Large language models (LLMs) are increasingly used in various domains, showing impressive potential on various tasks. Recently, reasoning LLMs have been proposed to improve the \textit{reasoning} or \textit{thinking} capabilities of LLMs to solve complex problems. Despite the promising results of reasoning LLMs, enhancing the multi-step reasoning capabilities of LLMs still remains a significant challenge. While existing optimization methods have advanced the LLM reasoning capabilities, they often treat reasoning trajectories as a whole, without considering the underlying critical steps within the trajectory. In this paper, we introduce \textbf{G}uided \textbf{P}ivotal \textbf{O}ptimization (GPO), a novel fine-tuning strategy that dives into the reasoning process to enable more effective improvements. GPO first identifies the `critical step' within a reasoning trajectory - a point that the model must carefully proceed so as to succeed at the problem. We locate the critical step by estimating the advantage function. GPO then resets the policy to the critical step and samples the new rollout and prioritizes learning process on those rollouts. This focus allows the model to learn more effectively from pivotal moments within the reasoning process to improve the reasoning performance. We demonstrate that GPO is not a standalone method, but rather a general strategy that can be integrated with various optimization methods to improve reasoning performance. Besides theoretical analysis, our experiments across challenging reasoning benchmarks show that GPO can consistently and significantly enhances the performance of existing optimization methods, showcasing its effectiveness and generalizability in improving LLM reasoning by concentrating on pivotal moments within the generation process.

We propose BlockScan, a customized Transformer for anomaly detection in blockchain transactions. Unlike existing methods that rely on rule-based systems or directly apply off-the-shelf large language models (LLMs), BlockScan introduces a series of customized designs to effectively model the unique data structure of blockchain transactions. First, a blockchain transaction is multi-modal, containing blockchain-specific tokens, texts, and numbers. We design a novel modularized tokenizer to handle these multi-modal inputs, balancing the information across different modalities. Second, we design a customized masked language modeling mechanism for pretraining the Transformer architecture, incorporating RoPE embedding and FlashAttention for handling longer sequences. Finally, we design a novel anomaly detection method based on the model outputs. We further provide theoretical analysis for the detection method of our system. Extensive evaluations on Ethereum and Solana transactions demonstrate BlockScan's exceptional capability in anomaly detection while maintaining a low false positive rate. Remarkably, BlockScan is the only method that successfully detects anomalous transactions on Solana with high accuracy, whereas all other approaches achieved very low or zero detection recall scores. This work sets a new benchmark for applying Transformer-based approaches in blockchain data analysis.

Deep reinforcement learning (DRL) is playing an increasingly important role in real-world applications. However, obtaining an optimally performing DRL agent for complex tasks, especially with sparse rewards, remains a significant challenge. The training of a DRL agent can be often trapped in a bottleneck without further progress. In this paper, we propose RICE, an innovative refining scheme for reinforcement learning that incorporates explanation methods to break through the training bottlenecks. The high-level idea of RICE is to construct a new initial state distribution that combines both the default initial states and critical states identified through explanation methods, thereby encouraging the agent to explore from the mixed initial states. Through careful design, we can theoretically guarantee that our refining scheme has a tighter sub-optimality bound. We evaluate RICE in various popular RL environments and real-world applications. The results demonstrate that RICE significantly outperforms existing refining schemes in enhancing agent performance.

Toxicity classification in textual content remains a significant problem. Data with labels from a single annotator fall short of capturing the diversity of human perspectives. Therefore, there is a growing need to incorporate crowdsourced annotations for training an effective toxicity classifier. Additionally, the standard approach to training a classifier using empirical risk minimization (ERM) may fail to address the potential shifts between the training set and testing set due to exploiting spurious correlations. This work introduces a novel bi-level optimization framework that integrates crowdsourced annotations with the soft-labeling technique and optimizes the soft-label weights by Group Distributionally Robust Optimization (GroupDRO) to enhance the robustness against out-of-distribution (OOD) risk. We theoretically prove the convergence of our bi-level optimization algorithm. Experimental results demonstrate that our approach outperforms existing baseline methods in terms of both average and worst-group accuracy, confirming its effectiveness in leveraging crowdsourced annotations to achieve more effective and robust toxicity classification.

Despite the promising performance of deep reinforcement learning (DRL) agents in many challenging scenarios, the black-box nature of these agents greatly limits their applications in critical domains. Prior research has proposed several explanation techniques to understand the deep learning-based policies in RL. Most existing methods explain why an agent takes individual actions rather than pinpointing the critical steps to its final reward. To fill this gap, we propose StateMask, a novel method to identify the states most critical to the agent's final reward. The high-level idea of StateMask is to learn a mask net that blinds a target agent and forces it to take random actions at some steps without compromising the agent's performance. Through careful design, we can theoretically ensure that the masked agent performs similarly to the original agent. We evaluate StateMask in various popular RL environments and show its superiority over existing explainers in explanation fidelity. We also show that StateMask has better utilities, such as launching adversarial attacks and patching policy errors.

Malware datasets inevitably contain incorrect labels due to the shortage of expertise and experience needed for sample labeling. Previous research demonstrated that a training dataset with incorrectly labeled samples would result in inaccurate model learning. To address this problem, researchers have proposed various noise learning methods to offset the impact of incorrectly labeled samples, and in image recognition and text mining applications, these methods demonstrated great success. In this work, we apply both representative and state-of-the-art noise learning methods to real-world malware classification tasks. We surprisingly observe that none of the existing methods could minimize incorrect labels’ impact. Through a carefully designed experiment, we discover that the inefficacy mainly results from extreme data imbalance and the high percentage of incorrectly labeled data samples. As such, we further propose a new noise learning method and name it after MORSE. Unlike existing methods, MORSE customizes and extends a state-of-the-art semi-supervised learning technique. It takes possibly incorrectly labeled data as unlabeled data and thus avoids their potential negative impact on model learning. In MORSE, we also integrate a sample re-weighting method that balances the training data usage in the model learning and thus handles the data imbalance challenge. We evaluate MORSE on both our synthesized and real-world datasets. We show that MORSE could significantly outperform existing noise learning methods and minimize the impact of incorrectly labeled data.

In a two-player deep reinforcement learning task, recent work shows an attacker could learn an adversarial policy that triggers a target agent to perform poorly and even react in an undesired way. However, its efficacy heavily relies upon the zero-sum assumption made in the two-player game. In this work, we propose a new adversarial learning algorithm. It addresses the problem by resetting the optimization goal in the learning process and designing a new surrogate optimization function. Our experiments show that our method significantly improves adversarial agents’ exploitability compared with the state-of-art attack. Besides, we also discover that our method could augment an agent with the ability to abuse the target game’s unfairness. Finally, we show that agents adversarially re-trained against our adversarial agents could obtain stronger adversary-resistance.

Recent research has confirmed the feasibility of backdoor attacks in deep reinforcement learning (RL) systems. However, the existing attacks require the ability to arbitrarily modify an agent's observation, constraining the application scope to simple RL systems such as Atari games. In this paper, we migrate backdoor attacks to more complex RL systems involving multiple agents and explore the possibility of triggering the backdoor without directly manipulating the agent's observation. As a proof of concept, we demonstrate that an adversary agent can trigger the backdoor of the victim agent with its own action in two-player competitive RL systems. We prototype and evaluate BackdooRL in four competitive environments. The results show that when the backdoor is activated, the winning rate of the victim drops by 17% to 37% compared to when not activated. The videos are hosted at https://github.com/wanglun1996/multi_agent_rl_backdoor_videos.

With the rapid development of deep reinforcement learning (DRL) techniques, there is an increasing need to understand and interpret DRL policies. While recent research has developed explanation methods to interpret how an agent determines its moves, they cannot capture the importance of actions/states to a game's final result. In this work, we propose a novel self-explainable model that augments a Gaussian process with a customized kernel function and an interpretable predictor. Together with the proposed model, we also develop a parameter learning procedure that leverages inducing points and variational inference to improve learning efficiency. Using our proposed model, we can predict an agent's final rewards from its game episodes and extract time step importance within episodes as strategy-level explanations for that agent. Through experiments on Atari and MuJoCo games, we verify the explanation fidelity of our method and demonstrate how to employ interpretation to understand agent behavior, discover policy vulnerabilities, remediate policy errors, and even defend against adversarial attacks.

Automatically describing open-domain videos with natural language are attracting increasing interest in the field of artificial intelligence. Most existing methods simply borrow ideas from image captioning and obtain a compact video representation from an ensemble of global image feature before feeding to an RNN decoder which outputs a sentence of variable length. However, it is not only arduous for the generator to focus on specific salient objects at different time given the global video representation, it is more formidable to capture the fine-grained motion information and the relation between moving instances for more subtle linguistic descriptions. In this paper, we propose a Trajectory Structured Attentional Encoder-Decoder (TSA-ED) neural network framework for more elaborate video captioning which works by integrating local spatial-temporal representation at trajectory level through structured attention mechanism. Our proposed method is based on a LSTM-based encoder-decoder framework, which incorporates an attention modeling scheme to adaptively learn the correlation between sentence structure and the moving objects in videos, and consequently generates more accurate and meticulous statement description in the decoding stage. Experimental results demonstrate that the feature representation and structured attention mechanism based on the trajectory cluster can efficiently obtain the local motion information in the video to help generate a more fine-grained video description, and achieve the state-of-the-art performance on the well-known Charades and MSVD datasets.

To avoid the exhaustive search over locations and scales, current state-of-the-art object detection systems usually involve a crucial component generating a batch of candidate object proposals from images. In this paper, we present a simple yet effective approach for segmenting object proposals via a deep architecture of recursive neural networks (ReNNs), which hierarchically groups regions for detecting object candidates over scales. Unlike traditional methods that mainly adopt fixed similarity measures for merging regions or finding object proposals, our approach adaptively learns the region merging similarity and the objectness measure during the process of hierarchical region grouping. Specifically, guided by a structured loss, the ReNN model jointly optimizes the cross-region similarity metric with the region merging process as well as the objectness prediction. During inference of the object proposal generation, we introduce randomness into the greedy search to cope with the ambiguity of grouping regions. Extensive experiments on standard benchmarks, e.g., PASCAL VOC and ImageNet, suggest that our approach is capable of producing object proposals with high recall while well preserving the object boundaries and outperforms other existing methods in both accuracy and efficiency.

Sketch portrait generation benefits a wide range of applications such as digital entertainment and law enforcement. Although plenty of efforts have been dedicated to this task, several issues still remain unsolved for generating vivid and detail-preserving personal sketch portraits. For example, quite a few artifacts may exist in synthesizing hairpins and glasses, and textural details may be lost in the regions of hair or mustache. Moreover, the generalization ability of current systems is somewhat limited since they usually require elaborately collecting a dictionary of examples or carefully tuning features/components. In this paper, we present a novel representation learning framework that generates an end-to-end photo-sketch mapping through structure and texture decomposition. In the training stage, we first decompose the input face photo into different components according to their representational contents (i.e., structural and textural parts) by using a pre-trained convolutional neural network (CNN). Then, we utilize a branched fully CNN for learning structural and textural representations, respectively. In addition, we design a sorted matching mean square error metric to measure texture patterns in the loss function. In the stage of sketch rendering, our approach automatically generates structural and textural representations for the input photo and produces the final result via a probabilistic fusion scheme. Extensive experiments on several challenging benchmarks suggest that our approach outperforms example-based synthesis algorithms in terms of both perceptual and objective metrics. In addition, the proposed method also has better generalization ability across data set without additional training.

Sketch-based face recognition is an interesting task in vision and multimedia research, yet it is quite challenging due to the great difference between face photos and sketches. In this paper, we propose a novel approach for photo-sketch generation, aiming to automatically transform face photos into detail-preserving personal sketches. Unlike the traditional models synthesizing sketches based on a dictionary of exemplars, we develop a fully convolutional network to learn the end-to-end photo-sketch mapping. Our approach takes whole face photos as inputs and directly generates the corresponding sketch images with efficient inference and learning, in which the architecture is stacked by only convolutional kernels of very small sizes. To well capture the person identity during the photo-sketch transformation, we define our optimization objective in the form of joint generative discriminative minimization. In particular, a discriminative regularization term is incorporated into the photo-sketch generation, enhancing the discriminability of the generated person sketches against other individuals. Extensive experiments on several standard benchmarks suggest that our approach outperforms other state-of-the-arts in both photo sketch generation and face sketch verification.