Don't Command, Cultivate: An Exploratory Study of System-2 Alignment

The o1 system card identifies the o1 models as the most robust within OpenAI, with their defining characteristic being the progression from rapid, intuitive thinking to slower, more deliberate reasoning. This observation motivated us to investigate the influence of System-2 thinking patterns on model safety. In our preliminary research, we conducted safety evaluations of the o1 model, including complex jailbreak attack scenarios using adversarial natural language prompts and mathematical encoding prompts. Our findings indicate that the o1 model demonstrates relatively improved safety performance; however, it still exhibits vulnerabilities, particularly against jailbreak attacks employing mathematical encoding. Through detailed case analysis, we identified specific patterns in the o1 model's responses. We also explored the alignment of System-2 safety in open-source models using prompt engineering and supervised fine-tuning techniques. Experimental results show that some simple methods to encourage the model to carefully scrutinize user requests are beneficial for model safety. Additionally, we proposed a implementation plan for process supervision to enhance safety alignment. The implementation details and experimental results will be provided in future versions.

Enhancing Cluster Resilience: LLM-agent Based Autonomous Intelligent Cluster Diagnosis System and Evaluation Framework

Recent advancements in Large Language Models (LLMs) and related technologies such as Retrieval-Augmented Generation (RAG) and Diagram of Thought (DoT) have enabled the creation of autonomous intelligent systems capable of performing cluster diagnostics and troubleshooting. By integrating these technologies with self-play methodologies, we have developed an LLM-agent system designed to autonomously diagnose and resolve issues within AI clusters. Our innovations include a knowledge base tailored for cluster diagnostics, enhanced LLM algorithms, practical deployment strategies for agents, and a benchmark specifically designed for evaluating LLM capabilities in this domain. Through extensive experimentation across multiple dimensions, we have demonstrated the superiority of our system in addressing the challenges faced in cluster diagnostics, particularly in detecting and rectifying performance issues more efficiently and accurately than traditional methods.

KG-FPQ: Evaluating Factuality Hallucination in LLMs with Knowledge Graph-based False Premise Questions

Recent studies have demonstrated that large language models (LLMs) are susceptible to being misled by false premise questions (FPQs), leading to errors in factual knowledge, know as factuality hallucination. Existing benchmarks that assess this vulnerability primarily rely on manual construction, resulting in limited scale and lack of scalability. In this work, we introduce an automated, scalable pipeline to create FPQs based on knowledge graphs (KGs). The first step is modifying true triplets extracted from KGs to create false premises. Subsequently, utilizing the state-of-the-art capabilities of GPTs, we generate semantically rich FPQs. Based on the proposed method, we present a comprehensive benchmark, the Knowledge Graph-based False Premise Questions (KG-FPQ), which contains approximately 178k FPQs across three knowledge domains, at six levels of confusability, and in two task formats. Using KG-FPQ, we conduct extensive evaluations on several representative LLMs and provide valuable insights. The KG-FPQ dataset and code are available at~https://github.com/yanxuzhu/KG-FPQ.

Explainable Graph Neural Networks Under Fire

Predictions made by graph neural networks (GNNs) usually lack interpretability due to their complex computational behavior and the abstract nature of graphs. In an attempt to tackle this, many GNN explanation methods have emerged. Their goal is to explain a model's predictions and thereby obtain trust when GNN models are deployed in decision critical applications. Most GNN explanation methods work in a post-hoc manner and provide explanations in the form of a small subset of important edges and/or nodes. In this paper we demonstrate that these explanations can unfortunately not be trusted, as common GNN explanation methods turn out to be highly susceptible to adversarial perturbations. That is, even small perturbations of the original graph structure that preserve the model's predictions may yield drastically different explanations. This calls into question the trustworthiness and practical utility of post-hoc explanation methods for GNNs. To be able to attack GNN explanation models, we devise a novel attack method dubbed \textit{GXAttack}, the first \textit{optimization-based} adversarial attack method for post-hoc GNN explanations under such settings. Due to the devastating effectiveness of our attack, we call for an adversarial evaluation of future GNN explainers to demonstrate their robustness.

Improving Weak-to-Strong Generalization with Scalable Oversight and Ensemble Learning

This paper presents a follow-up study to OpenAI's recent superalignment work on Weak-to-Strong Generalization (W2SG). Superalignment focuses on ensuring that high-level AI systems remain consistent with human values and intentions when dealing with complex, high-risk tasks. The W2SG framework has opened new possibilities for empirical research in this evolving field. Our study simulates two phases of superalignment under the W2SG framework: the development of general superhuman models and the progression towards superintelligence. In the first phase, based on human supervision, the quality of weak supervision is enhanced through a combination of scalable oversight and ensemble learning, reducing the capability gap between weak teachers and strong students. In the second phase, an automatic alignment evaluator is employed as the weak supervisor. By recursively updating this auto aligner, the capabilities of the weak teacher models are synchronously enhanced, achieving weak-to-strong supervision over stronger student models.We also provide an initial validation of the proposed approach for the first phase. Using the SciQ task as example, we explore ensemble learning for weak teacher models through bagging and boosting. Scalable oversight is explored through two auxiliary settings: human-AI interaction and AI-AI debate. Additionally, the paper discusses the impact of improved weak supervision on enhancing weak-to-strong generalization based on in-context learning. Experiment code and dataset will be released at https://github.com/ADaM-BJTU/W2SG.

CDEval: A Benchmark for Measuring the Cultural Dimensions of Large Language Models

As the scaling of Large Language Models (LLMs) has dramatically enhanced their capabilities, there has been a growing focus on the alignment problem to ensure their responsible and ethical use. While existing alignment efforts predominantly concentrate on universal values such as the HHH principle, the aspect of culture, which is inherently pluralistic and diverse, has not received adequate attention. This work introduces a new benchmark, CDEval, aimed at evaluating the cultural dimensions of LLMs. CDEval is constructed by incorporating both GPT-4's automated generation and human verification, covering six cultural dimensions across seven domains. Our comprehensive experiments provide intriguing insights into the culture of mainstream LLMs, highlighting both consistencies and variations across different dimensions and domains. The findings underscore the importance of integrating cultural considerations in LLM development, particularly for applications in diverse cultural settings. Through CDEval, we aim to broaden the horizon of LLM alignment research by including cultural dimensions, thus providing a more holistic framework for the future development and evaluation of LLMs. This benchmark serves as a valuable resource for cultural studies in LLMs, paving the way for more culturally aware and sensitive models.

An LLM-free Multi-dimensional Benchmark for MLLMs Hallucination Evaluation

Despite making significant progress in multi-modal tasks, current Multi-modal Large Language Models (MLLMs) encounter the significant challenge of hallucination, which may lead to harmful consequences. Therefore, evaluating MLLMs' hallucinations is becoming increasingly important in model improvement and practical application deployment. Previous works are limited in high evaluation costs (e.g., relying on humans or advanced LLMs) and insufficient evaluation dimensions (e.g., types of hallucination and task). In this paper, we propose an LLM-free multi-dimensional benchmark AMBER, which can be used to evaluate both generative task and discriminative task including object existence, object attribute and object relation hallucination. Based on AMBER, we design a low-cost and efficient evaluation pipeline. Additionally, we conduct a comprehensive evaluation and detailed analysis of mainstream MLLMs including GPT-4V(ision), and also give guideline suggestions for mitigating hallucinations. The data and code of AMBER are available at https://github.com/junyangwang0410/AMBER.

A Transfer-Learning-Based Prognosis Prediction Paradigm that Bridges Data Distribution Shift across EMR Datasets

Due to the limited information about emerging diseases, symptoms are hard to be noticed and recognized, so that the window for clinical intervention could be ignored. An effective prognostic model is expected to assist doctors in making right diagnosis and designing personalized treatment plan, so to promptly prevent unfavorable outcomes. However, in the early stage of a disease, limited data collection and clinical experiences, plus the concern out of privacy and ethics, may result in restricted data availability for reference, to the extent that even data labels are difficult to mark correctly. In addition, Electronic Medical Record (EMR) data of different diseases or of different sources of the same disease can prove to be having serious cross-dataset feature misalignment problems, greatly mutilating the efficiency of deep learning models. This article introduces a transfer learning method to build a transition model from source dataset to target dataset. By way of constraining the distribution shift of features generated in disparate domains, domain-invariant features that are exclusively relative to downstream tasks are captured, so to cultivate a unified domain-invariant encoder across various task domains to achieve better feature representation. Experimental results of several target tasks demonstrate that our proposed model outperforms competing baseline methods and has higher rate of training convergence, especially in dealing with limited data amount. A multitude of experiences have proven the efficacy of our method to provide more accurate predictions concerning newly emergent pandemics and other diseases.

Recovering a Molecule's 3D Dynamics from Liquid-phase Electron Microscopy Movies

The dynamics of biomolecules are crucial for our understanding of their functioning in living systems. However, current 3D imaging techniques, such as cryogenic electron microscopy (cryo-EM), require freezing the sample, which limits the observation of their conformational changes in real time. The innovative liquid-phase electron microscopy (liquid-phase EM) technique allows molecules to be placed in the native liquid environment, providing a unique opportunity to observe their dynamics. In this paper, we propose TEMPOR, a Temporal Electron MicroscoPy Object Reconstruction algorithm for liquid-phase EM that leverages an implicit neural representation (INR) and a dynamical variational auto-encoder (DVAE) to recover time series of molecular structures. We demonstrate its advantages in recovering different motion dynamics from two simulated datasets, 7bcq and Cas9. To our knowledge, our work is the first attempt to directly recover 3D structures of a temporally-varying particle from liquid-phase EM movies. It provides a promising new approach for studying molecules' 3D dynamics in structural biology.

ACE-HetEM for ab initio Heterogenous Cryo-EM 3D Reconstruction

Due to the extremely low signal-to-noise ratio (SNR) and unknown poses (projection angles and image translation) in cryo-EM experiments, reconstructing 3D structures from 2D images is very challenging. On top of these challenges, heterogeneous cryo-EM reconstruction also has an additional requirement: conformation classification. An emerging solution to this problem is called amortized inference, implemented using the autoencoder architecture or its variants. Instead of searching for the correct image-to-pose/conformation mapping for every image in the dataset as in non-amortized methods, amortized inference only needs to train an encoder that maps images to appropriate latent spaces representing poses or conformations. Unfortunately, standard amortized-inference-based methods with entangled latent spaces have difficulty learning the distribution of conformations and poses from cryo-EM images. In this paper, we propose an unsupervised deep learning architecture called "ACE-HetEM" based on amortized inference. To explicitly enforce the disentanglement of conformation classifications and pose estimations, we designed two alternating training tasks in our method: image-to-image task and pose-to-pose task. Results on simulated datasets show that ACE-HetEM has comparable accuracy in pose estimation and produces even better reconstruction resolution than non-amortized methods. Furthermore, we show that ACE-HetEM is also applicable to real experimental datasets.