Boyi Liu

Qianyi Zhang, Shijian Ma, Boyi Liu et al.

Robust and flexible leader-following is a critical capability for robots to integrate into human society. While existing methods struggle to generalize to leaders of arbitrary form and often fail when the leader temporarily leaves the robot's field of view, this work introduces a unified framework addressing both challenges. First, traditional detection models are replaced with a segmentation model, allowing the leader to be anything. To enhance recognition robustness, a distance frame buffer is implemented that stores leader embeddings at multiple distances, accounting for the unique characteristics of leader-following tasks. Second, a goal-aware adaptation mechanism is designed to govern robot planning states based on the leader's visibility and motion, complemented by a graph-based planner that generates candidate trajectories for each state, ensuring efficient following with obstacle avoidance. Simulations and real-world experiments with a legged robot follower and various leaders (human, ground robot, UAV, legged robot, stop sign) in both indoor and outdoor environments show competitive improvements in follow success rate, reduced visual loss duration, lower collision rate, and decreased leader-follower distance.

ByteDance Seed, Jiaze Chen, Tiantian Fan et al. · bytedance

We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.

Bytedance-Seed-Foundation-Code-Team, Yao Cheng, Jianfeng Chen et al. · bytedance

As the capabilities of code large language models (LLMs) continue to expand, their applications across diverse code intelligence domains are rapidly increasing. However, most existing datasets only evaluate limited application domains. To address this gap, we have developed a comprehensive code evaluation dataset FullStack Bench focusing on full-stack programming, which encompasses a wide range of application domains (e.g., basic programming, data analysis, software engineering, mathematics, and machine learning). Besides, to assess multilingual programming capabilities, in FullStack Bench, we design real-world instructions and corresponding unit test cases from 16 widely-used programming languages to reflect real-world usage scenarios rather than simple translations. Moreover, we also release an effective code sandbox execution tool (i.e., SandboxFusion) supporting various programming languages and packages to evaluate the performance of our FullStack Bench efficiently. Comprehensive experimental results on our FullStack Bench demonstrate the necessity and effectiveness of our FullStack Bench and SandboxFusion.

Chau Pham, Boyi Liu, Yingxiang Yang et al.

Discussion and debate among Large Language Models (LLMs) have gained considerable attention due to their potential to enhance the reasoning ability of LLMs. Although natural language is an obvious choice for communication due to LLM's language understanding capability, the token sampling step needed when generating natural language poses a potential risk of information loss, as it uses only one token to represent the model's belief across the entire vocabulary. In this paper, we introduce a communication regime named CIPHER (Communicative Inter-Model Protocol Through Embedding Representation) to address this issue. Specifically, we remove the token sampling step from LLMs and let them communicate their beliefs across the vocabulary through the expectation of the raw transformer output embeddings. Remarkably, by deviating from natural language, CIPHER offers an advantage of encoding a broader spectrum of information without any modification to the model weights, outperforming the state-of-the-art LLM debate methods using natural language by 0.5-5.0% across five reasoning tasks and multiple open-source LLMs of varying sizes. This showcases the superiority and robustness of embeddings as an alternative "language" for communication among LLMs. We anticipate that CIPHER will inspire further exploration for the design of interactions within LLM agent systems, offering a new direction that could significantly influence future developments in the field.

Ruixing Zhang, Liangzhe Han, Boyi Liu et al.

Recent years have witnessed a rapid growth of applying deep spatiotemporal methods in traffic forecasting. However, the prediction of origin-destination (OD) demands is still a challenging problem since the number of OD pairs is usually quadratic to the number of stations. In this case, most of the existing spatiotemporal methods fail to handle spatial relations on such a large scale. To address this problem, this paper provides a dynamic graph representation learning framework for OD demands prediction. In particular, a hierarchical memory updater is first proposed to maintain a time-aware representation for each node, and the representations are updated according to the most recently observed OD trips in continuous-time and multiple discrete-time ways. Second, a spatiotemporal propagation mechanism is provided to aggregate representations of neighbor nodes along a random spatiotemporal route which treats origin and destination as two different semantic entities. Last, an objective function is designed to derive the future OD demands according to the most recent node representations, and also to tackle the data sparsity problem in OD prediction. Extensive experiments have been conducted on two real-world datasets, and the experimental results demonstrate the superiority of the proposed method. The code and data are available at https://github.com/Rising0321/HMOD.

Zhihan Liu, Hao Hu, Shenao Zhang et al.

Large language models (LLMs) demonstrate impressive reasoning abilities, but translating reasoning into actions in the real world remains challenging. In particular, it remains unclear how to complete a given task provably within a minimum number of interactions with the external environment, e.g., through an internal mechanism of reasoning. To this end, we propose a principled framework with provable regret guarantees to orchestrate reasoning and acting, which we call "reason for future, act for now" (\texttt{RAFA}). Specifically, we design a prompt template for reasoning that learns from the memory buffer and plans a future trajectory over a long horizon ("reason for future"). At each step, the LLM agent takes the initial action of the planned trajectory ("act for now"), stores the collected feedback in the memory buffer, and reinvokes the reasoning routine to replan the future trajectory from the new state. The key idea is to cast reasoning in LLMs as learning and planning in Bayesian adaptive Markov decision processes (MDPs). Correspondingly, we prompt LLMs to form an updated posterior of the unknown environment from the memory buffer (learning) and generate an optimal trajectory for multiple future steps that maximizes a value function (planning). The learning and planning subroutines are performed in an "in-context" manner to emulate the actor-critic update for MDPs. Our theoretical analysis proves that the novel combination of long-term reasoning and short-term acting achieves a $\sqrt{T}$ regret. Here, $T$ denotes the number of online interactions. In particular, the regret bound highlights an intriguing interplay between the prior knowledge obtained through pretraining and the uncertainty reduction achieved by reasoning and acting. Our empirical validation shows that it outperforms various existing frameworks and achieves nearly perfect scores on a few benchmarks.

Zhaoyang Wang, Canwen Xu, Boyi Liu et al.

Recent advances in large language model (LLM) have empowered autonomous agents to perform complex tasks that require multi-turn interactions with tools and environments. However, scaling such agent training is limited by the lack of diverse and reliable environments. In this paper, we propose Agent World Model (AWM), a fully synthetic environment generation pipeline. Using this pipeline, we scale to 1,000 environments covering everyday scenarios, in which agents can interact with rich toolsets (35 tools per environment on average) and obtain high-quality observations. Notably, these environments are code-driven and backed by databases, providing more reliable and consistent state transitions than environments simulated by LLMs. Moreover, they enable more efficient agent interaction compared with collecting trajectories from realistic environments. To demonstrate the effectiveness of this resource, we perform large-scale reinforcement learning for multi-turn tool-use agents. Thanks to the fully executable environments and accessible database states, we can also design reliable reward functions. Experiments on three benchmarks show that training exclusively in synthetic environments, rather than benchmark-specific ones, yields strong out-of-distribution generalization. The code is available at https://github.com/Snowflake-Labs/agent-world-model.

Shenao Zhang, Boyi Liu, Zhaoran Wang et al.

ReParameterization (RP) Policy Gradient Methods (PGMs) have been widely adopted for continuous control tasks in robotics and computer graphics. However, recent studies have revealed that, when applied to long-term reinforcement learning problems, model-based RP PGMs may experience chaotic and non-smooth optimization landscapes with exploding gradient variance, which leads to slow convergence. This is in contrast to the conventional belief that reparameterization methods have low gradient estimation variance in problems such as training deep generative models. To comprehend this phenomenon, we conduct a theoretical examination of model-based RP PGMs and search for solutions to the optimization difficulties. Specifically, we analyze the convergence of the model-based RP PGMs and pinpoint the smoothness of function approximators as a major factor that affects the quality of gradient estimation. Based on our analysis, we propose a spectral normalization method to mitigate the exploding variance issue caused by long model unrolls. Our experimental results demonstrate that proper normalization significantly reduces the gradient variance of model-based RP PGMs. As a result, the performance of the proposed method is comparable or superior to other gradient estimators, such as the Likelihood Ratio (LR) gradient estimator. Our code is available at https://github.com/agentification/RP_PGM.

Jiayang Li, Jing Yu, Qianni Wang et al.

In a Stackelberg congestion game (SCG), a leader aims to maximize their own gain by anticipating and manipulating the equilibrium state at which the followers settle by playing a congestion game. Often formulated as bilevel programs, large-scale SCGs are well known for their intractability and complexity. Here, we attempt to tackle this computational challenge by marrying traditional methodologies with the latest differentiable programming techniques in machine learning. The core idea centers on replacing the lower-level equilibrium problem with a smooth evolution trajectory defined by the imitative logit dynamic (ILD), which we prove converges to the equilibrium of the congestion game under mild conditions. Building upon this theoretical foundation, we propose two new local search algorithms for SCGs. The first is a gradient descent algorithm that obtains the derivatives by unrolling ILD via differentiable programming. Thanks to the smoothness of ILD, the algorithm promises both efficiency and scalability. The second algorithm adds a heuristic twist by cutting short the followers' evolution trajectory. Behaviorally, this means that, instead of anticipating the followers' best response at equilibrium, the leader seeks to approximate that response by only looking ahead a limited number of steps. Our numerical experiments are carried out over various instances of classic SCG applications, ranging from toy benchmarks to large-scale real-world examples. The results show the proposed algorithms are reliable and scalable local solvers that deliver high-quality solutions with greater regularity and significantly less computational effort compared to the many incumbents included in our study.

Fengzhuo Zhang, Boyi Liu, Kaixin Wang et al.

The cooperative Multi-A gent R einforcement Learning (MARL) with permutation invariant agents framework has achieved tremendous empirical successes in real-world applications. Unfortunately, the theoretical understanding of this MARL problem is lacking due to the curse of many agents and the limited exploration of the relational reasoning in existing works. In this paper, we verify that the transformer implements complex relational reasoning, and we propose and analyze model-free and model-based offline MARL algorithms with the transformer approximators. We prove that the suboptimality gaps of the model-free and model-based algorithms are independent of and logarithmic in the number of agents respectively, which mitigates the curse of many agents. These results are consequences of a novel generalization error bound of the transformer and a novel analysis of the Maximum Likelihood Estimate (MLE) of the system dynamics with the transformer. Our model-based algorithm is the first provably efficient MARL algorithm that explicitly exploits the permutation invariance of the agents. Our improved generalization bound may be of independent interest and is applicable to other regression problems related to the transformer beyond MARL.

Yufeng Zhang, Boyi Liu, Qi Cai et al.

With the attention mechanism, transformers achieve significant empirical successes. Despite the intuitive understanding that transformers perform relational inference over long sequences to produce desirable representations, we lack a rigorous theory on how the attention mechanism achieves it. In particular, several intriguing questions remain open: (a) What makes a desirable representation? (b) How does the attention mechanism infer the desirable representation within the forward pass? (c) How does a pretraining procedure learn to infer the desirable representation through the backward pass? We observe that, as is the case in BERT and ViT, input tokens are often exchangeable since they already include positional encodings. The notion of exchangeability induces a latent variable model that is invariant to input sizes, which enables our theoretical analysis. - To answer (a) on representation, we establish the existence of a sufficient and minimal representation of input tokens. In particular, such a representation instantiates the posterior distribution of the latent variable given input tokens, which plays a central role in predicting output labels and solving downstream tasks. - To answer (b) on inference, we prove that attention with the desired parameter infers the latent posterior up to an approximation error, which is decreasing in input sizes. In detail, we quantify how attention approximates the conditional mean of the value given the key, which characterizes how it performs relational inference over long sequences. - To answer (c) on learning, we prove that both supervised and self-supervised objectives allow empirical risk minimization to learn the desired parameter up to a generalization error, which is independent of input sizes. Particularly, in the self-supervised setting, we identify a condition number that is pivotal to solving downstream tasks.

Jing Wang, Meichen Song, Feng Gao et al.

We initiate the study of how to perturb the reward in a zero-sum Markov game with two players to induce a desirable Nash equilibrium, namely arbitrating. Such a problem admits a bi-level optimization formulation. The lower level requires solving the Nash equilibrium under a given reward function, which makes the overall problem challenging to optimize in an end-to-end way. We propose a backpropagation scheme that differentiates through the Nash equilibrium, which provides the gradient feedback for the upper level. In particular, our method only requires a black-box solver for the (regularized) Nash equilibrium (NE). We develop the convergence analysis for the proposed framework with proper black-box NE solvers and demonstrate the empirical successes in two multi-agent reinforcement learning (MARL) environments.

Xiaoyin Chen, Canwen Xu, Yite Wang et al.

We introduce Learning to Self-Evolve (LSE), a reinforcement learning framework that trains large language models (LLMs) to improve their own contexts at test time. We situate LSE in the setting of test-time self-evolution, where a model iteratively refines its context from feedback on seen problems to perform better on new ones. Existing approaches rely entirely on the inherent reasoning ability of the model and never explicitly train it for this task. LSE reduces the multi-step evolution problem to a single-step RL objective, where each context edit is rewarded by the improvement in downstream performance. We pair this objective with a tree-guided evolution loop. On Text-to-SQL generation (BIRD) and general question answering (MMLU-Redux), a 4B-parameter model trained with LSE outperforms self-evolving policies powered by GPT-5 and Claude Sonnet 4.5, as well as prompt optimization methods including GEPA and TextGrad, and transfers to guide other models without additional training. Our results highlight the effectiveness of treating self-evolution as a learnable skill.

Mingkai Tang, Boyi Liu, Yuanhang Li et al.

Multi-agent path finding (MAPF) is the problem of moving agents to the goal vertex without collision. In the online MAPF problem, new agents may be added to the environment at any time, and the current agents have no information about future agents. The inability of existing online methods to reuse previous planning contexts results in redundant computation and reduces algorithm efficiency. Hence, we propose a three-level approach to solve online MAPF utilizing sustainable information, which can decrease its redundant calculations. The high-level solver, the Sustainable Replan algorithm (SR), manages the planning context and simulates the environment. The middle-level solver, the Sustainable Conflict-Based Search algorithm (SCBS), builds a conflict tree and maintains the planning context. The low-level solver, the Sustainable Reverse Safe Interval Path Planning algorithm (SRSIPP), is an efficient single-agent solver that uses previous planning context to reduce duplicate calculations. Experiments show that our proposed method has significant improvement in terms of computational efficiency. In one of the test scenarios, our algorithm can be 1.48 times faster than SOTA on average under different agent number settings.

Yi Zhan, Yang Sun, Han Weng et al.

In this paper, we propose a novel graph-based methodology to evaluate the functional correctness of SQL generation. Conventional metrics for assessing SQL code generation, such as matching-based and execution-based methods (e.g., exact set match and execution accuracy), are subject to two primary limitations. Firstly, the former fails to effectively assess functional correctness, as different SQL queries may possess identical functionalities. Secondly, the latter is susceptible to producing false positive samples in evaluations. Our proposed evaluation method, \texttt{FuncEvalGMN}, does not depend on the sufficient preparation of the test data, and it enables precise testing of the functional correctness of the code. Firstly, we parse SQL using a relational operator tree (ROT) called \textit{Relnode}, which contains rich semantic information from the perspective of logical execution.Then, we introduce a GNN-based approach for predicting the functional correctness of generated SQL. This approach incorporates global positional embeddings to address the limitations with the loss of topological information in conventional graph matching frameworks. As an auxiliary contribution, we propose a rule-based matching algorithm, Relnode Partial Matching (\texttt{RelPM}) as a baseline. Finally, we contribute a dataset, \texttt{Pair-Aug-Spider} with a training set and two testing sets, each comprising pairs of SQL codes to simulate various SQL code evaluation scenarios. The training set and one testing dataset focus on code generation using large language models (LLMs), while the other emphasizes SQL equivalence rewriting.

Boyi Liu, Zimu Zhou, Yongxin Tong

Personalized Federated Learning (PFL) enables collaboratively model training on decentralized, heterogeneous data while tailoring them to each client's unique distribution. However, existing PFL methods produce static models with a fixed tradeoff between accuracy and efficiency, limiting their applicability in environments where inference requirements vary with contexts and resource availability. Early-exit networks (EENs) offer adaptive inference by attaching intermediate classifiers. Yet integrating them into PFL is challenging due to client-wise heterogeneity and depth-wise interference arising from conflicting exit objectives. Prior studies fail to resolve both conflicts simultaneously, leading to suboptimal performance. In this paper, we propose CAFEDistill, a Conflict-Aware Federated Exit Distillation framework that jointly addresses these conflicts and extends PFL to early-exit networks. Through a progressive, depth-prioritized student coordination mechanism, CAFEDistill mitigates interference among shallow and deep exits while allowing effective personalized knowledge transfer across clients. Furthermore, it reduces communication overhead via a client-decoupled formulation. Extensive evaluations show that CAFEDistill outperforms the state-of-the-arts, achieving higher accuracy and reducing inference costs by 30.79%-46.86%.

Boyi Liu, Bingjie Yan, Yize Zhou et al.

Federated Learning (FL) creates an ecosystem for multiple agents to collaborate on building models with data privacy consideration. The method for contribution measurement of each agent in the FL system is critical for fair credits allocation but few are proposed. In this paper, we develop a real-time contribution measurement method FedCM that is simple but powerful. The method defines the impact of each agent, comprehensively considers the current round and the previous round to obtain the contribution rate of each agent with attention aggregation. Moreover, FedCM updates contribution every round, which enable it to perform in real-time. Real-time is not considered by the existing approaches, but it is critical for FL systems to allocate computing power, communication resources, etc. Compared to the state-of-the-art method, the experimental results show that FedCM is more sensitive to data quantity and data quality under the premise of real-time. Furthermore, we developed federated learning open-source software based on FedCM. The software has been applied to identify COVID-19 based on medical images.

Zhilin Liang, Yuxiang Wang, Zimu Zhou et al.

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by grounding generation in external knowledge to improve factuality and reduce hallucinations. Yet most deployments assume a centralized corpus, which is infeasible in privacy aware domains where knowledge remains siloed. This motivates federated RAG (FedRAG), where a central LLM server collaborates with distributed silos without sharing raw documents. In context RAG violates this requirement by transmitting verbatim documents, whereas parametric RAG encodes documents into lightweight adapters that merge with a frozen LLM at inference, avoiding raw-text exchange. We adopt the parametric approach but face two unique challenges induced by FedRAG: high storage and communication from per-document adapters, and destructive aggregation caused by indiscriminately merging multiple adapters. We present FedMosaic, the first federated RAG framework built on parametric adapters. FedMosaic clusters semantically related documents into multi-document adapters with document-specific masks to reduce overhead while preserving specificity, and performs selective adapter aggregation to combine only relevance-aligned, nonconflicting adapters. Experiments show that FedMosaic achieves an average 10.9% higher accuracy than state-of-the-art methods in four categories, while lowering storage costs by 78.8% to 86.3% and communication costs by 91.4%, and never sharing raw documents.

Han Zhong, Yutong Yin, Shenao Zhang et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, yet generating reliable reasoning processes remains a significant challenge. We present a unified probabilistic framework that formalizes LLM reasoning through a novel graphical model incorporating latent thinking processes and evaluation signals. Within this framework, we introduce the Bootstrapping Reinforced Thinking Process (BRiTE) algorithm, which works in two steps. First, it generates high-quality rationales by approximating the optimal thinking process through reinforcement learning, using a novel reward shaping mechanism. Second, it enhances the base LLM by maximizing the joint probability of rationale generation with respect to the model's parameters. Theoretically, we demonstrate BRiTE's convergence at a rate of $1/T$ with $T$ representing the number of iterations. Empirical evaluations on math and coding benchmarks demonstrate that our approach consistently improves performance across different base models without requiring human-annotated thinking processes. In addition, BRiTE demonstrates superior performance compared to existing algorithms that bootstrap thinking processes use alternative methods such as rejection sampling, and can even match or exceed the results achieved through supervised fine-tuning with human-annotated data.

Zhihan Liu, Shenao Zhang, Yongfei Liu et al.

Direct preference learning offers a promising and computation-efficient beyond supervised fine-tuning (SFT) for improving code generation in coding large language models (LMs). However, the scarcity of reliable preference data is a bottleneck for the performance of direct preference learning to improve the coding accuracy of code LMs. In this paper, we introduce \underline{\textbf{D}}irect Preference Learning with Only \underline{\textbf{S}}elf-Generated \underline{\textbf{T}}ests and \underline{\textbf{C}}ode (DSTC), a framework that leverages only self-generated code snippets and tests to construct reliable preference pairs such that direct preference learning can improve LM coding accuracy without external annotations. DSTC combines a minimax selection process and test-code concatenation to improve preference pair quality, reducing the influence of incorrect self-generated tests and enhancing model performance without the need for costly reward models. When applied with direct preference learning methods such as Direct Preference Optimization (DPO) and Kahneman-Tversky Optimization (KTO), DSTC yields stable improvements in coding accuracy (pass@1 score) across diverse coding benchmarks, including HumanEval, MBPP, and BigCodeBench, demonstrating both its effectiveness and scalability for models of various sizes. This approach autonomously enhances code generation accuracy across LLMs of varying sizes, reducing reliance on expensive annotated coding datasets.

Yexiao He, Ang Li, Boyi Liu et al.

Healthcare decision-making represents one of the most challenging domains for Artificial Intelligence (AI), requiring the integration of diverse knowledge sources, complex reasoning, and various external analytical tools. Current AI systems often rely on either task-specific models, which offer limited adaptability, or general language models without grounding with specialized external knowledge and tools. We introduce MedOrch, a novel framework that orchestrates multiple specialized tools and reasoning agents to provide comprehensive medical decision support. MedOrch employs a modular, agent-based architecture that facilitates the flexible integration of domain-specific tools without altering the core system. Furthermore, it ensures transparent and traceable reasoning processes, enabling clinicians to meticulously verify each intermediate step underlying the system's recommendations. We evaluate MedOrch across three distinct medical applications: Alzheimer's disease diagnosis, chest X-ray interpretation, and medical visual question answering, using authentic clinical datasets. The results demonstrate MedOrch's competitive performance across these diverse medical tasks. Notably, in Alzheimer's disease diagnosis, MedOrch achieves an accuracy of 93.26%, surpassing the state-of-the-art baseline by over four percentage points. For predicting Alzheimer's disease progression, it attains a 50.35% accuracy, marking a significant improvement. In chest X-ray analysis, MedOrch exhibits superior performance with a Macro AUC of 61.2% and a Macro F1-score of 25.5%. Moreover, in complex multimodal visual question answering (Image+Table), MedOrch achieves an accuracy of 54.47%. These findings underscore MedOrch's potential to advance healthcare AI by enabling reasoning-driven tool utilization for multimodal medical data processing and supporting intricate cognitive tasks in clinical decision-making.

Hao Wang, Boyi Liu, Yufeng Zhang et al.

Competition-level code generation tasks pose significant challenges for current state-of-the-art large language models (LLMs). For example, on the LiveCodeBench-Hard dataset, models such as O1-Mini and O1-Preview achieve pass@1 rates of only 0.366 and 0.143, respectively. While tree search techniques have proven effective in domains like mathematics and general coding, their potential in competition-level code generation remains under-explored. In this work, we propose a novel token-level tree search method specifically designed for code generation. Leveraging Qwen2.5-Coder-32B-Instruct, our approach achieves a pass rate of 0.305 on LiveCodeBench-Hard, surpassing the pass@100 performance of GPT4o-0513 (0.245). Furthermore, by integrating Chain-of-Thought (CoT) prompting, we improve our method's performance to 0.351, approaching O1-Mini's pass@1 rate. To ensure reproducibility, we report the average number of generations required per problem by our tree search method on the test set. Our findings underscore the potential of tree search to significantly enhance performance on competition-level code generation tasks. This opens up new possibilities for large-scale synthesis of challenging code problems supervised fine-tuning (SFT) data, advancing competition-level code generation tasks.

Zihao Li, Boyi Liu, Zhuoran Yang et al.

We study the Constrained Convex Markov Decision Process (MDP), where the goal is to minimize a convex functional of the visitation measure, subject to a convex constraint. Designing algorithms for a constrained convex MDP faces several challenges, including (1) handling the large state space, (2) managing the exploration/exploitation tradeoff, and (3) solving the constrained optimization where the objective and the constraint are both nonlinear functions of the visitation measure. In this work, we present a model-based algorithm, Variational Primal-Dual Policy Optimization (VPDPO), in which Lagrangian and Fenchel duality are implemented to reformulate the original constrained problem into an unconstrained primal-dual optimization. Moreover, the primal variables are updated by model-based value iteration following the principle of Optimism in the Face of Uncertainty (OFU), while the dual variables are updated by gradient ascent. Moreover, by embedding the visitation measure into a finite-dimensional space, we can handle large state spaces by incorporating function approximation. Two notable examples are (1) Kernelized Nonlinear Regulators and (2) Low-rank MDPs. We prove that with an optimistic planning oracle, our algorithm achieves sublinear regret and constraint violation in both cases and can attain the globally optimal policy of the original constrained problem.

Han Weng, Puzhen Wu, Longjie Cui et al.

Reinforcement learning (RL) has been widely adopted to enhance the performance of large language models (LLMs) on Text-to-SQL tasks. However, existing methods often rely on execution-based or LLM-based Bradley-Terry reward models. The former suffers from high execution latency caused by repeated database calls, whereas the latter imposes substantial GPU memory overhead, both of which significantly hinder the efficiency and scalability of RL pipelines. To this end, we propose a novel reward model framework for RL-based Text-to-SQL named Graph-Reward-SQL, which employs the GMNScore outcome reward model. We leverage SQL graph representations to provide accurate reward signals while significantly reducing time cost and GPU memory usage. Building on this foundation, we further introduce StepRTM, a stepwise reward model that provides intermediate supervision over Common Table Expression (CTE) subqueries. This encourages both functional correctness and readability of SQL. Extensive comparative and ablation experiments on standard benchmarks, including Spider and BIRD, demonstrate that our method consistently outperforms existing reward models.

Yufeng Zhang, Liyu Chen, Boyi Liu et al.

Recent advances in reinforcement learning (RL) algorithms aim to enhance the performance of language models at scale. Yet, there is a noticeable absence of a cost-effective and standardized testbed tailored to evaluating and comparing these algorithms. To bridge this gap, we present a generalized version of the 24-Puzzle: the $(N,K)$-Puzzle, which challenges language models to reach a target value $K$ with $N$ integers. We evaluate the effectiveness of established RL algorithms such as Proximal Policy Optimization (PPO), alongside novel approaches like Identity Policy Optimization (IPO) and Direct Policy Optimization (DPO).

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Boyi Liu, Jiayang Li, Zhuoran Yang et al.

To regulate a social system comprised of self-interested agents, economic incentives are often required to induce a desirable outcome. This incentive design problem naturally possesses a bilevel structure, in which a designer modifies the rewards of the agents with incentives while anticipating the response of the agents, who play a non-cooperative game that converges to an equilibrium. The existing bilevel optimization algorithms raise a dilemma when applied to this problem: anticipating how incentives affect the agents at equilibrium requires solving the equilibrium problem repeatedly, which is computationally inefficient; bypassing the time-consuming step of equilibrium-finding can reduce the computational cost, but may lead the designer to a sub-optimal solution. To address such a dilemma, we propose a method that tackles the designer's and agents' problems simultaneously in a single loop. Specifically, at each iteration, both the designer and the agents only move one step. Nevertheless, we allow the designer to gradually learn the overall influence of the incentives on the agents, which guarantees optimality after convergence. The convergence rate of the proposed scheme is also established for a broad class of games.

Jieren Cheng, Le Liu, Xiangyan Tang et al.

Unsupervised domain adaptation aims to train a classification model from the labeled source domain for the unlabeled target domain. Since the data distributions of the two domains are different, the model often performs poorly on the target domain. Existing methods align the feature distributions of the source and target domains and learn domain-invariant features to improve the performance of the model. However, the features are usually aligned as a whole, and the domain adaptation task fails to serve the classification, which will ignore the class information and lead to misalignment.In this paper, we investigate those features that should be used for domain alignment, introduce prior knowledge to extract foreground features to guide the domain adaptation task for classification tasks, and perform alignment in the local structure of objects. We propose a method called Foreground Object Structure Transfer(FOST). The key to FOST is the new clustering based condition, which combines the relative position relationship of foreground objects. Based on this conditions, FOST makes the data distribution of the same class more compact in geometry. In practice, since the label of the target domain is not available, we use the clustering information of the source domain to assign pseudo labels to the target domain samples, and then according to the source domain data prior knowledge guides those positive features to maximum the inter-class distance between different classes and mimimum the intra-class distance. Extensive experimental results on various benchmarks ($i.e.$ ImageCLEF-DA, Office-31, Office-Home, Visda-2017) under different domain adaptation settings prove that our FOST compares favorably against the existing state-of-the-art domain adaptation methods.

Zhaohua Zheng, Yize Zhou, Yilong Sun et al.

Federated learning plays an important role in the process of smart cities. With the development of big data and artificial intelligence, there is a problem of data privacy protection in this process. Federated learning is capable of solving this problem. This paper starts with the current developments of federated learning and its applications in various fields. We conduct a comprehensive investigation. This paper summarize the latest research on the application of federated learning in various fields of smart cities. In-depth understanding of the current development of federated learning from the Internet of Things, transportation, communications, finance, medical and other fields. Before that, we introduce the background, definition and key technologies of federated learning. Further more, we review the key technologies and the latest results. Finally, we discuss the future applications and research directions of federated learning in smart cities.

Boyi Liu, Lujia Wang, Xinquan Chen et al.

A technological revolution is occurring in the field of robotics with the data-driven deep learning technology. However, building datasets for each local robot is laborious. Meanwhile, data islands between local robots make data unable to be utilized collaboratively. To address this issue, the work presents Peer-Assisted Robotic Learning (PARL) in robotics, which is inspired by the peer-assisted learning in cognitive psychology and pedagogy. PARL implements data collaboration with the framework of cloud robotic systems. Both data and models are shared by robots to the cloud after semantic computing and training locally. The cloud converges the data and performs augmentation, integration, and transferring. Finally, fine tune this larger shared dataset in the cloud to local robots. Furthermore, we propose the DAT Network (Data Augmentation and Transferring Network) to implement the data processing in PARL. DAT Network can realize the augmentation of data from multi-local robots. We conduct experiments on a simplified self-driving task for robots (cars). DAT Network has a significant improvement in the augmentation in self-driving scenarios. Along with this, the self-driving experimental results also demonstrate that PARL is capable of improving learning effects with data collaboration of local robots.

Boyi Liu, Bingjie Yan, Yize Zhou et al.

AI plays an important role in COVID-19 identification. Computer vision and deep learning techniques can assist in determining COVID-19 infection with Chest X-ray Images. However, for the protection and respect of the privacy of patients, the hospital's specific medical-related data did not allow leakage and sharing without permission. Collecting such training data was a major challenge. To a certain extent, this has caused a lack of sufficient data samples when performing deep learning approaches to detect COVID-19. Federated Learning is an available way to address this issue. It can effectively address the issue of data silos and get a shared model without obtaining local data. In the work, we propose the use of federated learning for COVID-19 data training and deploy experiments to verify the effectiveness. And we also compare performances of four popular models (MobileNet, ResNet18, MoblieNet, and COVID-Net) with the federated learning framework and without the framework. This work aims to inspire more researches on federated learning about COVID-19.

Yixian Zhang, Jieren Chen, Boyi Liu et al.

With the spread and development of new epidemics, it is of great reference value to identify the changing trends of epidemics in public emotions. We designed and implemented the COVID-19 public opinion monitoring system based on time series thermal new word mining. A new word structure discovery scheme based on the timing explosion of network topics and a Chinese sentiment analysis method for the COVID-19 public opinion environment is proposed. Establish a "Scrapy-Redis-Bloomfilter" distributed crawler framework to collect data. The system can judge the positive and negative emotions of the reviewer based on the comments, and can also reflect the depth of the seven emotions such as Hopeful, Happy, and Depressed. Finally, we improved the sentiment discriminant model of this system and compared the sentiment discriminant error of COVID-19 related comments with the Jiagu deep learning model. The results show that our model has better generalization ability and smaller discriminant error. We designed a large data visualization screen, which can clearly show the trend of public emotions, the proportion of various emotion categories, keywords, hot topics, etc., and fully and intuitively reflect the development of public opinion.

Minghan Chen, Yilong Sun, Xueqing Cai et al.

The agricultural irrigation system is closely related to agricultural production. There are some problems in nowadays agricultural irrigation system, such as poor mobility, imprecision and high price. To address these issues, an intelligent irrigation robot is designed and implemented in this work. The robot achieves precise irrigation by the irrigation path planning algorithm which is improved by Bayesian theory. In the proposed algorithm, we utilize as much information as possible to achieve full coverage irrigation in the complex agricultural environment. Besides, we propose the maximum risk to avoid the problem of lack of inspection in certain areas. Finally, We carried out simulation experiments and field experiments to verify the robot and the algorithm. The experimental results indicate that the robot is capable of fulfilling the requirements of various agricultural irrigation tasks.

Boyi Liu, Lujia Wang, Ming Liu et al.

Humans are capable of learning a new behavior by observing others to perform the skill. Similarly, robots can also implement this by imitation learning. Furthermore, if with external guidance, humans can master the new behavior more efficiently. So, how can robots achieve this? To address the issue, we present a novel framework named FIL. It provides a heterogeneous knowledge fusion mechanism for cloud robotic systems. Then, a knowledge fusion algorithm in FIL is proposed. It enables the cloud to fuse heterogeneous knowledge from local robots and generate guide models for robots with service requests. After that, we introduce a knowledge transfer scheme to facilitate local robots acquiring knowledge from the cloud. With FIL, a robot is capable of utilizing knowledge from other robots to increase its imitation learning in accuracy and efficiency. Compared with transfer learning and meta-learning, FIL is more suitable to be deployed in cloud robotic systems. Finally, we conduct experiments of a self-driving task for robots (cars). The experimental results demonstrate that the shared model generated by FIL increases imitation learning efficiency of local robots in cloud robotic systems.

Boyi Liu, Lujia Wang, Ming Liu et al.

Humans are capable of learning a new behavior by observing others perform the skill. Robots can also implement this by imitation learning. Furthermore, if with external guidance, humans will master the new behavior more efficiently. So how can robots implement this? To address the issue, we present Federated Imitation Learning (FIL) in the paper. Firstly, a knowledge fusion algorithm deployed on the cloud for fusing knowledge from local robots is presented. Then, effective transfer learning methods in FIL are introduced. With FIL, a robot is capable of utilizing knowledge from other robots to increase its imitation learning. FIL considers information privacy and data heterogeneity when robots share knowledge. It is suitable to be deployed in cloud robotic systems. Finally, we conduct experiments of a simplified self-driving task for robots (cars). The experimental results demonstrate that FIL is capable of increasing imitation learning of local robots in cloud robotic systems.

Boyi Liu, Xiangyan Tang, Jieren Cheng et al.

Traffic flow forecasting is hot spot research of intelligent traffic system construction. The existing traffic flow prediction methods have problems such as poor stability, high data requirements, or poor adaptability. In this paper, we define the traffic data time singularity ratio in the dropout module and propose a combination prediction method based on the improved long short-term memory neural network and time series autoregressive integrated moving average model (SDLSTM-ARIMA), which is derived from the Recurrent Neural Networks (RNN) model. It compares the traffic data time singularity with the probability value in the dropout module and combines them at unequal time intervals to achieve an accurate prediction of traffic flow data. Then, we design an adaptive traffic flow embedded system that can adapt to Java, Python and other languages and other interfaces. The experimental results demonstrate that the method based on the SDLSTM - ARIMA model has higher accuracy than the similar method using only autoregressive integrated moving average or autoregressive. Our embedded traffic prediction system integrating computer vision, machine learning and cloud has the advantages such as high accuracy, high reliability and low cost. Therefore, it has a wide application prospect.

Boyi Liu, Qi Cai, Zhuoran Yang et al.

Proximal policy optimization and trust region policy optimization (PPO and TRPO) with actor and critic parametrized by neural networks achieve significant empirical success in deep reinforcement learning. However, due to nonconvexity, the global convergence of PPO and TRPO remains less understood, which separates theory from practice. In this paper, we prove that a variant of PPO and TRPO equipped with overparametrized neural networks converges to the globally optimal policy at a sublinear rate. The key to our analysis is the global convergence of infinite-dimensional mirror descent under a notion of one-point monotonicity, where the gradient and iterate are instantiated by neural networks. In particular, the desirable representation power and optimization geometry induced by the overparametrization of such neural networks allow them to accurately approximate the infinite-dimensional gradient and iterate.

Boyi Liu, Zhuhua Hu, Yaochi Zhao et al.

The Pyralidae pests, such as corn borer and rice leaf roller, are main pests in economic crops. The timely detection and identification of Pyralidae pests is a critical task for agriculturists and farmers. However, the traditional identification of pests by humans is labor intensive and inefficient. To tackle the challenges, a pest monitoring autonomous robot vehicle and a method to recognize Pyralidae pests are presented in this paper. Firstly, the robot on autonomous vehicle collects images by performing camera sensing in natural farm scene. Secondly, the total probability image can be obtained by using inverse histogram mapping, and then the object contour of Pyralidae pests can be extracted quickly and accurately with the constrained Otsu method. Finally, by employing Hu moment and the perimeter and area characteristics, the correct contours of objects can be drawn, and the recognition results can be obtained by comparing them with the reference templates of Pyralidae pests. Additionally, the moving speed of the mechanical arms on the vehicle can be adjusted adaptively by interacting with the recognition algorithm. The experimental results demonstrate that the robot vehicle can automatically capture pest images, and can achieve 94.3$\%$ recognition accuracy in natural farm planting scene.

Boyi Liu, Lujia Wang, Ming Liu

This paper was motivated by the problem of how to make robots fuse and transfer their experience so that they can effectively use prior knowledge and quickly adapt to new environments. To address the problem, we present a learning architecture for navigation in cloud robotic systems: Lifelong Federated Reinforcement Learning (LFRL). In the work, We propose a knowledge fusion algorithm for upgrading a shared model deployed on the cloud. Then, effective transfer learning methods in LFRL are introduced. LFRL is consistent with human cognitive science and fits well in cloud robotic systems. Experiments show that LFRL greatly improves the efficiency of reinforcement learning for robot navigation. The cloud robotic system deployment also shows that LFRL is capable of fusing prior knowledge. In addition, we release a cloud robotic navigation-learning website based on LFRL.

Yuan Xie, Boyi Liu, Qiang Liu et al.

When learning from a batch of logged bandit feedback, the discrepancy between the policy to be learned and the off-policy training data imposes statistical and computational challenges. Unlike classical supervised learning and online learning settings, in batch contextual bandit learning, one only has access to a collection of logged feedback from the actions taken by a historical policy, and expect to learn a policy that takes good actions in possibly unseen contexts. Such a batch learning setting is ubiquitous in online and interactive systems, such as ad platforms and recommendation systems. Existing approaches based on inverse propensity weights, such as Inverse Propensity Scoring (IPS) and Policy Optimizer for Exponential Models (POEM), enjoy unbiasedness but often suffer from large mean squared error. In this work, we introduce a new approach named Maximum Likelihood Inverse Propensity Scoring (MLIPS) for batch learning from logged bandit feedback. Instead of using the given historical policy as the proposal in inverse propensity weights, we estimate a maximum likelihood surrogate policy based on the logged action-context pairs, and then use this surrogate policy as the proposal. We prove that MLIPS is asymptotically unbiased, and moreover, has a smaller nonasymptotic mean squared error than IPS. Such an error reduction phenomenon is somewhat surprising as the estimated surrogate policy is less accurate than the given historical policy. Results on multi-label classification problems and a large- scale ad placement dataset demonstrate the empirical effectiveness of MLIPS. Furthermore, the proposed surrogate policy technique is complementary to existing error reduction techniques, and when combined, is able to consistently boost the performance of several widely used approaches.