading image details or prompt fidelity. Our procedure requires no retraining or modification to the base sampler and is compatible with common flow-matching solvers. Theoretically, our method is shown to monotonically increase a volume surrogate while, due to its geometric constraints, approximately preserving the marginal distribution. This provides a principled explanation for why generation quality is robustly maintained. Empirically, across multiple text-to-image settings under fixed sampling budgets, our method consistently improves diversity metrics such as the Vendi Score and Brisque over strong baselines, while upholding image quality and alignment.

Mingcong Lei, Honghao Cai, Zezhou Cui et al.

Embodied agents face persistent challenges in real-world environments, including partial observability, limited spatial reasoning, and high-latency multi-memory integration. We present RoboMemory, a brain-inspired framework that unifies Spatial, Temporal, Episodic, and Semantic memory under a parallelized architecture for efficient long-horizon planning and interactive environmental learning. A dynamic spatial knowledge graph (KG) ensures scalable and consistent memory updates, while a closed-loop planner with a critic module supports adaptive decision-making in dynamic settings. Experiments on EmbodiedBench show that RoboMemory, built on Qwen2.5-VL-72B-Ins, improves average success rates by 25% over its baseline and exceeds the closed-source state-of-the-art (SOTA) Gemini-1.5-Pro by 3%. Real-world trials further confirm its capacity for cumulative learning, with performance improving across repeated tasks. These results highlight RoboMemory as a scalable foundation for memory-augmented embodied intelligence, bridging the gap between cognitive neuroscience and robotic autonomy.