Yuhui Wang

Haozhe Liu, Mingchen Zhuge, Bing Li et al.

Recent work on deep reinforcement learning (DRL) has pointed out that algorithmic information about good policies can be extracted from offline data which lack explicit information about executed actions. For example, videos of humans or robots may convey a lot of implicit information about rewarding action sequences, but a DRL machine that wants to profit from watching such videos must first learn by itself to identify and recognize relevant states/actions/rewards. Without relying on ground-truth annotations, our new method called Deep State Identifier learns to predict returns from episodes encoded as videos. Then it uses a kind of mask-based sensitivity analysis to extract/identify important critical states. Extensive experiments showcase our method's potential for understanding and improving agent behavior. The source code and the generated datasets are available at https://github.com/AI-Initiative-KAUST/VideoRLCS.

Ming Zhang, Jiabao Zhuang, Wenqing Jing et al.

Deep Research Agents increasingly automate survey generation, yet whether they match human experts at retrieving essential papers and organizing them into expert-like taxonomies remains unclear. Existing benchmarks emphasize writing quality or citation correctness, while standard clustering metrics ignore hierarchical structure. We introduce TaxoBench, a benchmark of 72 highly-cited LLM surveys with expert-authored taxonomy trees and 3,815 papers mapped to paper categories. TaxoBench evaluates (1) retrieval via Recall/Precision/F1, and (2) organization at a leaf level (paper-to-category assignment) and a hierarchy level via novel metrics, namely Unordered Semantic Tree Edit Distance US-TED/US-NTED and Semantic Path Similarity Sem-Path. Two modes are supported: Deep Research (topic-only, end-to-end) and Bottom-Up (expert paper set provided, organization-only). To distinguish disagreement with a single expert reference from genuine model failure, we explicitly partition findings into capability-based (reference-free) and alignment-based (reference-dependent). Evaluating 7 Deep Research Agents and 12 frontier LLMs reveals a dual bottleneck: capability-side, the best agent retrieves only 20.92% of expert-cited papers, and 1,000 model taxonomies show 75.9% sibling overlap, 51.2% MECE violations, and 83.4% structural imbalance, all detectable without any reference; alignment-side, all 12 LLMs converge to Sem-Path 28--29%, well below 47--58% achieved by three independent human-annotator groups on the same paper sets. Our benchmark is publicly available at https://github.com/KongLongGeFDU/TaxoBench

Deyao Zhu, Yuhui Wang, Jürgen Schmidhuber et al.

Offline RL methods have been shown to reduce the need for environment interaction by training agents using offline collected episodes. However, these methods typically require action information to be logged during data collection, which can be difficult or even impossible in some practical cases. In this paper, we investigate the potential of using action-free offline datasets to improve online reinforcement learning, name this problem Reinforcement Learning with Action-Free Offline Pretraining (AFP-RL). We introduce Action-Free Guide (AF-Guide), a method that guides online training by extracting knowledge from action-free offline datasets. AF-Guide consists of an Action-Free Decision Transformer (AFDT) implementing a variant of Upside-Down Reinforcement Learning. It learns to plan the next states from the offline dataset, and a Guided Soft Actor-Critic (Guided SAC) that learns online with guidance from AFDT. Experimental results show that AF-Guide can improve sample efficiency and performance in online training thanks to the knowledge from the action-free offline dataset. Code is available at https://github.com/Vision-CAIR/AF-Guide.

Yashesh Gaur, Nick Kibre, Jian Xue et al.

Automatic Speech Recognition (ASR) systems typically yield output in lexical form. However, humans prefer a written form output. To bridge this gap, ASR systems usually employ Inverse Text Normalization (ITN). In previous works, Weighted Finite State Transducers (WFST) have been employed to do ITN. WFSTs are nicely suited to this task but their size and run-time costs can make deployment on embedded applications challenging. In this paper, we describe the development of an on-device ITN system that is streaming, lightweight & accurate. At the core of our system is a streaming transformer tagger, that tags lexical tokens from ASR. The tag informs which ITN category might be applied, if at all. Following that, we apply an ITN-category-specific WFST, only on the tagged text, to reliably perform the ITN conversion. We show that the proposed ITN solution performs equivalent to strong baselines, while being significantly smaller in size and retaining customization capabilities.

Ming Zhang, Qiyuan Peng, Yinxi Wei et al.

Evaluating large language models (LLMs) on natural-language logical reasoning is essential because rule-governed tasks require conclusions to follow strictly from stated premises. Many existing logical-reasoning benchmarks are generated by templating natural-language items from sampled formulas, provide only coarse or unaudited formal annotations, and are now quickly saturated by frontier reasoning models. We present LLMEval-Logic, a Chinese logical reasoning benchmark built from realistic situational scenarios. Its pipeline forward-authors and expert-audits natural-language items together with their reference formalizations, verifies annotated answers with Z3, constructs expert rubrics for natural-to-formal grading, and hardens selected items through a closed-loop adversarial workflow. The benchmark is released in two paired subsets: a 246-item Base subset shipped with 1,400 expert-developed rubric atoms, and a 190-item Hard subset with 938 multi-step sub-questions over closed model spaces. Evaluating 14 frontier LLMs on LLMEval-Logic reveals substantial gaps in current models: the best model reaches only 37.5% Hard Item Accuracy, and even with reference symbols the highest joint Z3+Rubric formalization score among evaluated models reaches only 60.16%. Our benchmark is publicly available at https://github.com/llmeval/LLMEval-Logic.

Yuhui Wang, Langran Deng, Weibo Wu et al.

A provably stable summation-by-parts simultaneous approximation term (SBP-SAT) finite-difference time-domain (FDTD) subgridding method without region split is proposed. By designing projection SBP operators tailored for embedded topological features and deriving the corresponding SAT boundary conditions, this approach guarantees long-time stability through discrete energy analysis. Unlike conventional SBP-SAT FDTD subgridding techniques that rely on aligned or multi-block configurations, the proposed method enables a direct coupling between an internal refined region and a single surrounding coarse-grid domain without introducing auxiliary blocks or causing domain fragmentation. Numerical results validate the efficiency, accuracy, and topological flexibility of the proposed method. Compared with existing multi-block SBP-SAT methods, this method effectively reduces computational complexity by minimizing SAT boundary conditions and improves calculation accuracy near grid interfaces.

Yuhui Wang, Zhixiong Yang, Ming Zhang et al.

In the maintenance of complex systems, fault trees are used to locate problems and provide targeted solutions. To enable fault trees stored as images to be directly processed by large language models, which can assist in tracking and analyzing malfunctions, we propose a novel textual representation of fault trees. Building on it, we construct a benchmark for multi-turn dialogue systems that emphasizes robust interaction in complex environments, evaluating a model's ability to assist in malfunction localization, which contains $3130$ entries and $40.75$ turns per entry on average. We train an end-to-end model to generate vague information to reflect user behavior and introduce long-range rollback and recovery procedures to simulate user error scenarios, enabling assessment of a model's integrated capabilities in task tracking and error recovery, and Gemini 2.5 pro archives the best performance.

Yuhui Wang, Yingxia Shao, Ang Li

In the era of big data, intellectual property-oriented scientific and technological resources show the trend of large data scale, high information density and low value density, which brings severe challenges to the effective use of intellectual property resources, and the demand for mining hidden information in intellectual property is increasing. This makes intellectual property-oriented science and technology resource portraits and analysis of evolution become the current research hotspot. This paper sorts out the construction method of intellectual property resource intellectual portrait and its pre-work property entity extraction and entity completion from the aspects of algorithm classification and general process, and directions for improvement of future methods.

Yuhui Wang, Junping Du, Yingxia Shao

Patent texts contain a large amount of entity information. Through named entity recognition, intellectual property entity information containing key information can be extracted from it, helping researchers to understand the patent content faster. Therefore, it is difficult for existing named entity extraction methods to make full use of the semantic information at the word level brought about by professional vocabulary changes. This paper proposes a method for extracting intellectual property entities based on Transformer and technical word information , and provides accurate word vector representation in combination with the BERT language method. In the process of word vector generation, the technical word information extracted by IDCNN is added to improve the understanding of intellectual property entities Representation ability. Finally, the Transformer encoder that introduces relative position encoding is used to learn the deep semantic information of the text from the sequence of word vectors, and realize entity label prediction. Experimental results on public datasets and annotated patent datasets show that the method improves the accuracy of entity recognition.

Dingwei Zhu, Zhiheng Xi, Shihan Dou et al.

Training reinforcement learning (RL) systems in real-world environments remains challenging due to noisy supervision and poor out-of-domain (OOD) generalization, especially in LLM post-training. Recent distributional RL methods improve robustness by modeling values with multiple quantile points, but they still learn each quantile independently as a scalar. This results in rough-grained value representations that lack fine-grained conditioning on state information, struggling under complex and OOD conditions. We propose DFPO (Distributional Value Flow Policy Optimization with Conditional Risk and Consistency Control), a robust distributional RL framework that models values as continuous flows across time steps. By scaling value modeling through learning of a value flow field instead of isolated quantile predictions, DFPO captures richer state information for more accurate advantage estimation. To stabilize training under noisy feedback, DFPO further integrates conditional risk control and consistency constraints along value flow trajectories. Experiments on dialogue, math reasoning, and scientific tasks show that DFPO outperforms PPO, FlowRL, and other robust baselines under noisy supervision, achieving improved training stability and generalization.

Xiangyu Peng, Zangwei Zheng, Chenhui Shen et al.

Video generation models have achieved remarkable progress in the past year. The quality of AI video continues to improve, but at the cost of larger model size, increased data quantity, and greater demand for training compute. In this report, we present Open-Sora 2.0, a commercial-level video generation model trained for only $200k. With this model, we demonstrate that the cost of training a top-performing video generation model is highly controllable. We detail all techniques that contribute to this efficiency breakthrough, including data curation, model architecture, training strategy, and system optimization. According to human evaluation results and VBench scores, Open-Sora 2.0 is comparable to global leading video generation models including the open-source HunyuanVideo and the closed-source Runway Gen-3 Alpha. By making Open-Sora 2.0 fully open-source, we aim to democratize access to advanced video generation technology, fostering broader innovation and creativity in content creation. All resources are publicly available at: https://github.com/hpcaitech/Open-Sora.

Dingwei Zhu, Zhiheng Xi, Shihan Dou et al.

Reinforcement learning (RL) has shown strong performance in LLM post-training, but real-world deployment often involves noisy or incomplete supervision. In such settings, complex and unreliable supervision signals can destabilize training and harm generalization. While existing approaches such as worst-case optimization (e.g., RFQI, CQL) and mean-based methods (e.g., PPO, GRPO) can improve stability, they often overlook generalization and may produce overly conservative policies, leading to uneven performance across diverse real scenarios. To this end, we introduce DVPO (Distributional Value Modeling with Risk-aware Policy Optimization), a new RL framework that combines conditional risk theory with distributional value modeling to better balance robustness and generalization. DVPO learns token-level value distributions to provide fine-grained supervision, and applies an asymmetric risk regularization to shape the distribution tails: it contracts the lower tail to dampen noisy negative deviations, while expanding the upper tail to preserve exploratory diversity. Across extensive experiments and analysis in multi-turn dialogue, math reasoning, and scientific QA, DVPO consistently outperforms PPO, GRPO, and robust Bellman-based PPO under noisy supervision, showing its potential for LLM post-training in the real-world.

Hongrui Zhang, Daiqing Wu, Yangyang Li et al.

Multimodal Emotion Recognition (MER) focuses on identifying and interpreting emotions from modality-compound inputs. Closely mirroring human cognitive processes in real-world environments, MER has drawn substantial attention from both academia and industry. Recently, a paradigm shift has been unveiled in MER, from leveraging small-scale, task-specific models to Large Language Models (LLMs). We refer to the latter as the MER-with-LLMs paradigm, which offers unprecedented generality, spurring numerous empirical attempts, even alongside speculation about LLMs' potential to achieve general emotional intelligence. However, with these new opportunities come new challenges, including the scarcity of emotionally annotated data, the affective gap both within and across modalities, and the opacity of affective interpretation. To systematically review existing research and guide future exploration, this paper categorizes prior works according to their focus on addressing these challenges into three directions: Affective Data Augmentation, Multimodal Affective Representation, and Multimodal Affective Reasoning. By thoroughly tracing the development, emerging trends, and remaining issues within each direction, this paper aims to provide a clear academic map of the MER-with-LLMs paradigm and foster its structured advancement.

Alfonso Amayuelas, Firas Laakom, Piotr Piękos et al.

The use of LLMs for code generation has naturally extended to code testing and evaluation. As codebases grow in size and complexity, so does the need for automated test generation. Current approaches for LLM-based test generation rely on strategies that maximize immediate coverage gain, a greedy approach that plateaus on code where reaching deep branches requires setup steps that individually yield zero new coverage. Drawing on principles of Bayesian exploration, we treat the program's branch structure as an unknown environment, and an evolving coverage map as a proxy probabilistic posterior representing what the LLM has discovered so far. Our method, CovQValue, feeds the coverage map back to the LLM, generates diverse candidate plans in parallel, and selects the most informative plan by LLM-estimated Q-values, seeking actions that balance immediate branch discovery with future reachability. Our method outperforms greedy selection on TestGenEval Lite, achieving 51-77% higher branch coverage across three popular LLMs and winning on 77-84% of targets. In addition, we build a benchmark for iterative test generation, RepoExploreBench, where they achieve 40-74%. These results show the potential of curiosity-driven planning methods for LLM-based exploration, enabling more effective discovery of program behavior through sequential interaction

Guangke Chen, Yuhui Wang, Shouling Ji et al.

Modern text-to-speech (TTS) systems, particularly those built on Large Audio-Language Models (LALMs), generate high-fidelity speech that faithfully reproduces input text and mimics specified speaker identities. While prior misuse studies have focused on speaker impersonation, this work explores a distinct content-centric threat: exploiting TTS systems to produce speech containing harmful content. Realizing such threats poses two core challenges: (1) LALM safety alignment frequently rejects harmful prompts, yet existing jailbreak attacks are ill-suited for TTS because these systems are designed to faithfully vocalize any input text, and (2) real-world deployment pipelines often employ input/output filters that block harmful text and audio. We present HARMGEN, a suite of five attacks organized into two families that address these challenges. The first family employs semantic obfuscation techniques (Concat, Shuffle) that conceal harmful content within text. The second leverages audio-modality exploits (Read, Spell, Phoneme) that inject harmful content through auxiliary audio channels while maintaining benign textual prompts. Through evaluation across five commercial LALMs-based TTS systems and three datasets spanning two languages, we demonstrate that our attacks substantially reduce refusal rates and increase the toxicity of generated speech. We further assess both reactive countermeasures deployed by audio-streaming platforms and proactive defenses implemented by TTS providers. Our analysis reveals critical vulnerabilities: deepfake detectors underperform on high-fidelity audio; reactive moderation can be circumvented by adversarial perturbations; while proactive moderation detects 57-93% of attacks. Our work highlights a previously underexplored content-centric misuse vector for TTS and underscore the need for robust cross-modal safeguards throughout training and deployment.

Jiacheng Liang, Tanqiu Jiang, Yuhui Wang et al.

This paper presents AutoRAN, the first framework to automate the hijacking of internal safety reasoning in large reasoning models (LRMs). At its core, AutoRAN pioneers an execution simulation paradigm that leverages a weaker but less-aligned model to simulate execution reasoning for initial hijacking attempts and iteratively refine attacks by exploiting reasoning patterns leaked through the target LRM's refusals. This approach steers the target model to bypass its own safety guardrails and elaborate on harmful instructions. We evaluate AutoRAN against state-of-the-art LRMs, including GPT-o3/o4-mini and Gemini-2.5-Flash, across multiple benchmarks (AdvBench, HarmBench, and StrongReject). Results show that AutoRAN achieves approaching 100% success rate within one or few turns, effectively neutralizing reasoning-based defenses even when evaluated by robustly aligned external models. This work reveals that the transparency of the reasoning process itself creates a critical and exploitable attack surface, highlighting the urgent need for new defenses that protect models' reasoning traces rather than merely their final outputs.

Ming Zhang, Yujiong Shen, Zelin Li et al.

Evaluating large language models (LLMs) in medicine is crucial because medical applications require high accuracy with little room for error. Current medical benchmarks have three main types: medical exam-based, comprehensive medical, and specialized assessments. However, these benchmarks have limitations in question design (mostly multiple-choice), data sources (often not derived from real clinical scenarios), and evaluation methods (poor assessment of complex reasoning). To address these issues, we present LLMEval-Med, a new benchmark covering five core medical areas, including 2,996 questions created from real-world electronic health records and expert-designed clinical scenarios. We also design an automated evaluation pipeline, incorporating expert-developed checklists into our LLM-as-Judge framework. Furthermore, our methodology validates machine scoring through human-machine agreement analysis, dynamically refining checklists and prompts based on expert feedback to ensure reliability. We evaluate 13 LLMs across three categories (specialized medical models, open-source models, and closed-source models) on LLMEval-Med, providing valuable insights for the safe and effective deployment of LLMs in medical domains. The dataset is released in https://github.com/llmeval/LLMEval-Med.

Qingyuan Wu, Simon Sinong Zhan, Yixuan Wang et al.

Reinforcement learning (RL) is challenging in the common case of delays between events and their sensory perceptions. State-of-the-art (SOTA) state augmentation techniques either suffer from state space explosion or performance degeneration in stochastic environments. To address these challenges, we present a novel Auxiliary-Delayed Reinforcement Learning (AD-RL) method that leverages auxiliary tasks involving short delays to accelerate RL with long delays, without compromising performance in stochastic environments. Specifically, AD-RL learns a value function for short delays and uses bootstrapping and policy improvement techniques to adjust it for long delays. We theoretically show that this can greatly reduce the sample complexity. On deterministic and stochastic benchmarks, our method significantly outperforms the SOTAs in both sample efficiency and policy performance. Code is available at https://github.com/QingyuanWuNothing/AD-RL.

Qingyuan Wu, Yuhui Wang, Simon Sinong Zhan et al.

Reinforcement Learning with Verified Reward (RLVR) has emerged as a critical paradigm for advancing the reasoning capabilities of Large Language Models (LLMs). Most existing RLVR methods, such as GRPO and its variants, ensure stable updates by constraining policy divergence through clipping likelihood ratios. This paper introduces a unified clipping framework that characterizes existing methods via a general notion of policy divergence, encompassing both likelihood ratios and Kullback-Leibler (KL) divergences and extending to alternative measures. The framework provides a principled foundation for systematically analyzing how different policy divergence measures affect exploration and performance. We further identify the KL3 estimator, a variance-reduced Monte Carlo estimator of the KL divergence, as a key policy divergence constraint. We theoretically demonstrate that the KL3-based constraint is mathematically equivalent to an asymmetric ratio-based clipping that reallocates probability mass toward high-confidence actions, promoting stronger exploration while retaining the simplicity of GRPO-style methods. Empirical results on mathematical reasoning benchmarks demonstrate that incorporating the KL3 estimator into GRPO improves both training stability and final performance, highlighting the importance of principled policy divergence constraints in policy optimization.

Yanning Dai, Yuhui Wang, Dylan R. Ashley et al.

Morphology-control co-design concerns the coupled optimization of an agent's body structure and control policy. This problem exhibits a bi-level structure, where the control dynamically adapts to the morphology to maximize performance. Existing methods typically neglect the control's adaptation dynamics by adopting a single-level formulation that treats the control policy as fixed when optimizing morphology. This can lead to inefficient optimization, as morphology updates may be misaligned with control adaptation. In this paper, we revisit the co-design problem from a game-theoretic perspective, modeling the intrinsic coupling between morphology and control as a novel variant of a Stackelberg game. We propose Stackelberg Proximal Policy Optimization (Stackelberg PPO), which explicitly incorporates the control's adaptation dynamics into morphology optimization. By modeling this intrinsic coupling, our method aligns morphology updates with control adaptation, thereby stabilizing training and improving learning efficiency. Experiments across diverse co-design tasks demonstrate that Stackelberg PPO outperforms standard PPO in both stability and final performance, opening the way for dramatically more efficient robotics designs.

Qingyuan Wu, Yuhui Wang, Simon Sinong Zhan et al.

Reinforcement learning (RL) with delays is challenging as sensory perceptions lag behind the actual events: the RL agent needs to estimate the real state of its environment based on past observations. State-of-the-art (SOTA) methods typically employ recursive, step-by-step forecasting of states. This can cause the accumulation of compounding errors. To tackle this problem, our novel belief estimation method, named Directly Forecasting Belief Transformer (DFBT), directly forecasts states from observations without incrementally estimating intermediate states step-by-step. We theoretically demonstrate that DFBT greatly reduces compounding errors of existing recursively forecasting methods, yielding stronger performance guarantees. In experiments with D4RL offline datasets, DFBT reduces compounding errors with remarkable prediction accuracy. DFBT's capability to forecast state sequences also facilitates multi-step bootstrapping, thus greatly improving learning efficiency. On the MuJoCo benchmark, our DFBT-based method substantially outperforms SOTA baselines. Code is available at https://github.com/QingyuanWuNothing/DFBT.

Ming Zhang, Yuhui Wang, Yujiong Shen et al.

Process-driven dialogue systems, which operate under strict predefined process constraints, are essential in customer service and equipment maintenance scenarios. Although Large Language Models (LLMs) have shown remarkable progress in dialogue and reasoning, they still struggle to solve these strictly constrained dialogue tasks. To address this challenge, we construct Process Flow Dialogue (PFDial) dataset, which contains 12,705 high-quality Chinese dialogue instructions derived from 440 flowcharts containing 5,055 process nodes. Based on PlantUML specification, each UML flowchart is converted into atomic dialogue units i.e., structured five-tuples. Experimental results demonstrate that a 7B model trained with merely 800 samples, and a 0.5B model trained on total data both can surpass 90% accuracy. Additionally, the 8B model can surpass GPT-4o up to 43.88% with an average of 11.00%. We further evaluate models' performance on challenging backward transitions in process flows and conduct an in-depth analysis of various dataset formats to reveal their impact on model performance in handling decision and sequential branches. The data is released in https://github.com/KongLongGeFDU/PFDial.

Joya Chen, Kai Xu, Yuhui Wang et al.

A standard hardware bottleneck when training deep neural networks is GPU memory. The bulk of memory is occupied by caching intermediate tensors for gradient computation in the backward pass. We propose a novel method to reduce this footprint - Dropping Intermediate Tensors (DropIT). DropIT drops min-k elements of the intermediate tensors and approximates gradients from the sparsified tensors in the backward pass. Theoretically, DropIT reduces noise on estimated gradients and therefore has a higher rate of convergence than vanilla-SGD. Experiments show that we can drop up to 90\% of the intermediate tensor elements in fully-connected and convolutional layers while achieving higher testing accuracy for Visual Transformers and Convolutional Neural Networks on various tasks (e.g., classification, object detection, instance segmentation). Our code and models are available at https://github.com/chenjoya/dropit.

Tanqiu Jiang, Yuhui Wang, Jiacheng Liang et al.

LLM agents are increasingly deployed in long-horizon, complex environments to solve challenging problems, but this expansion exposes them to long-horizon attacks that exploit multi-turn user-agent-environment interactions to achieve objectives infeasible in single-turn settings. To measure agent vulnerabilities to such risks, we present AgentLAB, the first benchmark dedicated to evaluating LLM agent susceptibility to adaptive, long-horizon attacks. Currently, AgentLAB supports five novel attack types including intent hijacking, tool chaining, task injection, objective drifting, and memory poisoning, spanning 28 realistic agentic environments, and 644 security test cases. Leveraging AgentLAB, we evaluate representative LLM agents and find that they remain highly susceptible to long-horizon attacks; moreover, defenses designed for single-turn interactions fail to reliably mitigate long-horizon threats. We anticipate that AgentLAB will serve as a valuable benchmark for tracking progress on securing LLM agents in practical settings. The benchmark is publicly available at https://tanqiujiang.github.io/AgentLAB_main.

Jiacheng Liang, Yuhui Wang, Tanqiu Jiang et al.

Mixture-of-Experts (MoE) language models introduce unique challenges for safety alignment due to their sparse routing mechanisms, which can enable degenerate optimization behaviors under standard full-parameter fine-tuning. In our preliminary experiments, we observe that naively applying full-parameter safety fine-tuning to MoE models can reduce attack success rates through routing or expert dominance effects, rather than by directly repairing Safety-Critical Experts. To address this challenge, we propose RASA, a routing-aware expert-level alignment framework that explicitly repairs Safety-Critical Experts while preventing routing-based bypasses. RASA identifies experts disproportionately activated by successful jailbreaks, selectively fine-tunes only these experts under fixed routing, and subsequently enforces routing consistency with safety-aligned contexts. Across two representative MoE architectures and a diverse set of jailbreak attacks, RASA achieves near-perfect robustness, strong cross-attack generalization, and substantially reduced over-refusal, while preserving general capabilities on benchmarks such as MMLU, GSM8K, and TruthfulQA. Our results suggest that robust MoE safety alignment benefits from targeted expert repair rather than global parameter updates, offering a practical and architecture-preserving alternative to prior approaches.

Yuhui Wang, Tanqiu Jiang, Jiacheng Liang et al.

As large language model (LLM)-powered agents are increasingly deployed to perform complex, real-world tasks, they face a growing class of attacks that exploit extended user-agent-environment interactions to pursue malicious objectives improbable in single-turn settings. Such long-horizon threats pose significant risks to the safe deployment of LLM agents in critical domains. In this paper, we present MAGE (Memory As Guardrail Enforcement), a novel defensive framework designed to counter a wide range of long-horizon threats. Inspired by the "shadow stack" abstraction in systems security, MAGE maintains a dedicated, safety-focused agentic memory that distills and retains safety-critical context across the agent's full execution trajectory, leveraging this shadow memory to proactively assess the risk of pending actions prior to their execution. Extensive evaluation demonstrates that MAGE substantially outperforms existing defenses across diverse long-horizon threats in detection accuracy, achieves early-stage detection for the majority of attacks, and introduces only negligible overhead to agent utility. To our best knowledge, MAGE represents the first framework to detect and mitigate long-horizon threats using an agentic memory approach, establishing a new paradigm for this critical challenge and opening promising directions for future research.

Tanqiu Jiang, Zian Wang, Jiacheng Liang et al.

Jailbreak attacks circumvent LLMs' built-in safeguards by concealing harmful queries within jailbreak prompts. While existing defenses primarily focus on mitigating the effects of jailbreak prompts, they often prove inadequate as jailbreak prompts can take arbitrary, adaptive forms. This paper presents RobustKV, a novel defense that adopts a fundamentally different approach by selectively removing critical tokens of harmful queries from key-value (KV) caches. Intuitively, for a jailbreak prompt to be effective, its tokens must achieve sufficient `importance' (as measured by attention scores), which inevitably lowers the importance of tokens in the concealed harmful query. Thus, by strategically evicting the KVs of the lowest-ranked tokens, RobustKV diminishes the presence of the harmful query in the KV cache, thus preventing the LLM from generating malicious responses. Extensive evaluation using benchmark datasets and models demonstrates that RobustKV effectively counters state-of-the-art jailbreak attacks while maintaining the LLM's general performance on benign queries. Moreover, RobustKV creates an intriguing evasiveness dilemma for adversaries, forcing them to balance between evading RobustKV and bypassing the LLM's built-in safeguards. This trade-off contributes to RobustKV's robustness against adaptive attacks. (warning: this paper contains potentially harmful content generated by LLMs.)

Yuhui Wang, Junaid Farooq

The advent of fifth generation (5G) networks has opened new avenues for enhancing connectivity, particularly in challenging environments like remote areas or disaster-struck regions. Unmanned aerial vehicles (UAVs) have been identified as a versatile tool in this context, particularly for improving network performance through the Integrated access and backhaul (IAB) feature of 5G. However, existing approaches to UAV-assisted network enhancement face limitations in dynamically adapting to varying user locations and network demands. This paper introduces a novel approach leveraging deep reinforcement learning (DRL) to optimize UAV placement in real-time, dynamically adjusting to changing network conditions and user requirements. Our method focuses on the intricate balance between fronthaul and backhaul links, a critical aspect often overlooked in current solutions. The unique contribution of this work lies in its ability to autonomously position UAVs in a way that not only ensures robust connectivity to ground users but also maintains seamless integration with central network infrastructure. Through various simulated scenarios, we demonstrate how our approach effectively addresses these challenges, enhancing coverage and network performance in critical areas. This research fills a significant gap in UAV-assisted 5G networks, providing a scalable and adaptive solution for future mobile networks.

Qingyuan Wu, Simon Sinong Zhan, Yixuan Wang et al.

In environments with delayed observation, state augmentation by including actions within the delay window is adopted to retrieve Markovian property to enable reinforcement learning (RL). However, state-of-the-art (SOTA) RL techniques with Temporal-Difference (TD) learning frameworks often suffer from learning inefficiency, due to the significant expansion of the augmented state space with the delay. To improve learning efficiency without sacrificing performance, this work introduces a novel framework called Variational Delayed Policy Optimization (VDPO), which reformulates delayed RL as a variational inference problem. This problem is further modelled as a two-step iterative optimization problem, where the first step is TD learning in the delay-free environment with a small state space, and the second step is behaviour cloning which can be addressed much more efficiently than TD learning. We not only provide a theoretical analysis of VDPO in terms of sample complexity and performance, but also empirically demonstrate that VDPO can achieve consistent performance with SOTA methods, with a significant enhancement of sample efficiency (approximately 50\% less amount of samples) in the MuJoCo benchmark.

Yuhui Wang, Rongyi Zhu, Ting Wang

Harmful fine-tuning attacks pose a major threat to the security of large language models (LLMs), allowing adversaries to compromise safety guardrails with minimal harmful data. While existing defenses attempt to reinforce LLM alignment, they fail to address models' inherent "trainability" on harmful data, leaving them vulnerable to stronger attacks with increased learning rates or larger harmful datasets. To overcome this critical limitation, we introduce SEAM, a novel alignment-enhancing defense that transforms LLMs into self-destructive models with intrinsic resilience to misalignment attempts. Specifically, these models retain their capabilities for legitimate tasks while exhibiting substantial performance degradation when fine-tuned on harmful data. The protection is achieved through a novel loss function that couples the optimization trajectories of benign and harmful data, enhanced with adversarial gradient ascent to amplify the self-destructive effect. To enable practical training, we develop an efficient Hessian-free gradient estimate with theoretical error bounds. Extensive evaluation across LLMs and datasets demonstrates that SEAM creates a no-win situation for adversaries: the self-destructive models achieve state-of-the-art robustness against low-intensity attacks and undergo catastrophic performance collapse under high-intensity attacks, rendering them effectively unusable. (warning: this paper contains potentially harmful content generated by LLMs.)

Jiacheng Liang, Yuhui Wang, Changjiang Li et al.

GraphRAG advances retrieval-augmented generation (RAG) by structuring external knowledge as multi-scale knowledge graphs, enabling language models to integrate both broad context and granular details in their generation. While GraphRAG has demonstrated success across domains, its security implications remain largely unexplored. To bridge this gap, this work examines GraphRAG's vulnerability to poisoning attacks, uncovering an intriguing security paradox: existing RAG poisoning attacks are less effective under GraphRAG than conventional RAG, due to GraphRAG's graph-based indexing and retrieval; yet, the same features also create new attack surfaces. We present GragPoison, a novel attack that exploits shared relations in the underlying knowledge graph to craft poisoning text capable of compromising multiple queries simultaneously. GragPoison employs three key strategies: (i) relation injection to introduce false knowledge, (ii) relation enhancement to amplify poisoning influence, and (iii) narrative generation to embed malicious content within coherent text. Empirical evaluation across diverse datasets and models shows that GragPoison substantially outperforms existing attacks in terms of effectiveness (up to 98% success rate) and scalability (using less than 68% poisoning text) on multiple variations of GraphRAG. We also explore potential defensive measures and their limitations, identifying promising directions for future research.

Zhiying Wang, Gang Sun, Yuhui Wang et al.

The Space-Air-Ground Integrated Network (SAGIN) framework is a crucial foundation for future networks, where satellites and aerial nodes assist in computational task offloading. The low-altitude economy, leveraging the flexibility and multifunctionality of Unmanned Aerial Vehicles (UAVs) in SAGIN, holds significant potential for development in areas such as communication and sensing. However, effective coordination is needed to streamline information exchange and enable efficient system resource allocation. In this paper, we propose a Clustering-based Multi-agent Deep Deterministic Policy Gradient (CMADDPG) algorithm to address the multi-UAV cooperative task scheduling challenges in SAGIN. The CMADDPG algorithm leverages dynamic UAV clustering to partition UAVs into clusters, each managed by a Cluster Head (CH) UAV, facilitating a distributed-centralized control approach. Within each cluster, UAVs delegate offloading decisions to the CH UAV, reducing intra-cluster communication costs and decision conflicts, thereby enhancing task scheduling efficiency. Additionally, by employing a multi-agent reinforcement learning framework, the algorithm leverages the extensive coverage of satellites to achieve centralized training and distributed execution of multi-agent tasks, while maximizing overall system profit through optimized task offloading decision-making. Simulation results reveal that the CMADDPG algorithm effectively optimizes resource allocation, minimizes queue delays, maintains balanced load distribution, and surpasses existing methods by achieving at least a 25\% improvement in system profit, showcasing its robustness and adaptability across diverse scenarios.

Hongjun Liu, Yinghao Zhu, Yuhui Wang et al.

Recent progress in multimodal large language models (MLLMs) has demonstrated promising performance on medical benchmarks and in preliminary trials as clinical assistants. Yet, our pilot audit of diagnostic cases uncovers a critical failure mode: instability in early evidence interpretation precedes hallucination, creating branching reasoning trajectories that cascade into globally inconsistent conclusions. This highlights the need for clinical reasoning agents that constrain stochasticity and hallucination while producing auditable decision flows. We introduce MedMMV, a controllable multimodal multi-agent framework for reliable and verifiable clinical reasoning. MedMMV stabilizes reasoning through diversified short rollouts, grounds intermediate steps in a structured evidence graph under the supervision of a Hallucination Detector, and aggregates candidate paths with a Combined Uncertainty scorer. On six medical benchmarks, MedMMV improves accuracy by up to 12.7% and, more critically, demonstrates superior reliability. Blind physician evaluations confirm that MedMMV substantially increases reasoning truthfulness without sacrificing informational content. By controlling instability through a verifiable, multi-agent process, our framework provides a robust path toward deploying trustworthy AI systems in high-stakes domains like clinical decision support.

Yuhui Wang, Changjiang Li, Guangke Chen et al.

Large reasoning models (LRMs) exhibit unprecedented capabilities in solving complex problems through Chain-of-Thought (CoT) reasoning. However, recent studies reveal that their final answers often contradict their own reasoning traces. We hypothesize that this inconsistency stems from two competing mechanisms for generating answers: CoT reasoning and memory retrieval. To test this hypothesis, we conduct controlled experiments that challenge LRMs with misleading cues during reasoning and/or corrupted answers during retrieval. Our results across models and datasets confirm that both mechanisms operate simultaneously, with their relative dominance influenced by multiple factors: problem domains, model scales, and fine-tuning approaches (e.g., reinforcement learning vs. distillation). The findings reveal a critical limitation in current reasoning fine-tuning paradigms: models can exploit the retrieval mechanism as a shortcut, effectively "hacking" the reward signal and undermining genuine reasoning development. To address this challenge, we introduce FARL, a novel fine-tuning framework that integrates memory unlearning with reinforcement learning. By carefully suppressing retrieval shortcuts during the fine-tuning process, FARL promotes reasoning-dominant behavior and enhances generalizable reasoning capabilities.

Rongyi Zhu, Yuhui Wang, Tanqiu Jiang et al.

Model steering represents a powerful technique that dynamically aligns large language models (LLMs) with human preferences during inference. However, conventional model-steering methods rely heavily on externally annotated data, not only limiting their adaptability to varying contexts but also tethering their effectiveness to annotation quality. In this paper, we present SIMS, the first self-improving model-steering framework that operates without relying on external supervision. At its core, SIMS autonomously generates and refines contrastive samples through iterative self-improvement cycles, enabling adaptive, context-specific steering. Additionally, SIMS employs novel strategies, including prompt ranking and contrast sampling, to further enhance steering efficacy. Extensive evaluation across diverse LLMs and benchmarks demonstrates that SIMS substantially outperforms existing methods in steering effectiveness and adaptability, highlighting self-improving model steering as a promising direction for future research on inference-time LLM alignment.

Yanning Dai, Chenyu Tang, Ruizhi Zhang et al.

Dynamic prediction of locomotor capacity after stroke could enable more individualized rehabilitation, yet current assessments largely provide static impairment scores and do not indicate whether patients can perform specific tasks such as slope walking or stair climbing. Here, we present a wearable-informed data-physics hybrid generative framework that reconstructs a stroke survivor's locomotor control from wearable inertial sensing and predicts task-conditioned post-stroke locomotion in new environments. From a single 20 m level-ground walking trial recorded by five IMUs, the framework personalizes a physics-based digital avatar using a healthy-motion prior and hybrid imitation learning, generating dynamically feasible, patient-specific movements for inclined walking and stair negotiation. Across 11 stroke inpatients, predicted postures reached 82.2% similarity for slopes and 69.9% for stairs, substantially exceeding a physics-only baseline. In a multicentre pilot randomized study (n = 21; 28 days), access to scenario-specific locomotion predictions to support task selection and difficulty titration was associated with larger gains in Fugl-Meyer lower-extremity scores than standard care (mean change 6.0 vs 3.7 points; $p < 0.05$). These results suggest that wearable-informed generative digital avatars may augment individualized gait rehabilitation planning and provide a pathway toward dynamically personalized post-stroke motor recovery strategies.

Ming Zhang, Kexin Tan, Yueyuan Huang et al.

Evaluating novelty is critical yet challenging in peer review, as reviewers must assess submissions against a vast, rapidly evolving literature. This report presents OpenNovelty, an LLM-powered agentic system for transparent, evidence-based novelty analysis. The system operates through four phases: (1) extracting the core task and contribution claims to generate retrieval queries; (2) retrieving relevant prior work based on extracted queries via semantic search engine; (3) constructing a hierarchical taxonomy of core-task-related work and performing contribution-level full-text comparisons against each contribution; and (4) synthesizing all analyses into a structured novelty report with explicit citations and evidence snippets. Unlike naive LLM-based approaches, \textsc{OpenNovelty} grounds all assessments in retrieved real papers, ensuring verifiable judgments. We deploy our system on 500+ ICLR 2026 submissions with all reports publicly available on our website, and preliminary analysis suggests it can identify relevant prior work, including closely related papers that authors may overlook. OpenNovelty aims to empower the research community with a scalable tool that promotes fair, consistent, and evidence-backed peer review.

Senjie Jin, Lu Chen, Zhiheng Xi et al.

Natural language chain-of-thought (N-CoT) and Program chain-of-thought (P-CoT) have emerged as two primary paradigms for large language models (LLMs) to solve mathematical reasoning problems. Current research typically endeavors to achieve unidirectional enhancement: P-CoT enhanced N-CoT or N-CoT enhanced P-CoT. In this paper, we seek to fully unleash the two paradigms' strengths for mutual enhancement and ultimately achieve simultaneous improvements. We conduct a detailed analysis of the error types across two paradigms, based on which we propose Parrot, a novel training pipeline for mathematical problems: 1) Three target-designed subtasks integrate sequential P-CoT and N-CoT generation. 2) A subtask hybrid training strategy to facilitate natural language semantic transferability. 3) The converted N-CoT auxiliary reward is designed to alleviate the sparse rewards in P-CoT optimization. Extensive experiments demonstrate that Parrot significantly enhances both the performance of N-CoT and P-CoT, especially on N-CoT. Using Parrot SFT, the N-CoT performance of LLaMA2 and CodeLLaMA achieve gains of +21.87 and +21.48 on MathQA over the RL baseline, which is resource-intensive.

Ming Zhang, Yujiong Shen, Jingyi Deng et al.

Existing evaluation of Large Language Models (LLMs) on static benchmarks is vulnerable to data contamination and leaderboard overfitting, critical issues that obscure true model capabilities. To address this, we introduce LLMEval-3, a framework for dynamic evaluation of LLMs. LLMEval-3 is built on a proprietary bank of 220k graduate-level questions, from which it dynamically samples unseen test sets for each evaluation run. Its automated pipeline ensures integrity via contamination-resistant data curation, a novel anti-cheating architecture, and a calibrated LLM-as-a-judge process achieving 90% agreement with human experts, complemented by a relative ranking system for fair comparison. An 20-month longitudinal study of nearly 50 leading models reveals a performance ceiling on knowledge memorization and exposes data contamination vulnerabilities undetectable by static benchmarks. The framework demonstrates exceptional robustness in ranking stability and consistency, providing strong empirical validation for the dynamic evaluation paradigm. LLMEval-3 offers a robust and credible methodology for assessing the true capabilities of LLMs beyond leaderboard scores, promoting the development of more trustworthy evaluation standards.

Yuhui Wang, Qingyuan Wu, Dylan R. Ashley et al.

The Value Iteration Network (VIN) is an end-to-end differentiable neural network architecture for planning. It exhibits strong generalization to unseen domains by incorporating a differentiable planning module that operates on a latent Markov Decision Process (MDP). However, VINs struggle to scale to long-term and large-scale planning tasks, such as navigating a 100x100 maze -- a task that typically requires thousands of planning steps to solve. We observe that this deficiency is due to two issues: the representation capacity of the latent MDP and the planning module's depth. We address these by augmenting the latent MDP with a dynamic transition kernel, dramatically improving its representational capacity, and, to mitigate the vanishing gradient problem, introduce an "adaptive highway loss" that constructs skip connections to improve gradient flow. We evaluate our method on 2D/3D maze navigation environments, continuous control, and the real-world Lunar rover navigation task. We find that our new method, named Dynamic Transition VIN (DT-VIN), scales to 5000 layers and solves challenging versions of the above tasks. Altogether, we believe that DT-VIN represents a concrete step forward in performing long-term large-scale planning in complex environments.

Yuhui Wang, Weida Li, Francesco Faccio et al.

Value iteration networks (VINs) enable end-to-end learning for planning tasks by employing a differentiable "planning module" that approximates the value iteration algorithm. However, long-term planning remains a challenge because training very deep VINs is difficult. To address this problem, we embed highway value iteration -- a recent algorithm designed to facilitate long-term credit assignment -- into the structure of VINs. This improvement augments the "planning module" of the VIN with three additional components: 1) an "aggregate gate," which constructs skip connections to improve information flow across many layers; 2) an "exploration module," crafted to increase the diversity of information and gradient flow in spatial dimensions; 3) a "filter gate" designed to ensure safe exploration. The resulting novel highway VIN can be trained effectively with hundreds of layers using standard backpropagation. In long-term planning tasks requiring hundreds of planning steps, deep highway VINs outperform both traditional VINs and several advanced, very deep NNs.

Mingchen Zhuge, Haozhe Liu, Francesco Faccio et al.

Both Minsky's "society of mind" and Schmidhuber's "learning to think" inspire diverse societies of large multimodal neural networks (NNs) that solve problems by interviewing each other in a "mindstorm." Recent implementations of NN-based societies of minds consist of large language models (LLMs) and other NN-based experts communicating through a natural language interface. In doing so, they overcome the limitations of single LLMs, improving multimodal zero-shot reasoning. In these natural language-based societies of mind (NLSOMs), new agents -- all communicating through the same universal symbolic language -- are easily added in a modular fashion. To demonstrate the power of NLSOMs, we assemble and experiment with several of them (having up to 129 members), leveraging mindstorms in them to solve some practical AI tasks: visual question answering, image captioning, text-to-image synthesis, 3D generation, egocentric retrieval, embodied AI, and general language-based task solving. We view this as a starting point towards much larger NLSOMs with billions of agents-some of which may be humans. And with this emergence of great societies of heterogeneous minds, many new research questions have suddenly become paramount to the future of artificial intelligence. What should be the social structure of an NLSOM? What would be the (dis)advantages of having a monarchical rather than a democratic structure? How can principles of NN economies be used to maximize the total reward of a reinforcement learning NLSOM? In this work, we identify, discuss, and try to answer some of these questions.

Yuhui Wang, Diane J. Cook

Time series data are valuable but are often inscrutable. Gaining trust in time series classifiers for finance, healthcare, and other critical applications may rely on creating interpretable models. Researchers have previously been forced to decide between interpretable methods that lack predictive power and deep learning methods that lack transparency. In this paper, we propose a novel Mimic algorithm that retains the predictive accuracy of the strongest classifiers while introducing interpretability. Mimic mirrors the learning method of an existing multivariate time series classifier while simultaneously producing a visual representation that enhances user understanding of the learned model. Experiments on 26 time series datasets support Mimic's ability to imitate a variety of time series classifiers visually and accurately.

Chao Wen, Miao Xu, Zhilin Zhang et al.

In online advertising, auto-bidding has become an essential tool for advertisers to optimize their preferred ad performance metrics by simply expressing high-level campaign objectives and constraints. Previous works designed auto-bidding tools from the view of single-agent, without modeling the mutual influence between agents. In this paper, we instead consider this problem from a distributed multi-agent perspective, and propose a general $\underline{M}$ulti-$\underline{A}$gent reinforcement learning framework for $\underline{A}$uto-$\underline{B}$idding, namely MAAB, to learn the auto-bidding strategies. First, we investigate the competition and cooperation relation among auto-bidding agents, and propose a temperature-regularized credit assignment to establish a mixed cooperative-competitive paradigm. By carefully making a competition and cooperation trade-off among agents, we can reach an equilibrium state that guarantees not only individual advertiser's utility but also the system performance (i.e., social welfare). Second, to avoid the potential collusion behaviors of bidding low prices underlying the cooperation, we further propose bar agents to set a personalized bidding bar for each agent, and then alleviate the revenue degradation due to the cooperation. Third, to deploy MAAB in the large-scale advertising system with millions of advertisers, we propose a mean-field approach. By grouping advertisers with the same objective as a mean auto-bidding agent, the interactions among the large-scale advertisers are greatly simplified, making it practical to train MAAB efficiently. Extensive experiments on the offline industrial dataset and Alibaba advertising platform demonstrate that our approach outperforms several baseline methods in terms of social welfare and revenue.

Yuhui Wang, Qingyuan Wu, Pengcheng He et al.

Most of the policy evaluation algorithms are based on the theories of Bellman Expectation and Optimality Equation, which derive two popular approaches - Policy Iteration (PI) and Value Iteration (VI). However, multi-step bootstrapping is often at cross-purposes with and off-policy learning in PI-based methods due to the large variance of multi-step off-policy correction. In contrast, VI-based methods are naturally off-policy but subject to one-step learning.In this paper, we deduce a novel multi-step Bellman Optimality Equation by utilizing a latent structure of multi-step bootstrapping with the optimal value function. Via this new equation, we derive a new multi-step value iteration method that converges to the optimal value function with exponential contraction rate $\mathcal{O}(γ^n)$ but only linear computational complexity. Moreover, it can naturally derive a suite of multi-step off-policy algorithms that can safely utilize data collected by arbitrary policies without correction.Experiments reveal that the proposed methods are reliable, easy to implement and achieve state-of-the-art performance on a series of standard benchmark datasets.

Yang You, Yuhui Wang, Huan Zhang et al.

Large-batch training is an efficient approach for current distributed deep learning systems. It has enabled researchers to reduce the ImageNet/ResNet-50 training from 29 hours to around 1 minute. In this paper, we focus on studying the limit of the batch size. We think it may provide a guidance to AI supercomputer and algorithm designers. We provide detailed numerical optimization instructions for step-by-step comparison. Moreover, it is important to understand the generalization and optimization performance of huge batch training. Hoffer et al. introduced "ultra-slow diffusion" theory to large-batch training. However, our experiments show contradictory results with the conclusion of Hoffer et al. We provide comprehensive experimental results and detailed analysis to study the limitations of batch size scaling and "ultra-slow diffusion" theory. For the first time we scale the batch size on ImageNet to at least a magnitude larger than all previous work, and provide detailed studies on the performance of many state-of-the-art optimization schemes under this setting. We propose an optimization recipe that is able to improve the top-1 test accuracy by 18% compared to the baseline.

Xinghu Yao, Chao Wen, Yuhui Wang et al.

Learning a stable and generalizable centralized value function (CVF) is a crucial but challenging task in multi-agent reinforcement learning (MARL), as it has to deal with the issue that the joint action space increases exponentially with the number of agents in such scenarios. This paper proposes an approach, named SMIX($λ$), to address the issue using an efficient off-policy centralized training method within a flexible learner search space. As importance sampling for such off-policy training is both computationally costly and numerically unstable, we proposed to use the $λ$-return as a proxy to compute the TD error. With this new loss function objective, we adopt a modified QMIX network structure as the base to train our model. By further connecting it with the ${Q(λ)}$ approach from an unified expectation correction viewpoint, we show that the proposed SMIX($λ$) is equivalent to ${Q(λ)}$ and hence shares its convergence properties, while without being suffered from the aforementioned curse of dimensionality problem inherent in MARL. Experiments on the StarCraft Multi-Agent Challenge (SMAC) benchmark demonstrate that our approach not only outperforms several state-of-the-art MARL methods by a large margin, but also can be used as a general tool to improve the overall performance of other CTDE-type algorithms by enhancing their CVFs.

Yuhui Wang, Hao He, Chao Wen et al.

Proximal policy optimization (PPO) is one of the most successful deep reinforcement-learning methods, achieving state-of-the-art performance across a wide range of challenging tasks. However, its optimization behavior is still far from being fully understood. In this paper, we show that PPO could neither strictly restrict the likelihood ratio as it attempts to do nor enforce a well-defined trust region constraint, which means that it may still suffer from the risk of performance instability. To address this issue, we present an enhanced PPO method, named Truly PPO. Two critical improvements are made in our method: 1) it adopts a new clipping function to support a rollback behavior to restrict the difference between the new policy and the old one; 2) the triggering condition for clipping is replaced with a trust region-based one, such that optimizing the resulted surrogate objective function provides guaranteed monotonic improvement of the ultimate policy performance. It seems, by adhering more truly to making the algorithm proximal - confining the policy within the trust region, the new algorithm improves the original PPO on both sample efficiency and performance.

Yuhui Wang, Hao He, Xiaoyang Tan

In real-world scenarios, the observation data for reinforcement learning with continuous control is commonly noisy and part of it may be dynamically missing over time, which violates the assumption of many current methods developed for this. We addressed the issue within the framework of partially observable Markov Decision Process (POMDP) using a model-based method, in which the transition model is estimated from the incomplete and noisy observations using a newly proposed surrogate loss function with local approximation, while the policy and value function is learned with the help of belief imputation. For the latter purpose, a generative model is constructed and is seamlessly incorporated into the belief updating procedure of POMDP, which enables robust execution even under a significant incompleteness and noise. The effectiveness of the proposed method is verified on a collection of benchmark tasks, showing that our approach outperforms several compared methods under various challenging scenarios.

Yuhui Wang, Hao He, Xiaoyang Tan et al.

Proximal policy optimization (PPO) is one of the most popular deep reinforcement learning (RL) methods, achieving state-of-the-art performance across a wide range of challenging tasks. However, as a model-free RL method, the success of PPO relies heavily on the effectiveness of its exploratory policy search. In this paper, we give an in-depth analysis on the exploration behavior of PPO, and show that PPO is prone to suffer from the risk of lack of exploration especially under the case of bad initialization, which may lead to the failure of training or being trapped in bad local optima. To address these issues, we proposed a novel policy optimization method, named Trust Region-Guided PPO (TRGPPO), which adaptively adjusts the clipping range within the trust region. We formally show that this method not only improves the exploration ability within the trust region but enjoys a better performance bound compared to the original PPO as well. Extensive experiments verify the advantage of the proposed method.