Haifeng Zhang

Muning Wen, Runji Lin, Hanjing Wang et al.

Transformer architectures have facilitated the development of large-scale and general-purpose sequence models for prediction tasks in natural language processing and computer vision, e.g., GPT-3 and Swin Transformer. Although originally designed for prediction problems, it is natural to inquire about their suitability for sequential decision-making and reinforcement learning problems, which are typically beset by long-standing issues involving sample efficiency, credit assignment, and partial observability. In recent years, sequence models, especially the Transformer, have attracted increasing interest in the RL communities, spawning numerous approaches with notable effectiveness and generalizability. This survey presents a comprehensive overview of recent works aimed at solving sequential decision-making tasks with sequence models such as the Transformer, by discussing the connection between sequential decision-making and sequence modeling, and categorizing them based on the way they utilize the Transformer. Moreover, this paper puts forth various potential avenues for future research intending to improve the effectiveness of large sequence models for sequential decision-making, encompassing theoretical foundations, network architectures, algorithms, and efficient training systems. As this article has been accepted by the Frontiers of Computer Science, here is an early version, and the most up-to-date version can be found at https://journal.hep.com.cn/fcs/EN/10.1007/s11704-023-2689-5

Runji Lin, Ye Li, Xidong Feng et al.

The pretrain-finetuning paradigm in large-scale sequence models has made significant progress in natural language processing and computer vision tasks. However, such a paradigm is still hindered by several challenges in Reinforcement Learning (RL), including the lack of self-supervised pretraining algorithms based on offline data and efficient fine-tuning/prompt-tuning over unseen downstream tasks. In this work, we explore how prompts can improve sequence modeling-based offline reinforcement learning (offline-RL) algorithms. Firstly, we propose prompt tuning for offline RL, where a context vector sequence is concatenated with the input to guide the conditional policy generation. As such, we can pretrain a model on the offline dataset with self-supervised loss and learn a prompt to guide the policy towards desired actions. Secondly, we extend our framework to Meta-RL settings and propose Contextual Meta Transformer (CMT); CMT leverages the context among different tasks as the prompt to improve generalization on unseen tasks. We conduct extensive experiments across three different offline-RL settings: offline single-agent RL on the D4RL dataset, offline Meta-RL on the MuJoCo benchmark, and offline MARL on the SMAC benchmark. Superior results validate the strong performance, and generality of our methods.

Yongcheng Zeng, Guoqing Liu, Weiyu Ma et al.

Fine-tuning pre-trained Large Language Models (LLMs) is essential to align them with human values and intentions. This process often utilizes methods like pairwise comparisons and KL divergence against a reference LLM, focusing on the evaluation of full answers generated by the models. However, the generation of these responses occurs in a token level, following a sequential, auto-regressive fashion. In this paper, we introduce Token-level Direct Preference Optimization (TDPO), a novel approach to align LLMs with human preferences by optimizing policy at the token level. Unlike previous methods, which face challenges in divergence efficiency, TDPO incorporates forward KL divergence constraints for each token, improving alignment and diversity. Utilizing the Bradley-Terry model for a token-based reward system, TDPO enhances the regulation of KL divergence, while preserving simplicity without the need for explicit reward modeling. Experimental results across various text tasks demonstrate TDPO's superior performance in balancing alignment with generation diversity. Notably, fine-tuning with TDPO strikes a better balance than DPO in the controlled sentiment generation and single-turn dialogue datasets, and significantly improves the quality of generated responses compared to both DPO and PPO-based RLHF methods. Our code is open-sourced at https://github.com/Vance0124/Token-level-Direct-Preference-Optimization.

Haifeng Zhang, Qinghui He, Xiuli Bi et al.

In recent years, the rapid development of generative artificial intelligence technology has significantly lowered the barrier to creating high-quality fake images, posing a serious challenge to information authenticity and credibility. Existing generated image detection methods typically enhance generalization through model architecture or network design. However, their generalization performance remains susceptible to data bias, as the training data may drive models to fit specific generative patterns and content rather than the common features shared by images from different generative models (asymmetric bias learning). To address this issue, we propose a Multi-dimensional Adversarial Feature Learning (MAFL) framework. The framework adopts a pretrained multimodal image encoder as the feature extraction backbone, constructs a real-fake feature learning network, and designs an adversarial bias-learning branch equipped with a multi-dimensional adversarial loss, forming an adversarial training mechanism between authenticity-discriminative feature learning and bias feature learning. By suppressing generation-pattern and content biases, MAFL guides the model to focus on the generative features shared across different generative models, thereby effectively capturing the fundamental differences between real and generated images, enhancing cross-model generalization, and substantially reducing the reliance on large-scale training data. Through extensive experimental validation, our method outperforms existing state-of-the-art approaches by 10.89% in accuracy and 8.57% in Average Precision (AP). Notably, even when trained with only 320 images, it can still achieve over 80% detection accuracy on public datasets.

Weiyu Ma, Qirui Mi, Yongcheng Zeng et al.

StarCraft II is a challenging benchmark for AI agents due to the necessity of both precise micro level operations and strategic macro awareness. Previous works, such as Alphastar and SCC, achieve impressive performance on tackling StarCraft II , however, still exhibit deficiencies in long term strategic planning and strategy interpretability. Emerging large language model (LLM) agents, such as Voyage and MetaGPT, presents the immense potential in solving intricate tasks. Motivated by this, we aim to validate the capabilities of LLMs on StarCraft II, a highly complex RTS game.To conveniently take full advantage of LLMs` reasoning abilities, we first develop textual StratCraft II environment, called TextStarCraft II, which LLM agent can interact. Secondly, we propose a Chain of Summarization method, including single frame summarization for processing raw observations and multi frame summarization for analyzing game information, providing command recommendations, and generating strategic decisions. Our experiment consists of two parts: first, an evaluation by human experts, which includes assessing the LLMs`s mastery of StarCraft II knowledge and the performance of LLM agents in the game; second, the in game performance of LLM agents, encompassing aspects like win rate and the impact of Chain of Summarization.Experiment results demonstrate that: 1. LLMs possess the relevant knowledge and complex planning abilities needed to address StarCraft II scenarios; 2. Human experts consider the performance of LLM agents to be close to that of an average player who has played StarCraft II for eight years; 3. LLM agents are capable of defeating the built in AI at the Harder(Lv5) difficulty level. We have open sourced the code and released demo videos of LLM agent playing StarCraft II.

Xue Yan, Yan Song, Xinyu Cui et al.

Large language models (LLMs) demonstrate their promise in tackling complicated practical challenges by combining action-based policies with chain of thought (CoT) reasoning. Having high-quality prompts on hand, however, is vital to the framework's effectiveness. Currently, these prompts are handcrafted utilising extensive human labor, resulting in CoT policies that frequently fail to generalise. Human intervention is also required to develop grounding functions that ensure low-level controllers appropriately process CoT reasoning. In this paper, we propose a comprehensive training framework for complex task-solving, incorporating human prior knowledge into the learning of action policies. To that purpose, we offer a new leader-follower bilevel framework that is capable of learning to ask relevant questions (prompts) and subsequently undertaking reasoning to guide the learning of actions. The prompt policy is employed to make introspective revisions based on historical findings, leading the CoT process to consider the anticipated goals and generate outputs that lead to decisive, high-performing actions. The action policy subsequently learns to comprehend and integrate the CoT outputs to take actions. Our empirical data reveal that our framework outperforms leading methods in $5$ decision-making tasks such as Overcooked and FourRoom.

Qirui Mi, Zhijian Ma, Mengyue Yang et al.

LLM-driven agents demonstrate strong performance in sequential decision-making but often rely on on-the-fly reasoning, re-deriving solutions even in recurring scenarios. This insufficient experience reuse leads to computational redundancy and execution instability. To bridge this gap, we propose ProcMEM, a framework that enables agents to autonomously learn procedural memory from interaction experiences without parameter updates. By formalizing a Skill-MDP, ProcMEM transforms passive episodic narratives into executable Skills defined by activation, execution, and termination conditions to ensure executability. To achieve reliable reusability without capability degradation, we introduce Non-Parametric PPO, which leverages semantic gradients for high-quality candidate generation and a PPO Gate for robust Skill verification. Through score-based maintenance, ProcMEM sustains compact, high-quality procedural memory. Experimental results across in-domain, cross-task, and cross-agent scenarios demonstrate that ProcMEM achieves superior reuse rates and significant performance gains with extreme memory compression. Visualized evolutionary trajectories and Skill distributions further reveal how ProcMEM transparently accumulates, refines, and reuses procedural knowledge to facilitate long-term autonomy.

Siyu Xia, Zekun Xu, Jiajun Chai et al.

Large Language Models (LLMs) based agents have demonstrated remarkable potential in autonomous task-solving across complex, open-ended environments. A promising approach for improving the reasoning capabilities of LLM agents is to better utilize prior experiences in guiding current decisions. However, LLMs acquire experience either through implicit memory via training, which suffers from catastrophic forgetting and limited interpretability, or explicit memory via prompting, which lacks adaptability. In this paper, we introduce a novel agent-centric, trainable, multi-layered graph memory framework and evaluate how context memory enhances the ability of LLMs to utilize parametric information. The graph abstracts raw agent trajectories into structured decision paths in a state machine and further distills them into high-level, human-interpretable strategic meta-cognition. In order to make memory adaptable, we propose a reinforcement-based weight optimization procedure that estimates the empirical utility of each meta-cognition based on reward feedback from downstream tasks. These optimized strategies are then dynamically integrated into the LLM agent's training loop through meta-cognitive prompting. Empirically, the learnable graph memory delivers robust generalization, improves LLM agents' strategic reasoning performance, and provides consistent benefits during Reinforcement Learning (RL) training.

Cheng Deng, Luoyang Sun, Jiwen Jiang et al.

While scaling laws have been continuously validated in large language models (LLMs) with increasing model parameters, the inherent tension between the inference demands of LLMs and the limited resources of edge devices poses a critical challenge to the development of edge intelligence. Recently, numerous small language models have emerged, aiming to distill the capabilities of LLMs into smaller footprints. However, these models often retain the fundamental architectural principles of their larger counterparts, still imposing considerable strain on the storage and bandwidth capacities of edge devices. In this paper, we introduce the PLM, a Peripheral Language Model, developed through a co-design process that jointly optimizes model architecture and edge system constraints. The PLM utilizes a Multi-head Latent Attention mechanism and employs the squared ReLU activation function to encourage sparsity, thereby reducing peak memory footprint during inference. During training, we collect and reorganize open-source datasets, implement a multi-phase training strategy, and empirically investigate the Warmup-Stable-Decay-Constant (WSDC) learning rate scheduler. Additionally, we incorporate Reinforcement Learning from Human Feedback (RLHF) by adopting the ARIES preference learning approach. Following a two-phase SFT process, this method yields performance gains of 2% in general tasks, 9% in the GSM8K task, and 11% in coding tasks. In addition to its novel architecture, evaluation results demonstrate that PLM outperforms existing small language models trained on publicly available data while maintaining the lowest number of activated parameters. Furthermore, deployment across various edge devices, including consumer-grade GPUs, mobile phones, and Raspberry Pis, validates PLM's suitability for peripheral applications. The PLM series models are publicly available at https://github.com/plm-team/PLM.

Xinyu Cui, Xuanfa Jin, Xue Yan et al.

Autonomous navigation in congested maritime environments is a critical capability for a wide range of real-world applications. However, it remains an unresolved challenge due to complex vessel interactions and significant environmental uncertainties. Existing methods often fail in practical deployment due to a substantial sim-to-real gap, which stems from imprecise simulation, inadequate situational awareness, and unsafe exploration strategies. To address these, we propose \textbf{Sim2Sea}, a comprehensive framework designed to bridge simulation and real-world execution. Sim2Sea advances in three key aspects. First, we develop a GPU-accelerated parallel simulator for scalable and accurate maritime scenario simulation. Second, we design a dual-stream spatiotemporal policy that handles complex dynamics and multi-modal perception, augmented with a velocity-obstacle-guided action masking mechanism to ensure safe and efficient exploration. Finally, a targeted domain randomization scheme helps bridge the sim-to-real gap. Simulation results demonstrate that our method achieves faster convergence and safer trajectories than established baselines. In addition, our policy trained purely in simulation successfully transfers zero-shot to a 17-ton unmanned vessel operating in real-world congested waters. These results validate the effectiveness of Sim2Sea in achieving reliable sim-to-real transfer for practical autonomous maritime navigation.

Siting Wang, Minnan Pei, Luoyang Sun et al.

Humans can directly imagine and manipulate visual images in their minds, a capability known as spatial visualization. While multi-modal Large Language Models (MLLMs) support imagination-based reasoning, spatial visualization remains insufficiently evaluated, typically embedded within broader mathematical and logical assessments. Existing evaluations often rely on IQ tests or math competitions that may overlap with training data, compromising assessment reliability. To this end, we introduce SpatialViz-Bench, a comprehensive multi-modal benchmark for spatial visualization with 12 tasks across 4 sub-abilities, comprising 1,180 automatically generated problems. Our evaluation of 33 state-of-the-art MLLMs not only reveals wide performance variations and demonstrates the benchmark's strong discriminative power, but also uncovers counter-intuitive findings: models show difficulty perception misaligned with human intuition, exhibit dramatic 2Dto-3D performance cliffs, default to formulaic derivation over visualization, and paradoxically suffer performance degradation from Chain-of-Thought prompting in open-source models. Through statistical and qualitative analysis of error types, SpatialViz-Bench demonstrates that state-of-the-art MLLMs continue to exhibit deficiencies in spatial visualization tasks, thereby addressing a significant lacuna in the field. The benchmark data and evaluation code are publicly available.

Yan Song, He Jiang, Zheng Tian et al.

Few multi-agent reinforcement learning (MARL) research on Google Research Football (GRF) focus on the 11v11 multi-agent full-game scenario and to the best of our knowledge, no open benchmark on this scenario has been released to the public. In this work, we fill the gap by providing a population-based MARL training pipeline and hyperparameter settings on multi-agent football scenario that outperforms the bot with difficulty 1.0 from scratch within 2 million steps. Our experiments serve as a reference for the expected performance of Independent Proximal Policy Optimization (IPPO), a state-of-the-art multi-agent reinforcement learning algorithm where each agent tries to maximize its own policy independently across various training configurations. Meanwhile, we open-source our training framework Light-MALib which extends the MALib codebase by distributed and asynchronized implementation with additional analytical tools for football games. Finally, we provide guidance for building strong football AI with population-based training and release diverse pretrained policies for benchmarking. The goal is to provide the community with a head start for whoever experiment their works on GRF and a simple-to-use population-based training framework for further improving their agents through self-play. The implementation is available at https://github.com/Shanghai-Digital-Brain-Laboratory/DB-Football.

Siting Wang, Xiaofeng Wang, Zheng Zhu et al.

Flow-based vision-language-action (VLA) models excel in embodied control but suffer from intractable likelihoods during multi-step sampling, hindering online reinforcement learning. We propose \textbf{\textit{$\boldsymbolπ$-StepNFT}} (Step-wise Negative-aware Fine-Tuning), a critic-and-likelihood-free framework that requires only a single forward pass per optimization step and eliminates auxiliary value networks. We identify that wider exploration spaces necessitate finer-grained, step-wise guidance for alignment. Empirically, $π$-StepNFT unlocks latent potential on LIBERO with competitive few-shot robustness. Moreover, it achieves superior generalization on ManiSkill, outperforming value-based baselines in OOD scenarios by preventing overfitting to multimodal features. This property offers a scalable solution promising for complex real-world applications.

Yongcheng Zeng, Xinyu Cui, Xuanfa Jin et al.

Self-Refinement refers to a model's ability to revise its own responses to produce improved outputs. This capability can also serve as a fundamental mechanism for Self-Improvement, for example, by reconstructing datasets with refined results to enhance intrinsic model performance. However, our comprehensive experiments reveal that large language models (LLMs) show no clear evidence of inherent Self-Refinement and may even experience response quality degradation after Self-Refinement. To address this issue, we propose EVOLVE, a simple and effective framework for eliciting and tracking the evolution of Self-Refinement through iterative training. We first explore optimization methods during training to activate the model's Self-Refinement capability. Then, at inference, we investigate various generation strategies to further enhance and utilize Self-Refinement while supplying the necessary data for training. Through synergistic optimization of training and inference stages, we continually evolve the model's Self-Refinement ability, enabling it to better refine its own responses. Moreover, we demonstrate the potential of leveraging Self-Refinement to achieve broader Self-Improvement of intrinsic model abilities. Experiments show that the evolved Self-Refinement ability enables the Llama-3.1-8B base model to surpass GPT-4o, achieving 62.3% length-controlled and 63.3% raw win rates on AlpacaEval 2, and 50.3% on Arena-Hard. It also generalizes effectively to out-of-domain reasoning tasks, improving performance on mathematical reasoning benchmarks such as GSM8K and MATH.

Chengyuan Zhu, Yiyuan Yang, Kaixiang Yang et al.

The application of artificial intelligence technology has greatly enhanced and fortified the safety of energy pipelines, particularly in safeguarding against external threats. The predominant methods involve the integration of intelligent sensors to detect external vibration, enabling the identification of event types and locations, thereby replacing manual detection methods. However, practical implementation has exposed a limitation in current methods - their constrained ability to accurately discern the spatial dimensions of external signals, which complicates the authentication of threat events. Our research endeavors to overcome the above issues by harnessing deep learning techniques to achieve a more fine-grained recognition and localization process. This refinement is crucial in effectively identifying genuine threats to pipelines, thus enhancing the safety of energy transportation. This paper proposes a radial threat estimation method for energy pipelines based on distributed optical fiber sensing technology. Specifically, we introduce a continuous multi-view and multi-domain feature fusion methodology to extract comprehensive signal features and construct a threat estimation and recognition network. The utilization of collected acoustic signal data is optimized, and the underlying principle is elucidated. Moreover, we incorporate the concept of transfer learning through a pre-trained model, enhancing both recognition accuracy and training efficiency. Empirical evidence gathered from real-world scenarios underscores the efficacy of our method, notably in its substantial reduction of false alarms and remarkable gains in recognition accuracy. More generally, our method exhibits versatility and can be extrapolated to a broader spectrum of recognition tasks and scenarios.

Qirui Mi, Qipeng Yang, Zijun Fan et al.

Artificial intelligence (AI) has become a powerful tool for economic research, enabling large-scale simulation and policy optimization. However, applying AI effectively requires simulation platforms for scalable training and evaluation-yet existing environments remain limited to simplified, narrowly scoped tasks, falling short of capturing complex economic challenges such as demographic shifts, multi-government coordination, and large-scale agent interactions. To address this gap, we introduce EconGym, a scalable and modular testbed that connects diverse economic tasks with AI algorithms. Grounded in rigorous economic modeling, EconGym implements 11 heterogeneous role types (e.g., households, firms, banks, governments), their interaction mechanisms, and agent models with well-defined observations, actions, and rewards. Users can flexibly compose economic roles with diverse agent algorithms to simulate rich multi-agent trajectories across 25+ economic tasks for AI-driven policy learning and analysis. Experiments show that EconGym supports diverse and cross-domain tasks-such as coordinating fiscal, pension, and monetary policies-and enables benchmarking across AI, economic methods, and hybrids. Results indicate that richer task composition and algorithm diversity expand the policy space, while AI agents guided by classical economic methods perform best in complex settings. EconGym also scales to 10k agents with high realism and efficiency.

Yue Deng, Weiyu Ma, Yuxin Fan et al.

StarCraft Multi-Agent Challenge (SMAC) has been one of the most commonly used experimental environments in multi-agent reinforcement learning (MARL), where the specific task is to control a set number of allied units to defeat enemy forces. Traditional MARL algorithms often require interacting with the environment for millions of steps to train a parametric model, of which the resulting policies are typically non-interpretable with weak transferability. In this paper, we introduce SMAC-R1 which is based on the Qwen2.5-7B-Base LLM distilled from DeepSeek-Coder-v2.5-236B. Similar to online reinforcement learning after behavior cloning in offline learning process, in our pipeline, agents leverage the DeepSeek LLM to generate decision tree code by providing task descriptions, and the agents are further self-reflected using feedback from the rewards provided by the environment. Based on that, we augment the generated scripts to fine-tune a small LLM, Qwen2.5-7B-Base, to distill the decision-making ability via Supervised Fine-Tuning (SFT) and enhance the script generation ability by the Group Relative Policy Optimization (GRPO) algorithm. We conduct experiments in the original 23 SMAC tasks and 10 newly-designed tasks to demonstrate that our method can produce high-quality, interpretable decision trees with minimal environmental exploration. Moreover, these scripts exhibit strong transferability, successfully applying to homogeneous SMAC environments without modification. We believe this approach offers a new direction for solving decision-making tasks and domain-specific LLM training pipelines in the future.

Qirui Mi, Mengyue Yang, Xiangning Yu et al.

Simulating collective decision-making involves more than aggregating individual behaviors; it emerges from dynamic interactions among individuals. While large language models (LLMs) offer strong potential for social simulation, achieving quantitative alignment with real-world data remains a key challenge. To bridge this gap, we propose the Mean-Field LLM (MF-LLM) framework, the first to incorporate mean field theory into LLM-based social simulation. MF-LLM models bidirectional interactions between individuals and the population through an iterative process, generating population signals to guide individual decisions, which in turn update the signals. This interplay produces coherent trajectories of collective behavior. To improve alignment with real-world data, we introduce IB-Tune, a novel fine-tuning method inspired by the Information Bottleneck principle, which retains population signals most predictive of future actions while filtering redundant history. Evaluated on a real-world social dataset, MF-LLM reduces KL divergence to human population distributions by 47\% compared to non-mean-field baselines, enabling accurate trend forecasting and effective intervention planning. Generalizing across 7 domains and 4 LLM backbones, MF-LLM provides a scalable, high-fidelity foundation for social simulation.

Weiyu Ma, Dongyu Xu, Shu Lin et al.

We present Adaptive Command, a novel framework integrating large language models (LLMs) with behavior trees for real-time strategic decision-making in StarCraft II. Our system focuses on enhancing human-AI collaboration in complex, dynamic environments through natural language interactions. The framework comprises: (1) an LLM-based strategic advisor, (2) a behavior tree for action execution, and (3) a natural language interface with speech capabilities. User studies demonstrate significant improvements in player decision-making and strategic adaptability, particularly benefiting novice players and those with disabilities. This work contributes to the field of real-time human-AI collaborative decision-making, offering insights applicable beyond RTS games to various complex decision-making scenarios.

Zhiyu Zhao, Haoxuan Li, Haifeng Zhang et al. · pku

When building a world model, a common assumption is that the environment has a single, unchanging underlying causal rule, like applying Newton's laws to every situation. In reality, what appears as a drifting causal mechanism is often the manifestation of a fixed underlying mechanism seen through a narrow observational window. This brings about a problem that, when building a world model, even subtle shifts in policy or environment states can alter the very observed causal mechanisms. In this work, we introduce the \textbf{Meta-Causal Graph} as world models, a minimal unified representation that efficiently encodes the transformation rules governing how causal structures shift across different latent world states. A single Meta-Causal Graph is composed of multiple causal subgraphs, each triggered by meta state, which is in the latent state space. Building on this representation, we introduce a \textbf{Causality-Seeking Agent} whose objectives are to (1) identify the meta states that trigger each subgraph, (2) discover the corresponding causal relationships by agent curiosity-driven intervention policy, and (3) iteratively refine the Meta-Causal Graph through ongoing curiosity-driven exploration and agent experiences. Experiments on both synthetic tasks and a challenging robot arm manipulation task demonstrate that our method robustly captures shifts in causal dynamics and generalizes effectively to previously unseen contexts.

Luoyang Sun, Cheng Deng, Jiwen Jiang et al.

Attention mechanisms underpin the success of large language models (LLMs), yet their substantial computational and memory overhead poses challenges for optimizing efficiency and performance. A critical bottleneck arises as KV cache and attention computations scale rapidly with text length, challenging deployment on hardware with limited computational and memory resources. We observe that attention mechanisms exhibit substantial redundancy, since the KV cache can be significantly compressed and attention maps across heads display high similarity, revealing that much of the computation and storage is unnecessary. Leveraging these insights, we propose \textbf{G}rouped-Head Laten\textbf{T} \textbf{A}ttention (GTA), a novel attention mechanism that reduces memory usage and computational complexity while maintaining performance. GTA comprises two components: (1) a shared attention map mechanism that reuses attention scores across multiple heads, decreasing the key cache size; and (2) a nonlinear value decoder with learned projections that compresses the value cache into a latent space, further cutting memory needs. GTA cuts attention computation FLOPs by up to \emph{62.5\%} versus Grouped-Query Attention and shrink the KV cache by up to \emph{70\%}, all while avoiding the extra overhead of Multi-Head Latent Attention to improve LLM deployment efficiency. Consequently, GTA models achieve a \emph{2x} increase in end-to-end inference speed, with prefill benefiting from reduced computational cost and decoding benefiting from the smaller cache footprint.

Luoyang Sun, Jiwen Jiang, Yifeng Ding et al.

Vision-Language-Action Models (VLAs) have emerged as a key paradigm of Physical AI and are increasingly deployed in autonomous vehicles, robots, and smart spaces. In these resource-constrained on-device settings, selecting an appropriate large language model (LLM) backbone is a critical challenge: models must balance accuracy with strict inference latency and hardware efficiency constraints. This makes hardware-software co-design a game-changing requirement for on-device LLM deployment, where each hardware platform demands a tailored architectural solution. We propose a hardware co-design law that jointly captures model accuracy and inference performance. Specifically, we model training loss as an explicit function of architectural hyperparameters and characterise inference latency via roofline modelling. We empirically evaluate 1,942 candidate architectures on NVIDIA Jetson Orin, training 170 selected models for 10B tokens each to fit a scaling law relating architecture to training loss. By coupling this scaling law with latency modelling, we establish a direct accuracy-latency correspondence and identify the Pareto frontier for hardware co-designed LLMs. We further formulate architecture search as a joint optimisation over precision and performance, deriving feasible design regions under industrial hardware and application budgets. Our approach reduces architecture selection from months to days. At the same latency as Qwen2.5-0.5B on the target hardware, our co-designed architecture achieves 19.42% lower perplexity on WikiText-2. To our knowledge, this is the first principled and operational framework for hardware co-design scaling laws in on-device LLM deployment. We will make the code and related checkpoints publicly available.

Heyang Ma, Qirui Mi, Qipeng Yang et al.

Economic decision-making depends not only on structured signals such as prices and taxes, but also on unstructured language, including peer dialogue and media narratives. While multi-agent reinforcement learning (MARL) has shown promise in optimizing economic decisions, it struggles with the semantic ambiguity and contextual richness of language. We propose LAMP (Language-Augmented Multi-Agent Policy), a framework that integrates language into economic decision-making and narrows the gap to real-world settings. LAMP follows a Think-Speak-Decide pipeline: (1) Think interprets numerical observations to extract short-term shocks and long-term trends, caching high-value reasoning trajectories; (2) Speak crafts and exchanges strategic messages based on reasoning, updating beliefs by parsing peer communications; and (3) Decide fuses numerical data, reasoning, and reflections into a MARL policy to optimize language-augmented decision-making. Experiments in economic simulation show that LAMP outperforms both MARL and LLM-only baselines in cumulative return (+63.5%, +34.0%), robustness (+18.8%, +59.4%), and interpretability. These results demonstrate the potential of language-augmented policies to deliver more effective and robust economic strategies.

Zhongwei Yu, Wannian Xia, Xue Yan et al.

Advanced large language models (LLMs) frequently reflect in reasoning chain-of-thoughts (CoTs), where they self-verify the correctness of current solutions and explore alternatives. However, given recent findings that LLMs detect limited errors in CoTs, how reflection contributes to empirical improvements remains unclear. To analyze this issue, in this paper, we present a minimalistic reasoning framework to support basic self-verifying reflection for small transformers without natural language, which ensures analytic clarity and reduces the cost of comprehensive experiments. Theoretically, we prove that self-verifying reflection guarantees improvements if verification errors are properly bounded. Experimentally, we show that tiny transformers, with only a few million parameters, benefit from self-verification in both training and reflective execution, reaching remarkable LLM-level performance in integer multiplication and Sudoku. Similar to LLM results, we find that reinforcement learning (RL) improves in-distribution performance and incentivizes frequent reflection for tiny transformers, yet RL mainly optimizes shallow statistical patterns without faithfully reducing verification errors. In conclusion, integrating generative transformers with discriminative verification inherently facilitates CoT reasoning, regardless of scaling and natural language.

Yongcheng Zeng, Zexu Sun, Bokai Ji et al.

Curriculum learning plays a crucial role in enhancing the training efficiency of large language models (LLMs) on reasoning tasks. However, existing methods often fail to adequately account for variations in prompt difficulty or rely on simplistic filtering mechanisms to select prompt datasets within a narrow criterion range, resulting in significant computational waste. In this work, we approach the problem from the perspective of reinforcement learning gradient optimization, offering a systematic and theoretical investigation into how to improve the training efficiency of LLMs. We identify two key factors influencing training efficiency: the selection of training prompts and the allocation of rollout quantities across different prompts. Our theoretical analysis reveals that the sampling distribution of prompts dictates the convergence rate of gradient descent, while the allocation of the rollout quantity influences the consistency and stability of overall gradient updates. Based on these insights, we propose CurES, an efficient training method that accelerates convergence and employs Bayesian posterior estimation to minimize computational overhead. Experiments demonstrate that our CurES outperforms Group Relative Policy Optimization (GRPO) by \textbf{+3.30} points and \textbf{+4.82} points with 1.5B and 7B models, respectively. Additionally, CurES exhibits faster convergence compared to baselines, including GRPO.

Xue Yan, Zijing Ou, Mengyue Yang et al.

Large language models (LLMs) have emerged as effective action policies for sequential decision-making (SDM) tasks due to their extensive prior knowledge. However, this broad yet general knowledge is often insufficient for specific decision-making tasks with limited task-related data, making it challenging to efficiently adapt LLMs to specific SDM tasks. To address this challenge, we propose a memory-driven self-improvement framework that combines LLM general prior knowledge with a compact memory of domain-specific experiences. Memory retains past interactions and associated Q-values, thereby capturing decision-relevant knowledge that facilitates accurate value estimation and informs the LLM prior refinement. The refined LLM prior, in turn, generates higher-reward trajectories that further enrich memory, forming a natural self-improvement framework where memory and LLM prior mutually reinforce each other. Experiments show that our memory-driven approach significantly outperforms both traditional RL and LLM-based baselines, e.g., improving performance by over 40\% on in-distribution tasks and over 75\% when generalized to unseen tasks in ALFWorld.

Ning Yang, Pengyu Wang, Guoqing Liu et al.

Safe Reinforcement Learning (RL) often faces significant issues such as constraint violations and instability, necessitating the use of constrained policy optimization, which seeks optimal policies while ensuring adherence to specific constraints like safety. Typically, constrained optimization problems are addressed by the Lagrangian method, a post-violation remedial approach that may result in oscillations and overshoots. Motivated by this, we propose a novel method named Proactive Constrained Policy Optimization (PCPO) that incorporates a preemptive penalty mechanism. This mechanism integrates barrier items into the objective function as the policy nears the boundary, imposing a cost. Meanwhile, we introduce a constraint-aware intrinsic reward to guide boundary-aware exploration, which is activated only when the policy approaches the constraint boundary. We establish theoretical upper and lower bounds for the duality gap and the performance of the PCPO update, shedding light on the method's convergence characteristics. Additionally, to enhance the optimization performance, we adopt a policy iteration approach. An interesting finding is that PCPO demonstrates significant stability in experiments. Experimental results indicate that the PCPO framework provides a robust solution for policy optimization under constraints, with important implications for future research and practical applications.

Weiyu Ma, Jiwen Jiang, Haobo Fu et al.

We present an adapter-based approach for tactical conditioning of StarCraft II AI agents. Current agents, while powerful, lack the ability to adapt their strategies based on high-level tactical directives. Our method freezes a pre-trained policy network (DI-Star) and attaches lightweight adapter modules to each action head, conditioned on a tactical tensor that encodes strategic preferences. By training these adapters with KL divergence constraints, we ensure the policy maintains core competencies while exhibiting tactical variations. Experimental results show our approach successfully modulates agent behavior across tactical dimensions including aggression, expansion patterns, and technology preferences, while maintaining competitive performance. Our method enables flexible tactical control with minimal computational overhead, offering practical strategy customization for complex real-time strategy games.

Zhaoxing Li, Xiaoming Zhang, Haifeng Zhang et al.

The integration of Large Language Models (LLMs) with Graph Neural Networks (GNNs) has recently been explored to enhance the capabilities of Text Attribute Graphs (TAGs). Most existing methods feed textual descriptions of the graph structure or neighbouring nodes' text directly into LLMs. However, these approaches often cause LLMs to treat structural information simply as general contextual text, thus limiting their effectiveness in graph-related tasks. In this paper, we introduce LanSAGNN (Language Semantic Anisotropic Graph Neural Network), a framework that extends the concept of anisotropic GNNs to the natural language level. This model leverages LLMs to extract tailor-made semantic information for node pairs, effectively capturing the unique interactions within node relationships. In addition, we propose an efficient dual-layer LLMs finetuning architecture to better align LLMs' outputs with graph tasks. Experimental results demonstrate that LanSAGNN significantly enhances existing LLM-based methods without increasing complexity while also exhibiting strong robustness against interference.

Zhiyu Zhao, Haifeng Zhang

The Control as Inference (CAI) framework has successfully transformed single-agent reinforcement learning (RL) by reframing control tasks as probabilistic inference problems. However, the extension of CAI to multi-agent, general-sum stochastic games (SGs) remains underexplored, particularly in decentralized settings where agents operate independently without centralized coordination. In this paper, we propose a novel variational inference framework tailored to decentralized multi-agent systems. Our framework addresses the challenges posed by non-stationarity and unaligned agent objectives, proving that the resulting policies form an $ε$-Nash equilibrium. Additionally, we demonstrate theoretical convergence guarantees for the proposed decentralized algorithms. Leveraging this framework, we instantiate multiple algorithms to solve for Nash equilibrium, mean-field Nash equilibrium, and correlated equilibrium, with rigorous theoretical convergence analysis.

Qirui Mi, Zhiyu Zhao, Chengdong Ma et al.

Macroeconomic outcomes emerge from individuals' decisions, making it essential to model how agents interact with macro policy via consumption, investment, and labor choices. We formulate this as a dynamic Stackelberg game: the government (leader) sets policies, and agents (followers) respond by optimizing their behavior over time. Unlike static models, this dynamic formulation captures temporal dependencies and strategic feedback critical to policy design. However, as the number of agents increases, explicitly simulating all agent-agent and agent-government interactions becomes computationally infeasible. To address this, we propose the Dynamic Stackelberg Mean Field Game (DSMFG) framework, which approximates these complex interactions via agent-population and government-population couplings. This approximation preserves individual-level feedback while ensuring scalability, enabling DSMFG to jointly model three core features of real-world policymaking: dynamic feedback, asymmetry, and large scale. We further introduce Stackelberg Mean Field Reinforcement Learning (SMFRL), a data-driven algorithm that learns the leader's optimal policies while maintaining personalized responses for individual agents. Empirically, we validate our approach in a large-scale simulated economy, where it scales to 1,000 agents (vs. 100 in prior work) and achieves a fourfold increase in GDP over classical economic methods and a nineteenfold improvement over the static 2022 U.S. federal income tax policy.

Xue Yan, Yali Du, Binxin Ru et al.

The Elo rating system is widely adopted to evaluate the skills of (chess) game and sports players. Recently it has been also integrated into machine learning algorithms in evaluating the performance of computerised AI agents. However, an accurate estimation of the Elo rating (for the top players) often requires many rounds of competitions, which can be expensive to carry out. In this paper, to improve the sample efficiency of the Elo evaluation (for top players), we propose an efficient online match scheduling algorithm. Specifically, we identify and match the top players through a dueling bandits framework and tailor the bandit algorithm to the gradient-based update of Elo. We show that it reduces the per-step memory and time complexity to constant, compared to the traditional likelihood maximization approaches requiring $O(t)$ time. Our algorithm has a regret guarantee of $\tilde{O}(\sqrt{T})$, sublinear in the number of competition rounds and has been extended to the multidimensional Elo ratings for handling intransitive games. We empirically demonstrate that our method achieves superior convergence speed and time efficiency on a variety of gaming tasks.

Xidong Feng, Bo Liu, Jie Ren et al.

Gradient-based Meta-RL (GMRL) refers to methods that maintain two-level optimisation procedures wherein the outer-loop meta-learner guides the inner-loop gradient-based reinforcement learner to achieve fast adaptations. In this paper, we develop a unified framework that describes variations of GMRL algorithms and points out that existing stochastic meta-gradient estimators adopted by GMRL are actually \textbf{biased}. Such meta-gradient bias comes from two sources: 1) the compositional bias incurred by the two-level problem structure, which has an upper bound of $\mathcal{O}\big(Kα^{K}\hatσ_{\text{In}}|τ|^{-0.5}\big)$ \emph{w.r.t.} inner-loop update step $K$, learning rate $α$, estimate variance $\hatσ^{2}_{\text{In}}$ and sample size $|τ|$, and 2) the multi-step Hessian estimation bias $\hatΔ_{H}$ due to the use of autodiff, which has a polynomial impact $\mathcal{O}\big((K-1)(\hatΔ_{H})^{K-1}\big)$ on the meta-gradient bias. We study tabular MDPs empirically and offer quantitative evidence that testifies our theoretical findings on existing stochastic meta-gradient estimators. Furthermore, we conduct experiments on Iterated Prisoner's Dilemma and Atari games to show how other methods such as off-policy learning and low-bias estimator can help fix the gradient bias for GMRL algorithms in general.

Linghui Meng, Muning Wen, Yaodong Yang et al.

Offline reinforcement learning leverages previously-collected offline datasets to learn optimal policies with no necessity to access the real environment. Such a paradigm is also desirable for multi-agent reinforcement learning (MARL) tasks, given the increased interactions among agents and with the enviroment. Yet, in MARL, the paradigm of offline pre-training with online fine-tuning has not been studied, nor datasets or benchmarks for offline MARL research are available. In this paper, we facilitate the research by providing large-scale datasets, and use them to examine the usage of the Decision Transformer in the context of MARL. We investigate the generalisation of MARL offline pre-training in the following three aspects: 1) between single agents and multiple agents, 2) from offline pretraining to the online fine-tuning, and 3) to that of multiple downstream tasks with few-shot and zero-shot capabilities. We start by introducing the first offline MARL dataset with diverse quality levels based on the StarCraftII environment, and then propose the novel architecture of multi-agent decision transformer (MADT) for effective offline learning. MADT leverages transformer's modelling ability of sequence modelling and integrates it seamlessly with both offline and online MARL tasks. A crucial benefit of MADT is that it learns generalisable policies that can transfer between different types of agents under different task scenarios. On StarCraft II offline dataset, MADT outperforms the state-of-the-art offline RL baselines. When applied to online tasks, the pre-trained MADT significantly improves sample efficiency, and enjoys strong performance both few-short and zero-shot cases. To our best knowledge, this is the first work that studies and demonstrates the effectiveness of offline pre-trained models in terms of sample efficiency and generalisability enhancements in MARL.

Chenguang Wang, Yaodong Yang, Oliver Slumbers et al.

In this paper, we introduce a two-player zero-sum framework between a trainable \emph{Solver} and a \emph{Data Generator} to improve the generalization ability of deep learning-based solvers for Traveling Salesman Problem (TSP). Grounded in \textsl{Policy Space Response Oracle} (PSRO) methods, our two-player framework outputs a population of best-responding Solvers, over which we can mix and output a combined model that achieves the least exploitability against the Generator, and thereby the most generalizable performance on different TSP tasks. We conduct experiments on a variety of TSP instances with different types and sizes. Results suggest that our Solvers achieve the state-of-the-art performance even on tasks the Solver never meets, whilst the performance of other deep learning-based Solvers drops sharply due to over-fitting. To demonstrate the principle of our framework, we study the learning outcome of the proposed two-player game and demonstrate that the exploitability of the Solver population decreases during training, and it eventually approximates the Nash equilibrium along with the Generator.

Jakub Grudzien Kuba, Muning Wen, Yaodong Yang et al.

Policy gradient (PG) methods are popular reinforcement learning (RL) methods where a baseline is often applied to reduce the variance of gradient estimates. In multi-agent RL (MARL), although the PG theorem can be naturally extended, the effectiveness of multi-agent PG (MAPG) methods degrades as the variance of gradient estimates increases rapidly with the number of agents. In this paper, we offer a rigorous analysis of MAPG methods by, firstly, quantifying the contributions of the number of agents and agents' explorations to the variance of MAPG estimators. Based on this analysis, we derive the optimal baseline (OB) that achieves the minimal variance. In comparison to the OB, we measure the excess variance of existing MARL algorithms such as vanilla MAPG and COMA. Considering using deep neural networks, we also propose a surrogate version of OB, which can be seamlessly plugged into any existing PG methods in MARL. On benchmarks of Multi-Agent MuJoCo and StarCraft challenges, our OB technique effectively stabilises training and improves the performance of multi-agent PPO and COMA algorithms by a significant margin.

Yunfei Liu, Yang Yang, Xianyu Chen et al.

Knowledge tracing (KT) defines the task of predicting whether students can correctly answer questions based on their historical response. Although much research has been devoted to exploiting the question information, plentiful advanced information among questions and skills hasn't been well extracted, making it challenging for previous work to perform adequately. In this paper, we demonstrate that large gains on KT can be realized by pre-training embeddings for each question on abundant side information, followed by training deep KT models on the obtained embeddings. To be specific, the side information includes question difficulty and three kinds of relations contained in a bipartite graph between questions and skills. To pre-train the question embeddings, we propose to use product-based neural networks to recover the side information. As a result, adopting the pre-trained embeddings in existing deep KT models significantly outperforms state-of-the-art baselines on three common KT datasets.

Liheng Chen, Hongyi Guo, Yali Du et al.

In many real-world problems, a team of agents need to collaborate to maximize the common reward. Although existing works formulate this problem into a centralized learning with decentralized execution framework, which avoids the non-stationary problem in training, their decentralized execution paradigm limits the agents' capability to coordinate. Inspired by the concept of correlated equilibrium, we propose to introduce a coordination signal to address this limitation, and theoretically show that following mild conditions, decentralized agents with the coordination signal can coordinate their individual policies as manipulated by a centralized controller. The idea of introducing coordination signal is to encapsulate coordinated strategies into the signals, and use the signals to instruct the collaboration in decentralized execution. To encourage agents to learn to exploit the coordination signal, we propose Signal Instructed Coordination (SIC), a novel coordination module that can be integrated with most existing MARL frameworks. SIC casts a common signal sampled from a pre-defined distribution to all agents, and introduces an information-theoretic regularization to facilitate the consistency between the observed signal and agents' policies. Our experiments show that SIC consistently improves performance over well-recognized MARL models in both matrix games and a predator-prey game with high-dimensional strategy space.

Haifeng Zhang, Zilong Guo, Han Cai et al.

With the rapid growth of the express industry, intelligent warehouses that employ autonomous robots for carrying parcels have been widely used to handle the vast express volume. For such warehouses, the warehouse layout design plays a key role in improving the transportation efficiency. However, this work is still done by human experts, which is expensive and leads to suboptimal results. In this paper, we aim to automate the warehouse layout designing process. We propose a two-layer evolutionary algorithm to efficiently explore the warehouse layout space, where an auxiliary objective fitness approximation model is introduced to predict the outcome of the designed warehouse layout and a two-layer population structure is proposed to incorporate the approximation model into the ordinary evolution framework. Empirical experiments show that our method can efficiently design effective warehouse layouts that outperform both heuristic-designed and vanilla evolution-designed warehouse layouts.

Yi Zhang, Houjun Huang, Haifeng Zhang et al.

In recent years, finger vein recognition has become an important sub-field in biometrics and been applied to real-world applications. The development of finger vein recognition algorithms heavily depends on large-scale real-world data sets. In order to motivate research on finger vein recognition, we released the largest finger vein data set up to now and hold finger vein recognition competitions based on our data set every year. In 2017, International Competition on Finger Vein Recognition(ICFVR) is held jointly with IJCB 2017. 11 teams registered and 10 of them joined the final evaluation. The winner of this year dramatically improved the EER from 2.64% to 0.483% compared to the winner of last year. In this paper, we introduce the process and results of ICFVR 2017 and give insights on development of state-of-art finger vein recognition algorithms.

Haifeng Zhang, Jun Wang, Zhiming Zhou et al.

In typical reinforcement learning (RL), the environment is assumed given and the goal of the learning is to identify an optimal policy for the agent taking actions through its interactions with the environment. In this paper, we extend this setting by considering the environment is not given, but controllable and learnable through its interaction with the agent at the same time. This extension is motivated by environment design scenarios in the real-world, including game design, shopping space design and traffic signal design. Theoretically, we find a dual Markov decision process (MDP) w.r.t. the environment to that w.r.t. the agent, and derive a policy gradient solution to optimizing the parametrized environment. Furthermore, discontinuous environments are addressed by a proposed general generative framework. Our experiments on a Maze game design task show the effectiveness of the proposed algorithms in generating diverse and challenging Mazes against various agent settings.

Haifeng Zhang, Yevgeniy Vorobeychik

Innovation diffusion has been studied extensively in a variety of disciplines, including sociology, economics, marketing, ecology, and computer science. Traditional literature on innovation diffusion has been dominated by models of aggregate behavior and trends. However, the agent-based modeling (ABM) paradigm is gaining popularity as it captures agent heterogeneity and enables fine-grained modeling of interactions mediated by social and geographic networks. While most ABM work on innovation diffusion is theoretical, empirically grounded models are increasingly important, particularly in guiding policy decisions. We present a critical review of empirically grounded agent-based models of innovation diffusion, developing a categorization of this research based on types of agent models as well as applications. By connecting the modeling methodologies in the fields of information and innovation diffusion, we suggest that the maximum likelihood estimation framework widely used in the former is a promising paradigm for calibration of agent-based models for innovation diffusion. Although many advances have been made to standardize ABM methodology, we identify four major issues in model calibration and validation, and suggest potential solutions.