Zhiheng Xi

I am a second-year master student at Fudan NLP Group of Computer School, Fudan University, where I work on natural language processing (NLP) and deep learning (DL). I am advised by Prof. Tao Gui, Prof. Qi Zhang, and Prof. Xuanjing Huang. Previously, I got my bachelor's degree from Nanjing University, advised by Prof. Jia Liu. I'm honored to be interning at Shanghai AI lab currently. In 2021, I completed a fantastic internship at Microsoft Azure, advised by Jiaye Wu and Hang Zhang.

Email  /  Google Scholar  /  Github  /  Twitter  /  Zhihu

Research

I have general interest in deep learning, natural language processing, and robust machine learning. Recently, I focus my research on large language models (LLMs) and LLM-based agents.

We propose AgentGym, a framework designed to help the community easily evaluate and develop generally-capable LLM-based agents. It features diverse interactive environments and tasks with a unified format. It supports real-time feedback and concurrency, and is easily scalable. It also includes a high-quality trajectory set AgentTraj and a benchmark suite AgentEval.

To study the evolution potential of general LLM-based agents, we propose a novel method, AgentEvol, to investigate the potential of agent self-evolution beyond previously seen data across tasks and environments.

In this paper, we propose \(R^3\): Learning Reasoning through Reverse Curriculum Reinforcement Learning (RL), a novel method that employs only outcome supervision to achieve the benefits of process supervision for large language models.

In this paper, we provide a comprehensive survey of LLM-based agents with 86 pages. We start by tracing the concept of agents from its philosophical origins to its development in AI. Next, the main body includes the construction of LLM-based agents, its extensive applications, and the essential concept of Agent society. Finally, we discuss a range of key topics and open problems within the field, e.g., scaling number of agentsand Agent-as-a-service.

Different from previous work like Chain-of-Thought (CoT) which enhance LLMs' reasoning performance from the answer/reasoning side , we start from the problem side and propose Self-Polish (SP). It is a novel method that facilitates the model’s reasoning by guiding it to progressively refine the given problems to be more comprehensible and solvable.

In this work, we draw inspiration from the operating mechanism of deep neural networks (DNNs) and biological brains, where neuronal activations are sparse and task-specific, and we use the connectivity patterns of neurons as a unique identifier (Task Embeddings) associated with the task. Experiments show that our method consistently outperforms other baselines in predicting inter-task transferability across data regimes and transfer settings, while keeping high efficiency in computation and storage.

Adversarial training, a strong algorithm to enhance model robustness, is typically more expensive than traditional fine-tuning because of the necessity to generate adversarial examples via gradient descent. Delving into the optimization process of adversarial training, we find that robust connectivity patterns emerge in the early training phase (typically 0.15~0.3 epochs), far before parameters converge. Inspired by this finding, we dig out robust early-bird tickets (i.e., subnetworks) to develop an efficient adversarial training method.

In this paper, we show that robust and non-robust instances in the training dataset, though are both important for test performance, have contrary impacts on robustness, which makes it possible to build a highly robust model by leveraging the training dataset in a more effective way. We propose a new method that can distinguish between robust instances from non-robust ones according to the model’s sensitivity to perturbations on individual instances during training. Surprisingly, we find that the model under standard training easily overfits the robust instances by relying on their simple patterns before the model completely learns their robust features. Finally, we propose a new mitigation algorithm to further release the potential of robust instances.

In this work, we propose a novel approach that can learn a consistent policy via RL across various data groups or domains. Given the challenges associated with acquiring group annotations, our method automatically classifies data into different groups, deliberately maximizing performance variance. Then, we optimize the policy to perform well on challenging groups. Lastly, leveraging the established groups, our approach adaptively adjusts the exploration space, allocating more learning capacity to more challenging data and preventing the model from over-optimizing on simpler data.

EasyJailbreak is an easy-to-use Python framework designed for researchers and developers focusing on LLM security. Specifically, EasyJailbreak decomposes the mainstream jailbreaking process into several iterable steps: initialize mutation seeds, select suitable seeds, add constraint, mutate, attack, and evaluate. On this basis, EasyJailbreak provides a component for each step, constructing a playground for further research and attempts. More details can be found in our paper.

In this report, we attempt to address issues of reward modeling in RLHF of LLMs. (1) From a data perspective, we propose a method to measure the strength of preferences within the data, based on a voting mechanism of multiple reward models. (2) From an algorithmic standpoint, we introduce contrastive learning to enhance the ability of reward models to distinguish between chosen and rejected responses, thereby improving model generalization. Furthermore, we employ meta-learning to enable the reward model to maintain the ability to differentiate subtle differences in out-of-distribution samples.

In this technical report, we dissect the framework of RLHF, re-evaluate the inner workings of PPO, and explore how the parts comprising PPO algorithms impact policy agent training. We identify policy constraints being the key factor for the effective implementation of the PPO algorithm. Therefore, we explore the PPO-max, an advanced version of PPO algorithm, to efficiently improve the training stability of the policy model. Based on our main results, we perform a comprehensive analysis of RLHF abilities compared with SFT models and ChatGPT.

In this paper, we utilize instructions in code style, which are more structural and less ambiguous, to replace typically natural language instructions. Through this conversion, we provide LLMs with more precise instructions and strengthen the robustness of LLMs. Moreover, under few-shot scenarios, we propose a novel method to compose in-context demonstrations using both clean and adversarial samples ({adversarial context method}) to further boost the robustness of the LLMs.

Miscellanea

• I'm passionate about FPS games, including Counter-Strike: Global Offensive (CS:GO / CS2) and CrossFire (CF).
• I love watching soccer and am a big fan of Mourinho.
• I also love watching basketball games and my favorite player is Kevin Durant.