Huaimin Wang’s research while affiliated with National University of Defense Technology and other places

Agents significantly enhance the capabilities of standalone Large Language Models (LLMs) by perceiving environments, making decisions, and executing actions. However, LLM agents still face challenges in tasks that require multiple decision-making steps. Estimating the value of actions in specific tasks is difficult when intermediate actions are neither appropriately rewarded nor penalized. In this paper, we propose leveraging a task-relevant Q-value model to guide action selection. Specifically, we first collect decision-making trajectories annotated with step-level Q values via Monte Carlo Tree Search (MCTS) and construct preference data. We then use another LLM to fit these preferences through step-level Direct Policy Optimization (DPO), which serves as the Q-value model. During inference, at each decision-making step, LLM agents select the action with the highest Q value before interacting with the environment. We apply our method to various open-source and API-based LLM agents, demonstrating that Q-value models significantly improve their performance. Notably, the performance of the agent built with Phi-3-mini-4k-instruct improved by 103% on WebShop and 75% on HotPotQA when enhanced with Q-value models, even surpassing GPT-4o-mini. Additionally, Q-value models offer several advantages, such as generalization to different LLM agents and seamless integration with existing prompting strategies.

Logit-based knowledge distillation (KD) is commonly used to mitigate catastrophic forgetting in class-incremental learning (CIL) caused by data distribution shifts. However, the strict match of logit values between student and teacher models conflicts with the cross-entropy (CE) loss objective of learning new classes, leading to significant recency bias (i.e. unfairness). To address this issue, we rethink the overlooked limitations of KD-based methods through empirical analysis. Inspired by our findings, we introduce a plug-and-play pre-process method that normalizes the logits of both the student and teacher across all classes, rather than just the old classes, before distillation. This approach allows the student to focus on both old and new classes, capturing intrinsic inter-class relations from the teacher. By doing so, our method avoids the inherent conflict between KD and CE, maintaining fairness between old and new classes. Additionally, recognizing that overconfident teacher predictions can hinder the transfer of inter-class relations (i.e., dark knowledge), we extend our method to capture intra-class relations among different instances, ensuring fairness within old classes. Our method integrates seamlessly with existing logit-based KD approaches, consistently enhancing their performance across multiple CIL benchmarks without incurring additional training costs.

Instruction tuning has emerged as a paramount method for tailoring the behaviors of LLMs. Recent work has unveiled the potential for LLMs to achieve high performance through fine-tuning with a limited quantity of high-quality instruction data. Building upon this approach, we further explore the impact of prompt's robustness on the selection of high-quality instruction data. This paper proposes a pioneering framework of high-quality online instruction data mining for instruction tuning, focusing on the impact of prompt's robustness on the data mining process. Our notable innovation, is to generate the adversarial instruction data by conducting the attack for the prompt of online instruction data. Then, we introduce an Adversarial Instruction-Following Difficulty metric to measure how much help the adversarial instruction data can provide to the generation of the corresponding response. Apart from it, we propose a novel Adversarial Instruction Output Embedding Consistency approach to select high-quality online instruction data. We conduct extensive experiments on two benchmark datasets to assess the performance. The experimental results serve to underscore the effectiveness of our proposed two methods. Moreover, the results underscore the critical practical significance of considering prompt's robustness.

The emergence of deep learning (DL) has provided great opportunities for the high‐throughput analysis of atomic‐resolution micrographs. However, the DL models trained by image patches in fixed size generally lack efficiency and flexibility when processing micrographs containing diversified atomic configurations. Herein, inspired by the similarity between the atomic structures and graphs, a few‐shot learning framework based on an equivariant graph neural network (EGNN) to analyze a library of atomic structures (e.g., vacancies, phases, grain boundaries, doping, etc.) is described, showing significantly promoted robustness and three orders of magnitude reduced computing parameters compared to the image‐driven DL models, which is especially evident for those aggregated vacancy lines with flexible lattice distortion. Besides, the intuitiveness of graphs enables quantitative and straightforward extraction of the atomic‐scale structural features in batches, thus statistically unveiling the self‐assembly dynamics of vacancy lines under electron beam irradiation. A versatile model toolkit is established by integrating EGNN sub‐models for single structure recognition to process images involving varied configurations in the form of a task chain, leading to the discovery of novel doping configurations with superior electrocatalytic properties for hydrogen evolution reactions. This work provides a powerful tool to explore structure diversity in a fast, accurate, and intelligent manner.

Class-incremental learning (CIL) aims to learn a family of classes incrementally with data available in order rather than training all data at once. One main drawback of CIL is that standard deep neural networks suffer from catastrophic forgetting (CF), especially when the model only has access to data from the current incremental step. Knowledge Distillation (KD) is a widely used technique that utilizes old models as the teacher model to alleviate CF. However, based on a case study, our investigation reveals that the vanilla KD is insufficient with a strict point-to-point restriction. Instead, a relaxed match between the teacher and student improves distillation performance and model stability. In this article, we propose a simple yet effective method to mitigate CF without any additional training costs or requiring any exemplars. Specifically, we apply the linear correlation between the features of the teacher and student to measure the distillation loss rather than vanilla point-to-point loss, which significantly improves the model stability. Then, we utilize label augmentation to improve feature generalization and save prototypes to alleviate classification bias further. The proposed method significantly outperforms state-of-the-art methods in the various settings of benchmarks, including CIFAR-100 and Tiny-ImageNet, demonstrating its effectiveness and robustness.

With advancements in AI infrastructure and Trusted Execution Environment (TEE) technology, Federated Learning as a Service (FLaaS) through JointCloud Computing (JCC) is promising to break through the resource constraints caused by heterogeneous edge devices in the traditional Federated Learning (FL) paradigm. Specifically, with the protection from TEE, data owners can achieve efficient model training with high-performance AI services in the cloud. By providing additional FL services, cloud service providers can achieve collaborative learning among data owners. However, FLaaS still faces three challenges, i.e., i) low training performance caused by heterogeneous data among data owners, ii) high communication overhead among different clouds (i.e., data centers), and iii) lack of efficient resource scheduling strategies to balance training time and cost. To address these challenges, this paper presents a novel asynchronous FL approach named NebulaFL for collaborative model training among multiple clouds. To address data heterogeneity issues, NebulaFL adopts a version control-based asynchronous FL training scheme in each data center to balance training time among data owners. To reduce communication overhead, NebulaFL adopts a decentralized model rotation mechanism to achieve effective knowledge sharing among data centers. To balance training time and cost, NebulaFL integrates a reward-guided strategy for data owners selection and resource scheduling. The experimental results demonstrate that, compared to the state-of-the-art FL methods, NebulaFL can achieve up to 5.71\% accuracy improvement. In addition, NebulaFL can reduce up to 50% communication overhead and 61.94% costs under a target accuracy.

In sparse extrinsic reward settings, reinforcement learning remains a challenge despite increasing interest in this field. Existing approaches suggest that intrinsic rewards can alleviate issues caused by reward sparsity. However, many studies overlook the critical role of temporal information, essential for human curiosity. This article introduces a novel intrinsic reward mechanism inspired by human learning processes, where curiosity is evaluated by comparing current observations with historical knowledge. Our method involves training a self-supervised prediction model, periodically saving snapshots of the model parameters, and employing the nuclear norm to assess the temporal inconsistency between predictions from different snapshots as intrinsic rewards. Additionally, we propose a variational weighting mechanism to adaptively assign weights to the snapshots, enhancing the model’s robustness and performance. Experimental results across various benchmark environments demonstrate the efficacy of our approach, which outperforms other state-of-the-art methods without incurring additional training costs and exhibits higher noise tolerance. Our findings indicate that leveraging temporal information in intrinsic rewards can significantly improve exploration performance, motivating future research to develop more robust and accurate reward systems for reinforcement learning.

Large Language Models (LLMs) leverage external tools primarily through generating the API request to enhance task completion efficiency. The accuracy of API request generation significantly determines the capability of LLMs to accomplish tasks. Due to the inherent hallucinations within the LLM, it is difficult to efficiently and accurately generate the correct API request. Current research uses prompt-based feedback to facilitate the LLM-based API request generation. However, existing methods lack factual information and are insufficiently detailed. To address these issues, we propose AutoFeedback, an LLM-based framework for efficient and accurate API request generation, with a Static Scanning Component (SSC) and a Dynamic Analysis Component (DAC). SSC incorporates errors detected in the API requests as pseudo-facts into the feedback, enriching the factual information. DAC retrieves information from API documentation, enhancing the level of detail in feedback. Based on this two components, Autofeedback implementes two feedback loops during the process of generating API requests by the LLM. Extensive experiments demonstrate that it significantly improves accuracy of API request generation and reduces the interaction cost. AutoFeedback achieves an accuracy of 100.00\% on a real-world API dataset and reduces the cost of interaction with GPT-3.5 Turbo by 23.44\%, and GPT-4 Turbo by 11.85\%.