Zihan Ding

Kewei Xu, Xiaoben Lu, Shuofei Qiao et al.

Real-world data analysis is inherently iterative, yet existing benchmarks mostly evaluate isolated or short interactive tasks, leaving agents' ability to track evolving analytical context over long horizons untested. We introduce LongDS, a benchmark for long-horizon, multi-turn data analysis where agents must maintain, update, restore, and compose evolving analytical states. LongDS comprises 68 tasks constructed from real-world Kaggle notebooks, spanning 2,225 turns across six domains including Geoscience, Business, and Education. Tasks are designed around state-evolution patterns (e.g., counterfactual perturbation, rollback, multi-state composition), with an average dependency span of 11.3 turns. Evaluating five state-of-the-art models, we find that the best model reaches only 48.45% average accuracy, performance drops nearly 47 points from early to late turns, and long-horizon errors account for 52%--69% of failures. Further analysis shows that additional agent steps do not necessarily improve performance, suggesting that the key bottleneck is maintaining a correct analytical state rather than increasing interaction budget. We release LongDS to support research on reliable long-horizon agentic data analysis. Code and data will be released at https://github.com/zjunlp/DataMind.

Luting Wang, Yi Liu, Penghui Du et al.

Open-vocabulary object detection aims to provide object detectors trained on a fixed set of object categories with the generalizability to detect objects described by arbitrary text queries. Previous methods adopt knowledge distillation to extract knowledge from Pretrained Vision-and-Language Models (PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal cropping and single-level feature mimicking processes, they suffer from information destruction during knowledge extraction and inefficient knowledge transfer. To remedy these limitations, we propose an Object-Aware Distillation Pyramid (OADP) framework, including an Object-Aware Knowledge Extraction (OAKE) module and a Distillation Pyramid (DP) mechanism. When extracting object knowledge from PVLMs, the former adaptively transforms object proposals and adopts object-aware mask attention to obtain precise and complete knowledge of objects. The latter introduces global and block distillation for more comprehensive knowledge transfer to compensate for the missing relation information in object distillation. Extensive experiments show that our method achieves significant improvement compared to current methods. Especially on the MS-COCO dataset, our OADP framework reaches $35.6$ mAP$^{\text{N}}_{50}$, surpassing the current state-of-the-art method by $3.3$ mAP$^{\text{N}}_{50}$. Code is released at https://github.com/LutingWang/OADP.

Zihan Ding, Yuanpei Chen, Allen Z. Ren et al. · baidu

Generating human-like behavior on robots is a great challenge especially in dexterous manipulation tasks with robotic hands. Scripting policies from scratch is intractable due to the high-dimensional control space, and training policies with reinforcement learning (RL) and manual reward engineering can also be hard and lead to unnatural motions. Leveraging the recent progress on RL from Human Feedback, we propose a framework that learns a universal human prior using direct human preference feedback over videos, for efficiently tuning the RL policies on 20 dual-hand robot manipulation tasks in simulation, without a single human demonstration. A task-agnostic reward model is trained through iteratively generating diverse polices and collecting human preference over the trajectories; it is then applied for regularizing the behavior of polices in the fine-tuning stage. Our method empirically demonstrates more human-like behaviors on robot hands in diverse tasks including even unseen tasks, indicating its generalization capability.

Yuqing Wang, Zhijie Lin, Ceyuan Yang et al.

Unified multimodal models (UMMs) aim to handle perception and generation in a single model. Yet existing UMMs still rely on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck. Naively removing it introduces a quality gap, as the model must learn both high-level structure and low-level details from raw pixels. In this paper, we propose Representation Forcing (RF), a technique that closes this gap by making representation prediction a native capability of the model. Concretely, RF forces the decoder to autoregressively predict visual representations as intermediate tokens before pixels; these tokens then stay in context to guide pixel diffusion within the same backbone. By turning representations from perception outputs into generation targets, RF eliminates the need for any external generative latent space. We find that RF benefits both understanding and generation. On image generation, our pixel-space model with RF matches state-of-the-art VAE-based unified models. On image understanding, pixel-space RF generally outperforms its VAE-based variant. Together, these results offer an effective step toward end-to-end, bottleneck-free UMMs.

Zihan Ding, Tianrui Hui, Junshi Huang et al.

Referring video object segmentation aims to predict foreground labels for objects referred by natural language expressions in videos. Previous methods either depend on 3D ConvNets or incorporate additional 2D ConvNets as encoders to extract mixed spatial-temporal features. However, these methods suffer from spatial misalignment or false distractors due to delayed and implicit spatial-temporal interaction occurring in the decoding phase. To tackle these limitations, we propose a Language-Bridged Duplex Transfer (LBDT) module which utilizes language as an intermediary bridge to accomplish explicit and adaptive spatial-temporal interaction earlier in the encoding phase. Concretely, cross-modal attention is performed among the temporal encoder, referring words and the spatial encoder to aggregate and transfer language-relevant motion and appearance information. In addition, we also propose a Bilateral Channel Activation (BCA) module in the decoding phase for further denoising and highlighting the spatial-temporal consistent features via channel-wise activation. Extensive experiments show our method achieves new state-of-the-art performances on four popular benchmarks with 6.8% and 6.9% absolute AP gains on A2D Sentences and J-HMDB Sentences respectively, while consuming around 7x less computational overhead.

Zihan Ding, Dijia Su, Qinghua Liu et al.

This paper proposes new, end-to-end deep reinforcement learning algorithms for learning two-player zero-sum Markov games. Different from prior efforts on training agents to beat a fixed set of opponents, our objective is to find the Nash equilibrium policies that are free from exploitation by even the adversarial opponents. We propose (a) Nash-DQN algorithm, which integrates the deep learning techniques from single DQN into the classic Nash Q-learning algorithm for solving tabular Markov games; (b) Nash-DQN-Exploiter algorithm, which additionally adopts an exploiter to guide the exploration of the main agent. We conduct experimental evaluation on tabular examples as well as various two-player Atari games. Our empirical results demonstrate that (i) the policies found by many existing methods including Neural Fictitious Self Play and Policy Space Response Oracle can be prone to exploitation by adversarial opponents; (ii) the output policies of our algorithms are robust to exploitation, and thus outperform existing methods.

Zihan Ding, Zi-han Ding, Tianrui Hui et al.

Panoptic Narrative Grounding (PNG) is an emerging task whose goal is to segment visual objects of things and stuff categories described by dense narrative captions of a still image. The previous two-stage approach first extracts segmentation region proposals by an off-the-shelf panoptic segmentation model, then conducts coarse region-phrase matching to ground the candidate regions for each noun phrase. However, the two-stage pipeline usually suffers from the performance limitation of low-quality proposals in the first stage and the loss of spatial details caused by region feature pooling, as well as complicated strategies designed for things and stuff categories separately. To alleviate these drawbacks, we propose a one-stage end-to-end Pixel-Phrase Matching Network (PPMN), which directly matches each phrase to its corresponding pixels instead of region proposals and outputs panoptic segmentation by simple combination. Thus, our model can exploit sufficient and finer cross-modal semantic correspondence from the supervision of densely annotated pixel-phrase pairs rather than sparse region-phrase pairs. In addition, we also propose a Language-Compatible Pixel Aggregation (LCPA) module to further enhance the discriminative ability of phrase features through multi-round refinement, which selects the most compatible pixels for each phrase to adaptively aggregate the corresponding visual context. Extensive experiments show that our method achieves new state-of-the-art performance on the PNG benchmark with 4.0 absolute Average Recall gains.

Tianrui Hui, Zihan Ding, Junshi Huang et al.

Panoptic narrative grounding (PNG) aims to segment things and stuff objects in an image described by noun phrases of a narrative caption. As a multimodal task, an essential aspect of PNG is the visual-linguistic interaction between image and caption. The previous two-stage method aggregates visual contexts from offline-generated mask proposals to phrase features, which tend to be noisy and fragmentary. The recent one-stage method aggregates only pixel contexts from image features to phrase features, which may incur semantic misalignment due to lacking object priors. To realize more comprehensive visual-linguistic interaction, we propose to enrich phrases with coupled pixel and object contexts by designing a Phrase-Pixel-Object Transformer Decoder (PPO-TD), where both fine-grained part details and coarse-grained entity clues are aggregated to phrase features. In addition, we also propose a PhraseObject Contrastive Loss (POCL) to pull closer the matched phrase-object pairs and push away unmatched ones for aggregating more precise object contexts from more phrase-relevant object tokens. Extensive experiments on the PNG benchmark show our method achieves new state-of-the-art performance with large margins.

Zihan Ding, Chi Jin

Score-based generative models like the diffusion model have been testified to be effective in modeling multi-modal data from image generation to reinforcement learning (RL). However, the inference process of diffusion model can be slow, which hinders its usage in RL with iterative sampling. We propose to apply the consistency model as an efficient yet expressive policy representation, namely consistency policy, with an actor-critic style algorithm for three typical RL settings: offline, offline-to-online and online. For offline RL, we demonstrate the expressiveness of generative models as policies from multi-modal data. For offline-to-online RL, the consistency policy is shown to be more computational efficient than diffusion policy, with a comparable performance. For online RL, the consistency policy demonstrates significant speedup and even higher average performances than the diffusion policy.

Hongyu Li, Tianrui Hui, Zihan Ding et al.

Panoptic narrative grounding (PNG), whose core target is fine-grained image-text alignment, requires a panoptic segmentation of referred objects given a narrative caption. Previous discriminative methods achieve only weak or coarse-grained alignment by panoptic segmentation pretraining or CLIP model adaptation. Given the recent progress of text-to-image Diffusion models, several works have shown their capability to achieve fine-grained image-text alignment through cross-attention maps and improved general segmentation performance. However, the direct use of phrase features as static prompts to apply frozen Diffusion models to the PNG task still suffers from a large task gap and insufficient vision-language interaction, yielding inferior performance. Therefore, we propose an Extractive-Injective Phrase Adapter (EIPA) bypass within the Diffusion UNet to dynamically update phrase prompts with image features and inject the multimodal cues back, which leverages the fine-grained image-text alignment capability of Diffusion models more sufficiently. In addition, we also design a Multi-Level Mutual Aggregation (MLMA) module to reciprocally fuse multi-level image and phrase features for segmentation refinement. Extensive experiments on the PNG benchmark show that our method achieves new state-of-the-art performance.

Zhiyuan Yao, Zihan Ding

This paper investigates the network load balancing problem in data centers (DCs) where multiple load balancers (LBs) are deployed, using the multi-agent reinforcement learning (MARL) framework. The challenges of this problem consist of the heterogeneous processing architecture and dynamic environments, as well as limited and partial observability of each LB agent in distributed networking systems, which can largely degrade the performance of in-production load balancing algorithms in real-world setups. Centralised-training-decentralised-execution (CTDE) RL scheme has been proposed to improve MARL performance, yet it incurs -- especially in distributed networking systems, which prefer distributed and plug-and-play design scheme -- additional communication and management overhead among agents. We formulate the multi-agent load balancing problem as a Markov potential game, with a carefully and properly designed workload distribution fairness as the potential function. A fully distributed MARL algorithm is proposed to approximate the Nash equilibrium of the game. Experimental evaluations involve both an event-driven simulator and real-world system, where the proposed MARL load balancing algorithm shows close-to-optimal performance in simulations, and superior results over in-production LBs in the real-world system.

Zihan Ding, Kexin Jin, Jonas Latz et al.

Deep neural networks and other modern machine learning models are often susceptible to adversarial attacks. Indeed, an adversary may often be able to change a model's prediction through a small, directed perturbation of the model's input - an issue in safety-critical applications. Adversarially robust machine learning is usually based on a minmax optimisation problem that minimises the machine learning loss under maximisation-based adversarial attacks. In this work, we study adversaries that determine their attack using a Bayesian statistical approach rather than maximisation. The resulting Bayesian adversarial robustness problem is a relaxation of the usual minmax problem. To solve this problem, we propose Abram - a continuous-time particle system that shall approximate the gradient flow corresponding to the underlying learning problem. We show that Abram approximates a McKean-Vlasov process and justify the use of Abram by giving assumptions under which the McKean-Vlasov process finds the minimiser of the Bayesian adversarial robustness problem. We discuss two ways to discretise Abram and show its suitability in benchmark adversarial deep learning experiments.

Zihan Ding, Xiaoxi Wei, Yidan Xu

Human consciousness has been a long-lasting mystery for centuries, while machine intelligence and consciousness is an arduous pursuit. Researchers have developed diverse theories for interpreting the consciousness phenomenon in human brains from different perspectives and levels. This paper surveys several main branches of consciousness theories originating from different subjects including information theory, quantum physics, cognitive psychology, physiology and computer science, with the aim of bridging these theories from a computational perspective. It also discusses the existing evaluation metrics of consciousness and possibility for current computational models to be conscious. Breaking the mystery of consciousness can be an essential step in building general artificial intelligence with computing machines.

Yuheng Zheng, Zihan Ding

This paper establishes a new and comprehensive theoretical analysis for the application of reinforcement learning (RL) in high-frequency market making. We bridge the modern RL theory and the continuous-time statistical models in high-frequency financial economics. Different with most existing literature on methodological research about developing various RL methods for market making problem, our work is a pilot to provide the theoretical analysis. We target the effects of sampling frequency, and find an interesting tradeoff between error and complexity of RL algorithm when tweaking the values of the time increment $Δ$ $-$ as $Δ$ becomes smaller, the error will be smaller but the complexity will be larger. We also study the two-player case under the general-sum game framework and establish the convergence of Nash equilibrium to the continuous-time game equilibrium as $Δ\rightarrow0$. The Nash Q-learning algorithm, which is an online multi-agent RL method, is applied to solve the equilibrium. Our theories are not only useful for practitioners to choose the sampling frequency, but also very general and applicable to other high-frequency financial decision making problems, e.g., optimal executions, as long as the time-discretization of a continuous-time markov decision process is adopted. Monte Carlo simulation evidence support all of our theories.

Hao Ju, Shaofei Huang, Hongyu Li et al.

Current mobile manipulation research predominantly follows an instruction-driven paradigm, where agents rely on predefined textual commands to execute tasks. However, this setting confines agents to a passive role, limiting their autonomy and ability to react to dynamic environmental events. To address these limitations, we introduce sound-triggered mobile manipulation, where agents must actively perceive and interact with sound-emitting objects without explicit action instructions. To support these tasks, we develop Habitat-Echo, a data platform that integrates acoustic rendering with physical interaction. We further propose a baseline comprising a high-level task planner and low-level policy models to complete these tasks. Extensive experiments show that the proposed baseline empowers agents to actively detect and respond to auditory events, eliminating the need for case-by-case instructions. Notably, in the challenging dual-source scenario, the agent successfully isolates the primary source from overlapping acoustic interference to execute the first interaction, and subsequently proceeds to manipulate the secondary object, verifying the robustness of the baseline.

Wenzhe Li, Zihan Ding, Seth Karten et al.

Recent advances in reinforcement learning (RL) heavily rely on a variety of well-designed benchmarks, which provide environmental platforms and consistent criteria to evaluate existing and novel algorithms. Specifically, in multi-agent RL (MARL), a plethora of benchmarks based on cooperative games have spurred the development of algorithms that improve the scalability of cooperative multi-agent systems. However, for the competitive setting, a lightweight and open-sourced benchmark with challenging gaming dynamics and visual inputs has not yet been established. In this work, we present FightLadder, a real-time fighting game platform, to empower competitive MARL research. Along with the platform, we provide implementations of state-of-the-art MARL algorithms for competitive games, as well as a set of evaluation metrics to characterize the performance and exploitability of agents. We demonstrate the feasibility of this platform by training a general agent that consistently defeats 12 built-in characters in single-player mode, and expose the difficulty of training a non-exploitable agent without human knowledge and demonstrations in two-player mode. FightLadder provides meticulously designed environments to address critical challenges in competitive MARL research, aiming to catalyze a new era of discovery and advancement in the field. Videos and code at https://sites.google.com/view/fightladder/home.

Yuhang Song, Andrzej Wojcicki, Thomas Lukasiewicz et al.

Learning agents that are not only capable of taking tests, but also innovating is becoming a hot topic in AI. One of the most promising paths towards this vision is multi-agent learning, where agents act as the environment for each other, and improving each agent means proposing new problems for others. However, existing evaluation platforms are either not compatible with multi-agent settings, or limited to a specific game. That is, there is not yet a general evaluation platform for research on multi-agent intelligence. To this end, we introduce Arena, a general evaluation platform for multi-agent intelligence with 35 games of diverse logics and representations. Furthermore, multi-agent intelligence is still at the stage where many problems remain unexplored. Therefore, we provide a building toolkit for researchers to easily invent and build novel multi-agent problems from the provided game set based on a GUI-configurable social tree and five basic multi-agent reward schemes. Finally, we provide Python implementations of five state-of-the-art deep multi-agent reinforcement learning baselines. Along with the baseline implementations, we release a set of 100 best agents/teams that we can train with different training schemes for each game, as the base for evaluating agents with population performance. As such, the research community can perform comparisons under a stable and uniform standard. All the implementations and accompanied tutorials have been open-sourced for the community at https://sites.google.com/view/arena-unity/.

Zihan Ding, Amy Zhang, Yuandong Tian et al.

We introduce Diffusion World Model (DWM), a conditional diffusion model capable of predicting multistep future states and rewards concurrently. As opposed to traditional one-step dynamics models, DWM offers long-horizon predictions in a single forward pass, eliminating the need for recursive queries. We integrate DWM into model-based value estimation, where the short-term return is simulated by future trajectories sampled from DWM. In the context of offline reinforcement learning, DWM can be viewed as a conservative value regularization through generative modeling. Alternatively, it can be seen as a data source that enables offline Q-learning with synthetic data. Our experiments on the D4RL dataset confirm the robustness of DWM to long-horizon simulation. In terms of absolute performance, DWM significantly surpasses one-step dynamics models with a $44\%$ performance gain, and is comparable to or slightly surpassing their model-free counterparts.

Ceyuan Yang, Zhijie Lin, Yang Zhao et al.

We present Omni, a unified multimodal model natively trained on diverse modalities, including text, images, videos, 3D geometry, and hidden representations. We find that such training enables Context Unrolling, where the model explicitly reasons across multiple modal representations before producing predictions. This process enables the model to aggregate complementary information across heterogeneous modalities, facilitating a more faithful approximation of the shared multimodal knowledge manifold and improving downstream reasoning fidelity. As a result, Omni achieves strong performance on both multimodal generation and understanding benchmarks, while demonstrating advanced multimodal reasoning capabilities, including in-context generation of text, image, video, and 3D geometry.

Chengshuai Shi, Wenzhe Li, Xinran Liang et al.

Given the rapidly growing capabilities of vision-language models (VLMs), extending them to interactive decision-making tasks such as video games has emerged as a promising frontier. However, existing approaches either rely on large-scale supervised fine-tuning (SFT) on human trajectories or apply reinforcement learning (RL) only in relatively short-horizon settings (typically around 20--30 turns). In this work, we study RL-based training of VLMs for long-horizon decision-making in Super Mario Land, a visually grounded environment requiring 100+ turns of interaction with coordinated perception, reasoning, and action. We begin with a systematic investigation of key algorithmic components and propose an adapted variant of PPO with a lightweight turn-level critic, which substantially improves training stability and sample efficiency over critic-free methods such as GRPO and Reinforce++. We further show that pretrained VLMs provide strong action priors, significantly improving sample efficiency during RL training and reducing the need for manual design choices such as action engineering, compared to classical deep RL trained from scratch. Building on these insights, we introduce Odysseus, an open training framework for VLM agents, achieving substantial gains across multiple levels of the game and at least 3 times average game progresses than frontier models. Moreover, the trained models exhibit consistent improvements under both in-game and cross-game generalization settings, while maintaining general-domain capabilities. Overall, our results identify key ingredients for making RL stable and effective in long-horizon, multi-modal settings, and provide practical guidance for developing VLMs as embodied agents.

Zihan Ding, Ziyuan Yang, Yi Zhang

The rapid spread of memes makes harmful content detection increasingly crucial, as effective identification can curb the circulation of misinformation. However, existing methods rely heavily on high-volume annotated data, which leads to substantial training costs and limited generalization. To address these challenges, we propose PrismAgent, a zero-shot, multi-agent, interpretable framework. PrismAgent conceptualizes this task as a criminal case investigation, employing four specialized agents responsible for the analysis, investigation, prosecution, and judgment stages within a structured collaborative workflow. In the first stage, the analyst agent paraphrases each meme under benevolent and malicious assumptions to probe its underlying intent. The investigator agent then retrieves supporting evidence from an unannotated dataset and constructs contextual interpretations for the meme and its variants. Next, the prosecutor agent performs three independent preliminary judgments by pairing the original meme with each of the three interpretations. Finally, the judge agent deliberates across all evidence to render a final verdict. Moreover, PrismAgent's explicit multi-stage reasoning chain makes the model inherently interpretable, as every intermediate step is explicitly explained rather than only producing a final detection result. Extensive experiments on three public datasets show that PrismAgent significantly outperforms existing zero-shot detection methods.

Zihan Ding, Ziyuan Yang, Yi Zhang

Multimodal controversy detection (MCD) identifies controversial content in videos and their associated user comments, to support risk management for social video platforms.Prior research frames MCD as a static representation learning task, where features are directly extracted from videos and their accompanying comments. However, these methods fail to capture the diverse perspectives and evaluations from different audience groups. Inspired by the real-world process of content dissemination among audiences, we propose AuDisAgent, a training-free multi-agent framework that reformulates MCD as a dynamic propagation process.Our framework explicitly models audience dissemination through a structured multi-agent system. First, three specialized Screening Agents (Video Agent, Comment Agent, and Interaction Agent) conduct initial assessments from visual, textual, and cross-modal perspectives, respectively. For samples where the three agents cannot reach a consensus, a Viewing Panel Agent is activated to simulate post-screening discussions among audiences with diverse backgrounds and stances. This mechanism models how different audience groups interpret and react to the same content, uncovering latent controversial content that may emerge during the dissemination process. Finally, an Arbitration Agent renders the final judgment based on the complete reasoning chain from the preceding steps.In addition, to address the "cold-start" scenario where newly released videos have few or no comments, we design a Comment Bootstrapping Strategy that leverages historical public comments from semantically similar videos as the initial comment context. Extensive experiments on a public dataset demonstrate that our framework significantly outperforms existing state-of-the-art (SOTA) methods in both rich-comment and limited-comment scenarios.

Anjian Li, Zihan Ding, Adji Bousso Dieng et al.

Optimal trajectory design is computationally expensive for nonlinear and high-dimensional dynamical systems. The challenge arises from the non-convex nature of the optimization problem with multiple local optima, which usually requires a global search. Traditional numerical solvers struggle to find diverse solutions efficiently without appropriate initial guesses. In this paper, we introduce DiffuSolve, a general diffusion model-based solver for non-convex trajectory optimization. An expressive diffusion model is trained on pre-collected locally optimal solutions and efficiently samples initial guesses, which then warm-starts numerical solvers to fine-tune the feasibility and optimality. We also present DiffuSolve+, a novel constrained diffusion model with an additional loss in training that further reduces the problem constraint violations of diffusion samples. Experimental evaluations on three tasks verify the improved robustness, diversity, and a 2$\times$ to 11$\times$ increase in computational efficiency with our proposed method, which generalizes well to trajectory optimization problems of varying challenges.

Linqing Zhong, Chen Gao, Zihan Ding et al.

The Zero-Shot Object Navigation (ZSON) task requires embodied agents to find a previously unseen object by navigating in unfamiliar environments. Such a goal-oriented exploration heavily relies on the ability to perceive, understand, and reason based on the spatial information of the environment. However, current LLM-based approaches convert visual observations to language descriptions and reason in the linguistic space, leading to the loss of spatial information. In this paper, we introduce TopV-Nav, an MLLM-based method that directly reasons on the top-view map with sufficient spatial information. To fully unlock the MLLM's spatial reasoning potential in top-view perspective, we propose the Adaptive Visual Prompt Generation (AVPG) method to adaptively construct semantically-rich top-view map. It enables the agent to directly utilize spatial information contained in the top-view map to conduct thorough reasoning. Besides, we design a Dynamic Map Scaling (DMS) mechanism to dynamically zoom top-view map at preferred scales, enhancing local fine-grained reasoning. Additionally, we devise a Potential Target Driven (PTD) mechanism to predict and to utilize target locations, facilitating global and human-like exploration. Experiments on MP3D and HM3D datasets demonstrate the superiority of our TopV-Nav.

Si Liu, Zihan Ding, Jiahui Fu et al.

The objective of the collaborative vehicle-to-everything perception task is to enhance the individual vehicle's perception capability through message communication among neighboring traffic agents. Previous methods focus on achieving optimal performance within bandwidth limitations and typically adopt BEV maps as the basic collaborative message units. However, we demonstrate that collaboration with dense representations is plagued by object feature destruction during message packing, inefficient message aggregation for long-range collaboration, and implicit structure representation communication. To tackle these issues, we introduce a brand new message unit, namely point cluster, designed to represent the scene sparsely with a combination of low-level structure information and high-level semantic information. The point cluster inherently preserves object information while packing messages, with weak relevance to the collaboration range, and supports explicit structure modeling. Building upon this representation, we propose a novel framework V2X-PC for collaborative perception. This framework includes a Point Cluster Packing (PCP) module to keep object feature and manage bandwidth through the manipulation of cluster point numbers. As for effective message aggregation, we propose a Point Cluster Aggregation (PCA) module to match and merge point clusters associated with the same object. To further handle time latency and pose errors encountered in real-world scenarios, we propose parameter-free solutions that can adapt to different noisy levels without finetuning. Experiments on two widely recognized collaborative perception benchmarks showcase the superior performance of our method compared to the previous state-of-the-art approaches relying on BEV maps.

Seth Karten, Wenzhe Li, Zihan Ding et al.

We present the LLM Economist, a novel framework that uses agent-based modeling to design and assess economic policies in strategic environments with hierarchical decision-making. At the lower level, bounded rational worker agents -- instantiated as persona-conditioned prompts sampled from U.S. Census-calibrated income and demographic statistics -- choose labor supply to maximize text-based utility functions learned in-context. At the upper level, a planner agent employs in-context reinforcement learning to propose piecewise-linear marginal tax schedules anchored to the current U.S. federal brackets. This construction endows economic simulacra with three capabilities requisite for credible fiscal experimentation: (i) optimization of heterogeneous utilities, (ii) principled generation of large, demographically realistic agent populations, and (iii) mechanism design -- the ultimate nudging problem -- expressed entirely in natural language. Experiments with populations of up to one hundred interacting agents show that the planner converges near Stackelberg equilibria that improve aggregate social welfare relative to Saez solutions, while a periodic, persona-level voting procedure furthers these gains under decentralized governance. These results demonstrate that large language model-based agents can jointly model, simulate, and govern complex economic systems, providing a tractable test bed for policy evaluation at the societal scale to help build better civilizations.

Zihan Ding, Chi Jin, Difan Liu et al.

Diffusion probabilistic models have shown significant progress in video generation; however, their computational efficiency is limited by the large number of sampling steps required. Reducing sampling steps often compromises video quality or generation diversity. In this work, we introduce a distillation method that combines variational score distillation and consistency distillation to achieve few-step video generation, maintaining both high quality and diversity. We also propose a latent reward model fine-tuning approach to further enhance video generation performance according to any specified reward metric. This approach reduces memory usage and does not require the reward to be differentiable. Our method demonstrates state-of-the-art performance in few-step generation for 10-second videos (128 frames at 12 FPS). The distilled student model achieves a score of 82.57 on VBench, surpassing the teacher model as well as baseline models Gen-3, T2V-Turbo, and Kling. One-step distillation accelerates the teacher model's diffusion sampling by up to 278.6 times, enabling near real-time generation. Human evaluations further validate the superior performance of our 4-step student models compared to teacher model using 50-step DDIM sampling.

Zhongwen Xu, Zihan Ding

We revisit policy-gradient optimization for Large Language Models (LLMs) from a single-stream perspective. Prevailing group-based methods like GRPO reduce variance with on-the-fly baselines but suffer from critical flaws: frequent degenerate groups erase learning signals, and synchronization barriers hinder scalability. We introduce Single-stream Policy Optimization (SPO), which eliminates these issues by design. SPO replaces per-group baselines with a persistent, KL-adaptive value tracker and normalizes advantages globally across the batch, providing a stable, low-variance learning signal for every sample. Being group-free, SPO enables higher throughput and scales effectively in long-horizon or tool-integrated settings where generation times vary. Furthermore, the persistent value tracker naturally enables an adaptive curriculum via prioritized sampling. Experiments using Qwen3-8B show that SPO converges more smoothly and attains higher accuracy than GRPO, while eliminating computation wasted on degenerate groups. Ablation studies confirm that SPO's gains stem from its principled approach to baseline estimation and advantage normalization, offering a more robust and efficient path for LLM reasoning. Across five hard math benchmarks with Qwen3 8B, SPO improves the average maj@32 by +3.4 percentage points (pp) over GRPO, driven by substantial absolute point gains on challenging datasets, including +7.3 pp on BRUMO 25, +4.4 pp on AIME 25, +3.3 pp on HMMT 25, and achieves consistent relative gain in pass@$k$ across the evaluated $k$ values. SPO's success challenges the prevailing trend of adding incidental complexity to RL algorithms, highlighting a path where fundamental principles, not architectural workarounds, drive the next wave of progress in LLM reasoning.

Taiye Chen, Xun Hu, Zihan Ding et al.

Foundational world models must be both interactive and preserve spatiotemporal coherence for effective future planning with action choices. However, present models for long video generation have limited inherent world modeling capabilities due to two main challenges: compounding errors and insufficient memory mechanisms. We enhance image-to-video models with interactive capabilities through additional action conditioning and autoregressive framework, and reveal that compounding error is inherently irreducible in autoregressive video generation, while insufficient memory mechanism leads to incoherence of world models. We propose video retrieval augmented generation (VRAG) with explicit global state conditioning, which significantly reduces long-term compounding errors and increases spatiotemporal consistency of world models. In contrast, naive autoregressive generation with extended context windows and retrieval-augmented generation prove less effective for video generation, primarily due to the limited in-context learning capabilities of current video models. Our work illuminates the fundamental challenges in video world models and establishes a comprehensive benchmark for improving video generation models with internal world modeling capabilities.

Taiye Chen, Zihan Ding, Anjian Li et al.

Recent advancements in video generation have demonstrated the potential of using video diffusion models as world models, with autoregressive generation of infinitely long videos through masked conditioning. However, such models, usually with local full attention, lack effective memory compression and retrieval for long-term generation beyond the window size, leading to issues of forgetting and spatiotemporal inconsistencies. To enhance the retention of historical information within a fixed memory budget, we introduce a recurrent neural network (RNN) into the diffusion transformer framework. Specifically, a diffusion model incorporating LSTM with attention achieves comparable performance to state-of-the-art RNN blocks, such as TTT and Mamba2. Moreover, existing diffusion-RNN approaches often suffer from performance degradation due to training-inference gap or the lack of overlap across windows. To address these limitations, we propose a novel Recurrent Autoregressive Diffusion (RAD) framework, which executes frame-wise autoregression for memory update and retrieval, consistently across training and inference time. Experiments on Memory Maze and Minecraft datasets demonstrate the superiority of RAD for long video generation, highlighting the efficiency of LSTM in sequence modeling.

Zihan Ding, Xinyi Wang, Junlong Chen et al.

Creators struggle to edit long-form, narrative-rich videos not because of UI complexity, but due to the cognitive demands of searching, storyboarding, and sequencing hours of footage. Existing transcript- or embedding-based methods fall short for creative workflows, as models struggle to track characters, infer motivations, and connect dispersed events. We present a prompt-driven, modular editing system that helps creators restructure multi-hour content through free-form prompts rather than timelines. At its core is a semantic indexing pipeline that builds a global narrative via temporal segmentation, guided memory compression, and cross-granularity fusion, producing interpretable traces of plot, dialogue, emotion, and context. Users receive cinematic edits while optionally refining transparent intermediate outputs. Evaluated on 400+ videos with expert ratings, QA, and preference studies, our system scales prompt-driven editing, preserves narrative coherence, and balances automation with creator control.

Zihan Ding, Chi Jin

This handbook offers a unified perspective on diffusion models, encompassing diffusion probabilistic models, score-based generative models, consistency models, rectified flow, and related methods. By standardizing notations and aligning them with code implementations, it aims to bridge the "paper-to-code" gap and facilitate robust implementations and fair comparisons. The content encompasses the fundamentals of diffusion models, the pre-training process, and various post-training methods. Post-training techniques include model distillation and reward-based fine-tuning. Designed as a practical guide, it emphasizes clarity and usability over theoretical depth, focusing on widely adopted approaches in generative modeling with diffusion models.

Anjian Li, Zihan Ding, Adji Bousso Dieng et al.

The diffusion model has shown success in generating high-quality and diverse solutions to trajectory optimization problems. However, diffusion models with neural networks inevitably make prediction errors, which leads to constraint violations such as unmet goals or collisions. This paper presents a novel constraint-aware diffusion model for trajectory optimization. We introduce a novel hybrid loss function for training that minimizes the constraint violation of diffusion samples compared to the groundtruth while recovering the original data distribution. Our model is demonstrated on tabletop manipulation and two-car reach-avoid problems, outperforming traditional diffusion models in minimizing constraint violations while generating samples close to locally optimal solutions.

Zhiyuan Yao, Zihan Ding, Thomas Clausen

This paper presents the network load balancing problem, a challenging real-world task for multi-agent reinforcement learning (MARL) methods. Traditional heuristic solutions like Weighted-Cost Multi-Path (WCMP) and Local Shortest Queue (LSQ) are less flexible to the changing workload distributions and arrival rates, with a poor balance among multiple load balancers. The cooperative network load balancing task is formulated as a Dec-POMDP problem, which naturally induces the MARL methods. To bridge the reality gap for applying learning-based methods, all methods are directly trained and evaluated on an emulation system from moderate-to large-scale. Experiments on realistic testbeds show that the independent and "selfish" load balancing strategies are not necessarily the globally optimal ones, while the proposed MARL solution has a superior performance over different realistic settings. Additionally, the potential difficulties of MARL methods for network load balancing are analysed, which helps to draw the attention of the learning and network communities to such challenges.

Zhiyuan Yao, Zihan Ding, Thomas Heide Clausen

Network load balancers are central components in data centers, that distributes workloads across multiple servers and thereby contribute to offering scalable services. However, when load balancers operate in dynamic environments with limited monitoring of application server loads, they rely on heuristic algorithms that require manual configurations for fairness and performance. To alleviate that, this paper proposes a distributed asynchronous reinforcement learning mechanism to-with no active load balancer state monitoring and limited network observations-improve the fairness of the workload distribution achieved by a load balancer. The performance of proposed mechanism is evaluated and compared with stateof-the-art load balancing algorithms in a simulator, under configurations with progressively increasing complexities. Preliminary results show promise in RLbased load balancing algorithms, and identify additional challenges and future research directions, including reward function design and model scalability.

Zihan Ding

Domain randomization (DR) cannot provide optimal policies for adapting the learning agent to the dynamics of the environment, although it can generalize sub-optimal policies to work in a transferred domain. In this paper, we present Universal Policy with Embedding System Identification (UPESI) as an implicit system identification (SI) approach with universal policies (UPs), as a learning-based control method to execute optimal actions adaptively in environments with various dynamic properties. Previous approaches of SI for adaptive policies either conduct explicit SI, which is testified to be an ill-posed problem, or suffer from low efficiency without leveraging the simulation oracle. We propose to conduct SI in the embedding space of system dynamics by leveraging a learned forward dynamics model, and use Bayesian optimization for the SI process given transition data in a new environment. The identified embeddings are applied as additional input to the UP to enable its dynamics adaptability. Experiments demonstrate the advantageous performances of our proposed UP with embedding SI over standard DR and conventional SI approaches on both low-dimensional and high-dimensional simulation tasks.

Tianrui Hui, Shaofei Huang, Si Liu et al.

Language-queried video actor segmentation aims to predict the pixel-level mask of the actor which performs the actions described by a natural language query in the target frames. Existing methods adopt 3D CNNs over the video clip as a general encoder to extract a mixed spatio-temporal feature for the target frame. Though 3D convolutions are amenable to recognizing which actor is performing the queried actions, it also inevitably introduces misaligned spatial information from adjacent frames, which confuses features of the target frame and yields inaccurate segmentation. Therefore, we propose a collaborative spatial-temporal encoder-decoder framework which contains a 3D temporal encoder over the video clip to recognize the queried actions, and a 2D spatial encoder over the target frame to accurately segment the queried actors. In the decoder, a Language-Guided Feature Selection (LGFS) module is proposed to flexibly integrate spatial and temporal features from the two encoders. We also propose a Cross-Modal Adaptive Modulation (CMAM) module to dynamically recombine spatial- and temporal-relevant linguistic features for multimodal feature interaction in each stage of the two encoders. Our method achieves new state-of-the-art performance on two popular benchmarks with less computational overhead than previous approaches.

Jie Ren, Yewen Li, Zihan Ding et al.

Deep reinforcement learning (DRL) has successfully solved various problems recently, typically with a unimodal policy representation. However, grasping distinguishable skills for some tasks with non-unique optima can be essential for further improving its learning efficiency and performance, which may lead to a multimodal policy represented as a mixture-of-experts (MOE). To our best knowledge, present DRL algorithms for general utility do not deploy this method as policy function approximators due to the potential challenge in its differentiability for policy learning. In this work, we propose a probabilistic mixture-of-experts (PMOE) implemented with a Gaussian mixture model (GMM) for multimodal policy, together with a novel gradient estimator for the indifferentiability problem, which can be applied in generic off-policy and on-policy DRL algorithms using stochastic policies, e.g., Soft Actor-Critic (SAC) and Proximal Policy Optimisation (PPO). Experimental results testify the advantage of our method over unimodal polices and two different MOE methods, as well as a method of option frameworks, based on the above two types of DRL algorithms, on six MuJoCo tasks. Different gradient estimations for GMM like the reparameterisation trick (Gumbel-Softmax) and the score-ratio trick are also compared with our method. We further empirically demonstrate the distinguishable primitives learned with PMOE and show the benefits of our method in terms of exploration.

Haoqi Yuan, Ruihai Wu, Andrew Zhao et al.

Learning an accurate model of the environment is essential for model-based control tasks. Existing methods in robotic visuomotor control usually learn from data with heavily labelled actions, object entities or locations, which can be demanding in many cases. To cope with this limitation, we propose a method, dubbed DMotion, that trains a forward model from video data only, via disentangling the motion of controllable agent to model the transition dynamics. An object extractor and an interaction learner are trained in an end-to-end manner without supervision. The agent's motions are explicitly represented using spatial transformation matrices containing physical meanings. In the experiments, DMotion achieves superior performance on learning an accurate forward model in a Grid World environment, as well as a more realistic robot control environment in simulation. With the accurate learned forward models, we further demonstrate their usage in model predictive control as an effective approach for robotic manipulations.

Zihan Ding, Ya-Yen Tsai, Wang Wei Lee et al.

Reinforcement Learning (RL) methods have been widely applied for robotic manipulations via sim-to-real transfer, typically with proprioceptive and visual information. However, the incorporation of tactile sensing into RL for contact-rich tasks lacks investigation. In this paper, we model a tactile sensor in simulation and study the effects of its feedback in RL-based robotic control via a zero-shot sim-to-real approach with domain randomization. We demonstrate that learning and controlling with feedback from tactile sensor arrays at the gripper, both in simulation and reality, can enhance grasping stability, which leads to a significant improvement in robotic manipulation performance for a door opening task. In real-world experiments, the door open angle was increased by 45% on average for transferred policies with tactile sensing over those without it.

Ya-Yen Tsai, Hui Xu, Zihan Ding et al.

Reinforcement learning (RL) has demonstrated great success in the past several years. However, most of the scenarios focus on simulated environments. One of the main challenges of transferring the policy learned in a simulated environment to real world, is the discrepancy between the dynamics of the two environments. In prior works, Domain Randomization (DR) has been used to address the reality gap for both robotic locomotion and manipulation tasks. In this paper, we propose Domain Randomization Optimization IDentification (DROID), a novel framework to exploit single-shot human demonstration for identifying the simulator's distribution of dynamics parameters, and apply it to training a policy on a door opening task. Our results show that the proposed framework can identify the difference in dynamics between the simulated and the real worlds, and thus improve policy transfer by optimizing the simulator's randomization ranges. We further illustrate that based on these same identified parameters, our method can generalize the learned policy to different but related tasks.

Zihan Ding, Pablo Hernandez-Leal, Gavin Weiguang Ding et al.

Deep Reinforcement Learning (DRL) has recently achieved significant advances in various domains. However, explaining the policy of RL agents still remains an open problem due to several factors, one being the complexity of explaining neural networks decisions. Recently, a group of works have used decision-tree-based models to learn explainable policies. Soft decision trees (SDTs) and discretized differentiable decision trees (DDTs) have been demonstrated to achieve both good performance and share the benefit of having explainable policies. In this work, we further improve the results for tree-based explainable RL in both performance and explainability. Our proposal, Cascading Decision Trees (CDTs) apply representation learning on the decision path to allow richer expressivity. Empirical results show that in both situations, where CDTs are used as policy function approximators or as imitation learners to explain black-box policies, CDTs can achieve better performances with more succinct and explainable models than SDTs. As a second contribution our study reveals limitations of explaining black-box policies via imitation learning with tree-based explainable models, due to its inherent instability.

Yaodong Yang, Chengdong Ma, Zihan Ding et al.

Tremendous advances have been made in multiagent reinforcement learning (MARL). MARL corresponds to the learning problem in a multiagent system in which multiple agents learn simultaneously. It is an interdisciplinary field of study with a long history that includes game theory, machine learning, stochastic control, psychology, and optimization. Despite great successes in MARL, there is a lack of a self-contained overview of the literature that covers game-theoretic foundations of modern MARL methods and summarizes the recent advances. The majority of existing surveys are outdated and do not fully cover the recent developments since 2010. In this work, we provide a monograph on MARL that covers both the fundamentals and the latest developments on the research frontier. The goal of this monograph is to provide a self-contained assessment of the current state-of-the-art MARL techniques from a game-theoretic perspective. We expect this work to serve as a stepping stone for both new researchers who are about to enter this fast-growing field and experts in the field who want to obtain a panoramic view and identify new directions based on recent advances.

Zihan Ding, Tianyang Yu, Yanhua Huang et al.

Many researchers and developers are exploring for adopting Deep Reinforcement Learning (DRL) techniques in their applications. They however often find such an adoption challenging. Existing DRL libraries provide poor support for prototyping DRL agents (i.e., models), customising the agents, and comparing the performance of DRL agents. As a result, the developers often report low efficiency in developing DRL agents. In this paper, we introduce RLzoo, a new DRL library that aims to make the development of DRL agents efficient. RLzoo provides developers with (i) high-level yet flexible APIs for prototyping DRL agents, and further customising the agents for best performance, (ii) a model zoo where users can import a wide range of DRL agents and easily compare their performance, and (iii) an algorithm that can automatically construct DRL agents with custom components (which are critical to improve agent's performance in custom applications). Evaluation results show that RLzoo can effectively reduce the development cost of DRL agents, while achieving comparable performance with existing DRL libraries.

Eugene Valassakis, Zihan Ding, Edward Johns

Zero-shot sim-to-real transfer of tasks with complex dynamics is a highly challenging and unsolved problem. A number of solutions have been proposed in recent years, but we have found that many works do not present a thorough evaluation in the real world, or underplay the significant engineering effort and task-specific fine tuning that is required to achieve the published results. In this paper, we dive deeper into the sim-to-real transfer challenge, investigate why this is such a difficult problem, and present objective evaluations of a number of transfer methods across a range of real-world tasks. Surprisingly, we found that a method which simply injects random forces into the simulation performs just as well as more complex methods, such as those which randomise the simulator's dynamics parameters, or adapt a policy online using recurrent network architectures.

Zihan Ding, Nathan F. Lepora, Edward Johns

Deep learning and reinforcement learning methods have been shown to enable learning of flexible and complex robot controllers. However, the reliance on large amounts of training data often requires data collection to be carried out in simulation, with a number of sim-to-real transfer methods being developed in recent years. In this paper, we study these techniques for tactile sensing using the TacTip optical tactile sensor, which consists of a deformable tip with a camera observing the positions of pins inside this tip. We designed a model for soft body simulation which was implemented using the Unity physics engine, and trained a neural network to predict the locations and angles of edges when in contact with the sensor. Using domain randomisation techniques for sim-to-real transfer, we show how this framework can be used to accurately predict edges with less than 1 mm prediction error in real-world testing, without any real-world data at all.

Zihan Ding, Xiao-Yang Liu, Miao Yin et al.

Deep generative models have been successfully applied to many applications. However, existing works experience limitations when generating large images (the literature usually generates small images, e.g. 32 * 32 or 128 * 128). In this paper, we propose a novel scheme, called deep tensor adversarial generative nets (TGAN), that generates large high-quality images by exploring tensor structures. Essentially, the adversarial process of TGAN takes place in a tensor space. First, we impose tensor structures for concise image representation, which is superior in capturing the pixel proximity information and the spatial patterns of elementary objects in images, over the vectorization preprocess in existing works. Secondly, we propose TGAN that integrates deep convolutional generative adversarial networks and tensor super-resolution in a cascading manner, to generate high-quality images from random distributions. More specifically, we design a tensor super-resolution process that consists of tensor dictionary learning and tensor coefficients learning. Finally, on three datasets, the proposed TGAN generates images with more realistic textures, compared with state-of-the-art adversarial autoencoders. The size of the generated images is increased by over 8.5 times, namely 374 * 374 in PASCAL2.

Xiao-Yang Liu, Zihan Ding, Sem Borst et al.

Intelligent Transportation Systems (ITSs) are envisioned to play a critical role in improving traffic flow and reducing congestion, which is a pervasive issue impacting urban areas around the globe. Rapidly advancing vehicular communication and edge cloud computation technologies provide key enablers for smart traffic management. However, operating viable real-time actuation mechanisms on a practically relevant scale involves formidable challenges, e.g., policy iteration and conventional Reinforcement Learning (RL) techniques suffer from poor scalability due to state space explosion. Motivated by these issues, we explore the potential for Deep Q-Networks (DQN) to optimize traffic light control policies. As an initial benchmark, we establish that the DQN algorithms yield the "thresholding" policy in a single-intersection. Next, we examine the scalability properties of DQN algorithms and their performance in a linear network topology with several intersections along a main artery. We demonstrate that DQN algorithms produce intelligent behavior, such as the emergence of "greenwave" patterns, reflecting their ability to learn favorable traffic light actuations.