Jie Fu

Zejun Li, Ye Wang, Mengfei Du et al.

Recent years have witnessed remarkable progress in the development of large vision-language models (LVLMs). Benefiting from the strong language backbones and efficient cross-modal alignment strategies, LVLMs exhibit surprising capabilities to perceive visual signals and perform visually grounded reasoning. However, the capabilities of LVLMs have not been comprehensively and quantitatively evaluate. Most existing multi-modal benchmarks require task-oriented input-output formats, posing great challenges to automatically assess the free-form text output of LVLMs. To effectively leverage the annotations available in existing benchmarks and reduce the manual effort required for constructing new benchmarks, we propose to re-formulate existing benchmarks into unified LVLM-compatible formats. Through systematic data collection and reformulation, we present the ReForm-Eval benchmark, offering substantial data for evaluating various capabilities of LVLMs. Based on ReForm-Eval, we conduct extensive experiments, thoroughly analyze the strengths and weaknesses of existing LVLMs, and identify the underlying factors. Our benchmark and evaluation framework will be open-sourced as a cornerstone for advancing the development of LVLMs.

Zihan Qiu, Zhen Liu, Shuicheng Yan et al.

It has been shown that semi-parametric methods, which combine standard neural networks with non-parametric components such as external memory modules and data retrieval, are particularly helpful in data scarcity and out-of-distribution (OOD) scenarios. However, existing semi-parametric methods mostly depend on independent raw data points - this strategy is difficult to scale up due to both high computational costs and the incapacity of current attention mechanisms with a large number of tokens. In this paper, we introduce a novel heterogeneous memory augmentation approach for neural networks which, by introducing learnable memory tokens with attention mechanism, can effectively boost performance without huge computational overhead. Our general-purpose method can be seamlessly combined with various backbones (MLP, CNN, GNN, and Transformer) in a plug-and-play manner. We extensively evaluate our approach on various image and graph-based tasks under both in-distribution (ID) and OOD conditions and show its competitive performance against task-specific state-of-the-art methods. Code is available at \url{https://github.com/qiuzh20/HMA}.

Xinyi Wei, Shuo Han, Ahmed H. Hemida et al.

This paper investigates backdoor attack planning in stochastic control systems modeled as Markov Decision Processes (MDPs). A backdoor attack involves an adversary deploying a policy that performs well in the original MDP to pass testing, but behaves maliciously at runtime when combined with a trigger that perturbs system dynamics. We consider a sophisticated attacker capable of jointly optimizing the backdoor policy and its trigger using only a blackbox simulator. During execution, the attacker has access only to partial observations of the system state and is restricted to introduce small perturbations to the system's transition dynamics. We formulate the attack planning problem as a constrained Markov game with an augmented state space and two players: Player 0 learns a backdoor policy that maximizes attack rewards when the trigger is active. However, when the trigger is inactive, the backdoor policy behaves near-optimally in the original MDP; Player 1 designs a finite-memory, observation-based trigger to activate the attack. We propose a switching gradient-based optimization algorithm to jointly solve for the backdoor policy and trigger. Experiments on a case study demonstrate the effectiveness of our method in achieving stealthy and successful backdoor attacks, and how the attack performance varies under different parameters related to the stealthiness of the backdoor attack.

Rongao Li, Jie Fu, Bo-Wen Zhang et al.

We introduce TACO, an open-source, large-scale code generation dataset, with a focus on the optics of algorithms, designed to provide a more challenging training dataset and evaluation benchmark in the field of code generation models. TACO includes competition-level programming questions that are more challenging, to enhance or evaluate problem understanding and reasoning abilities in real-world programming scenarios. There are 25433 and 1000 coding problems in training and test set, as well as up to 1.55 million diverse solution answers. Moreover, each TACO problem includes several fine-grained labels such as task topics, algorithms, programming skills, and difficulty levels, providing a more precise reference for the training and evaluation of code generation models. The dataset and evaluation scripts are available on Hugging Face Hub (https://huggingface.co/datasets/BAAI/TACO) and Github (https://github.com/FlagOpen/TACO).

Ruiyuan Zhang, Jiaxiang Liu, Zexi Li et al.

Geometric fracture assembly presents a challenging practical task in archaeology and 3D computer vision. Previous methods have focused solely on assembling fragments based on semantic information, which has limited the quantity of objects that can be effectively assembled. Therefore, there is a need to develop a scalable framework for geometric fracture assembly without relying on semantic information. To improve the effectiveness of assembling geometric fractures without semantic information, we propose a co-creation space comprising several assemblers capable of gradually and unambiguously assembling fractures. Additionally, we introduce a novel loss function, i.e., the geometric-based collision loss, to address collision issues during the fracture assembly process and enhance the results. Our framework exhibits better performance on both PartNet and Breaking Bad datasets compared to existing state-of-the-art frameworks. Extensive experiments and quantitative comparisons demonstrate the effectiveness of our proposed framework, which features linear computational complexity, enhanced abstraction, and improved generalization. Our code is publicly available at https://github.com/Ruiyuan-Zhang/CCS.

Haoxiang Ma, Chongyang Shi, Shuo Han et al.

Covert planning refers to a class of constrained planning problems where an agent aims to accomplish a task with minimal information leaked to a passive observer to avoid detection. However, existing methods of covert planning often consider deterministic environments or do not exploit the observer's imperfect information. This paper studies how covert planning can leverage the coupling of stochastic dynamics and the observer's imperfect observation to achieve optimal task performance without being detected. Specifically, we employ a Markov decision process to model the interaction between the agent and its stochastic environment, and a partial observation function to capture the leaked information to a passive observer. Assuming the observer employs hypothesis testing to detect if the observation deviates from a nominal policy, the covert planning agent aims to maximize the total discounted reward while keeping the probability of being detected as an adversary below a given threshold. We prove that finite-memory policies are more powerful than Markovian policies in covert planning. Then, we develop a primal-dual proximal policy gradient method with a two-time-scale update to compute a (locally) optimal covert policy. We demonstrate the effectiveness of our methods using a stochastic gridworld example. Our experimental results illustrate that the proposed method computes a policy that maximizes the adversary's expected reward without violating the detection constraint, and empirically demonstrates how the environmental noises can influence the performance of the covert policies.

Shuai Zhao, Leilei Gan, Luu Anh Tuan et al. · mit

Recently, various parameter-efficient fine-tuning (PEFT) strategies for application to language models have been proposed and successfully implemented. However, this raises the question of whether PEFT, which only updates a limited set of model parameters, constitutes security vulnerabilities when confronted with weight-poisoning backdoor attacks. In this study, we show that PEFT is more susceptible to weight-poisoning backdoor attacks compared to the full-parameter fine-tuning method, with pre-defined triggers remaining exploitable and pre-defined targets maintaining high confidence, even after fine-tuning. Motivated by this insight, we developed a Poisoned Sample Identification Module (PSIM) leveraging PEFT, which identifies poisoned samples through confidence, providing robust defense against weight-poisoning backdoor attacks. Specifically, we leverage PEFT to train the PSIM with randomly reset sample labels. During the inference process, extreme confidence serves as an indicator for poisoned samples, while others are clean. We conduct experiments on text classification tasks, five fine-tuning strategies, and three weight-poisoning backdoor attack methods. Experiments show near 100% success rates for weight-poisoning backdoor attacks when utilizing PEFT. Furthermore, our defensive approach exhibits overall competitive performance in mitigating weight-poisoning backdoor attacks.

Songtao Feng, Jie Fu

Reinforcement learning from human feedback (RLHF) has achieved great empirical success in aligning large language models (LLMs) with human preference, and it is of great importance to study the statistical efficiency of RLHF algorithms from a theoretical perspective. In this work, we consider the online RLHF setting where the preference data is revealed during the learning process and study action value function approximation. We design a model-free posterior sampling algorithm for online RLHF inspired by Thompson sampling and provide its theoretical guarantee. Specifically, we adopt Bellman eluder (BE) dimension as the complexity measure of the function class and establish $O(\sqrt{T})$ regret bound for the proposed algorithm with other multiplicative factor depending on the horizon, BE dimension and the $log$-bracketing number of the function class. Further, in the analysis, we first establish the concentration-type inequality of the squared Bellman error bound based on the maximum likelihood estimator (MLE) generalization bound, which plays the crucial rules in obtaining the eluder-type regret bound and may be of independent interest.

Haiteng Zhao, Shengchao Liu, Chang Ma et al.

Molecule property prediction has gained significant attention in recent years. The main bottleneck is the label insufficiency caused by expensive lab experiments. In order to alleviate this issue and to better leverage textual knowledge for tasks, this study investigates the feasibility of employing natural language instructions to accomplish molecule-related tasks in a zero-shot setting. We discover that existing molecule-text models perform poorly in this setting due to inadequate treatment of instructions and limited capacity for graphs. To overcome these issues, we propose GIMLET, which unifies language models for both graph and text data. By adopting generalized position embedding, our model is extended to encode both graph structures and instruction text without additional graph encoding modules. GIMLET also decouples encoding of the graph from tasks instructions in the attention mechanism, enhancing the generalization of graph features across novel tasks. We construct a dataset consisting of more than two thousand molecule tasks with corresponding instructions derived from task descriptions. We pretrain GIMLET on the molecule tasks along with instructions, enabling the model to transfer effectively to a broad range of tasks. Experimental results demonstrate that GIMLET significantly outperforms molecule-text baselines in instruction-based zero-shot learning, even achieving closed results to supervised GNN models on tasks such as toxcast and muv.

Xuan Liu, Cagdas Onal, Jie Fu

In this paper, we present a model-free learning-based control scheme for the soft snake robot to improve its contact-aware locomotion performance in a cluttered environment. The control scheme includes two cooperative controllers: A bio-inspired controller (C1) that controls both the steering and velocity of the soft snake robot, and an event-triggered regulator (R2) that controls the steering of the snake in anticipation of obstacle contacts and during contact. The inputs from the two controllers are composed as the input to a Matsuoka CPG network to generate smooth and rhythmic actuation inputs to the soft snake. To enable stable and efficient learning with two controllers, we develop a game-theoretic process, fictitious play, to train C1 and R2 with a shared potential-field-based reward function for goal tracking tasks. The proposed approach is tested and evaluated in the simulator and shows significant improvement of locomotion performance in the obstacle-based environment comparing to two baseline controllers.

Jie Fu

We present a formal language for specifying qualitative preferences over temporal goals and a preference-based planning method in stochastic systems. Using automata-theoretic modeling, the proposed specification allows us to express preferences over different sets of outcomes, where each outcome describes a set of temporal sequences of subgoals. We define the value of preference satisfaction given a stochastic process over possible outcomes and develop an algorithm for time-constrained probabilistic planning in labeled Markov decision processes where an agent aims to maximally satisfy its preference formula within a pre-defined finite time duration. We present experimental results using a stochastic gridworld example and discuss possible extensions of the proposed preference model.

Haoxiang Ma, Jie Fu

Attention control is a key cognitive ability for humans to select information relevant to the current task. This paper develops a computational model of attention and an algorithm for attention-based probabilistic planning in Markov decision processes. In attention-based planning, the robot decides to be in different attention modes. An attention mode corresponds to a subset of state variables monitored by the robot. By switching between different attention modes, the robot actively perceives task-relevant information to reduce the cost of information acquisition and processing, while achieving near-optimal task performance. Though planning with attention-based active perception inevitably introduces partial observations, a partially observable MDP formulation makes the problem computational expensive to solve. Instead, our proposed method employs a hierarchical planning framework in which the robot determines what to pay attention to and for how long the attention should be sustained before shifting to other information sources. During the attention sustaining phase, the robot carries out a sub-policy, computed from an abstraction of the original MDP given the current attention. We use an example where a robot is tasked to capture a set of intruders in a stochastic gridworld. The experimental results show that the proposed method enables information- and computation-efficient optimal planning in stochastic environments.

Zhentian Qian, Kartik Patath, Jie Fu et al.

It is often desirable to capture and map semantic information of an environment during simultaneous localization and mapping (SLAM). Such semantic information can enable a robot to better distinguish places with similar low-level geometric and visual features and perform high-level tasks that use semantic information about objects to be manipulated and environments to be navigated. While semantic SLAM has gained increasing attention, there is little research on semanticlevel data association based on semantic objects, i.e., object-level data association. In this paper, we propose a novel object-level data association algorithm based on bag of words algorithm, formulated as a maximum weighted bipartite matching problem. With object-level data association solved, we develop a quadratic-programming-based semantic object initialization scheme using dual quadric and introduce additional constraints to improve the success rate of object initialization. The integrated semantic-level SLAM system can achieve high-accuracy object-level data association and real-time semantic mapping as demonstrated in the experiments. The online semantic map building and semantic-level localization capabilities facilitate semantic-level mapping and task planning in a priori unknown environment.

Abhishek N. Kulkarni, Jie Fu

In this chapter, we present an approach using formal methods to synthesize reactive defense strategy in a cyber network, equipped with a set of decoy systems. We first generalize formal graphical security models--attack graphs--to incorporate defender's countermeasures in a game-theoretic model, called an attack-defend game on graph. This game captures the dynamic interactions between the defender and the attacker and their defense/attack objectives in formal logic. Then, we introduce a class of hypergames to model asymmetric information created by decoys in the attacker-defender interactions. Given qualitative security specifications in formal logic, we show that the solution concepts from hypergames and reactive synthesis in formal methods can be extended to synthesize effective dynamic defense strategy using cyber deception. The strategy takes the advantages of the misperception of the attacker to ensure security specification is satisfied, which may not be satisfiable when the information is symmetric.

Shaohua Li, Xiuchao Sui, Jie Fu et al.

Convolutional Neural Networks (CNNs) are known to be brittle under various image transformations, including rotations, scalings, and changes of lighting conditions. We observe that the features of a transformed image are drastically different from the ones of the original image. To make CNNs more invariant to transformations, we propose "Feature Lenses", a set of ad-hoc modules that can be easily plugged into a trained model (referred to as the "host model"). Each individual lens reconstructs the original features given the features of a transformed image under a particular transformation. These lenses jointly counteract feature distortions caused by various transformations, thus making the host model more robust without retraining. By only updating lenses, the host model is freed from iterative updating when facing new transformations absent in the training data; as feature semantics are preserved, downstream applications, such as classifiers and detectors, automatically gain robustness without retraining. Lenses are trained in a self-supervised fashion with no annotations, by minimizing a novel "Top-K Activation Contrast Loss" between lens-transformed features and original features. Evaluated on ImageNet, MNIST-rot, and CIFAR-10, Feature Lenses show clear advantages over baseline methods.

Zhentian Qian, Jie Fu, Quanyan Zhu

In this paper, we study the problem of assessing the effectiveness of a proactive defense-by-detection policy with a network-based moving target defense. We model the network system using a probabilistic attack graph--a graphical security model. Given a network system with a proactive defense strategy, an intelligent attacker needs to perform reconnaissance repeatedly to learn about the locations of intrusion detection systems and re-plan optimally to reach the target while avoiding detection. To compute the attacker's strategy for security evaluation, we develop a receding-horizon planning algorithm using a risk-sensitive Markov decision process with a time-varying reward function. Finally, we implement both defense and attack strategies in a synthetic network and analyze how the frequency of network randomization and the number of detection systems can influence the success rate of the attacker. This study provides insights for designing proactive defense strategies against online and multi-stage attacks by a resourceful attacker.

Xuan Liu, Renato Gasoto, Cagdas Onal et al.

In this paper, we present a new locomotion control method for soft robot snakes. Inspired by biological snakes, our control architecture is composed of two key modules: A deep reinforcement learning (RL) module for achieving adaptive goal-tracking behaviors with changing goals, and a central pattern generator (CPG) system with Matsuoka oscillators for generating stable and diverse locomotion patterns. The two modules are interconnected into a closed-loop system: The RL module, analogizing the locomotion region located in the midbrain of vertebrate animals, regulates the input to the CPG system given state feedback from the robot. The output of the CPG system is then translated into pressure inputs to pneumatic actuators of the soft snake robot. Based on the fact that the oscillation frequency and wave amplitude of the Matsuoka oscillator can be independently controlled under different time scales, we further adapt the option-critic framework to improve the learning performance measured by optimality and data efficiency. The performance of the proposed controller is experimentally validated with both simulated and real soft snake robots.

Renato Gasoto, Miles Macklin, Xuan Liu et al.

In this work we present a framework that is capable of accurately representing soft robotic actuators in a multiphysics environment in real-time. We propose a constraint-based dynamics model of a 1-dimensional pneumatic soft actuator that accounts for internal pressure forces, as well as the effect of actuator latency and damping under inflation and deflation and demonstrate its accuracy a full soft robotic snake with the composition of multiple 1D actuators. We verify our model's accuracy in static deformation and dynamic locomotion open-loop control experiments. To achieve real-time performance we leverage the parallel computation power of GPUs to allow interactive control and feedback.

Xuan Liu, Jie Fu

In hierarchical planning for Markov decision processes (MDPs), temporal abstraction allows planning with macro-actions that take place at different time scale in form of sequential composition. In this paper, we propose a novel approach to compositional reasoning and hierarchical planning for MDPs under temporal logic constraints. In addition to sequential composition, we introduce a composition of policies based on generalized logic composition: Given sub-policies for sub-tasks and a new task expressed as logic compositions of subtasks, a semi-optimal policy, which is optimal in planning with only sub-policies, can be obtained by simply composing sub-polices. Thus, a synthesis algorithm is developed to compute optimal policies efficiently by planning with primitive actions, policies for sub-tasks, and the compositions of sub-policies, for maximizing the probability of satisfying temporal logic specifications. We demonstrate the correctness and efficiency of the proposed method in stochastic planning examples with a single agent and multiple task specifications.

Michael L. Littman, Ufuk Topcu, Jie Fu et al.

We propose a new task-specification language for Markov decision processes that is designed to be an improvement over reward functions by being environment independent. The language is a variant of Linear Temporal Logic (LTL) that is extended to probabilistic specifications in a way that permits approximations to be learned in finite time. We provide several small environments that demonstrate the advantages of our geometric LTL (GLTL) language and illustrate how it can be used to specify standard reinforcement-learning tasks straightforwardly.

Jie Fu

In this paper, we propose a sampling-based planning and optimal control method of nonlinear systems under non-differentiable constraints. Motivated by developing scalable planning algorithms, we consider the optimal motion plan to be a feedback controller that can be approximated by a weighted sum of given bases. Given this approximate optimal control formulation, our main contribution is to introduce importance sampling, specifically, model-reference adaptive search algorithm, to iteratively compute the optimal weight parameters, i.e., the weights corresponding to the optimal policy function approximation given chosen bases. The key idea is to perform the search by iteratively estimating a parametrized distribution which converges to a Dirac's Delta that infinitely peaks on the global optimal weights. Then, using this direct policy search, we incorporated trajectory-based verification to ensure that, for a class of nonlinear systems, the obtained policy is not only optimal but robust to bounded disturbances. The correctness and efficiency of the methods are demonstrated through numerical experiments including linear systems with a nonlinear cost function and motion planning for a Dubins car.

Jie Fu, Ufuk Topcu

This paper studies an optimal control problem for continuous-time stochastic systems subject to reachability objectives specified in a subclass of metric interval temporal logic specifications, a temporal logic with real-time constraints. We propose a probabilistic method for synthesizing an optimal control policy that maximizes the probability of satisfying a specification based on a discrete approximation of the underlying stochastic system. First, we show that the original problem can be formulated as a stochastic optimal control problem in a state space augmented with finite memory and states of some clock variables. Second, we present a numerical method for computing an optimal policy with which the given specification is satisfied with the maximal probability in point-based semantics in the discrete approximation of the underlying system. We show that the policy obtained in the discrete approximation converges to the optimal one for satisfying the specification in the continuous or dense-time semantics as the discretization becomes finer in both state and time. Finally, we illustrate our approach with a robotic motion planning example.

Jie Fu, Shuo Han, Ufuk Topcu

Optimal control synthesis in stochastic systems with respect to quantitative temporal logic constraints can be formulated as linear programming problems. However, centralized synthesis algorithms do not scale to many practical systems. To tackle this issue, we propose a decomposition-based distributed synthesis algorithm. By decomposing a large-scale stochastic system modeled as a Markov decision process into a collection of interacting sub-systems, the original control problem is formulated as a linear programming problem with a sparse constraint matrix, which can be solved through distributed optimization methods. Additionally, we propose a decomposition algorithm which automatically exploits, if exists, the modular structure in a given large-scale system. We illustrate the proposed methods through robotic motion planning examples.

Jie Fu, Ufuk Topcu

In systems in which control authority is shared by an autonomous controller and a human operator, it is important to find solutions that achieve a desirable system performance with a reasonable workload for the human operator. We formulate a shared autonomy system capable of capturing the interaction and switching control between an autonomous controller and a human operator, as well as the evolution of the operator's cognitive state during control execution. To trade-off human's effort and the performance level, e.g., measured by the probability of satisfying the underlying temporal logic specification, a two-stage policy synthesis algorithm is proposed for generating Pareto efficient coordination and control policies with respect to user specified weights. We integrate the Tchebychev scalarization method for multi-objective optimization methods to obtain a better coverage of the set of Pareto efficient solutions than linear scalarization methods.

Jie Fu, Ufuk Topcu

We consider synthesis of control policies that maximize the probability of satisfying given temporal logic specifications in unknown, stochastic environments. We model the interaction between the system and its environment as a Markov decision process (MDP) with initially unknown transition probabilities. The solution we develop builds on the so-called model-based probably approximately correct Markov decision process (PAC-MDP) methodology. The algorithm attains an $\varepsilon$-approximately optimal policy with probability $1-δ$ using samples (i.e. observations), time and space that grow polynomially with the size of the MDP, the size of the automaton expressing the temporal logic specification, $\frac{1}{\varepsilon}$, $\frac{1}δ$ and a finite time horizon. In this approach, the system maintains a model of the initially unknown MDP, and constructs a product MDP based on its learned model and the specification automaton that expresses the temporal logic constraints. During execution, the policy is iteratively updated using observation of the transitions taken by the system. The iteration terminates in finitely many steps. With high probability, the resulting policy is such that, for any state, the difference between the probability of satisfying the specification under this policy and the optimal one is within a predefined bound.