AI alignment aims to make AI systems behave in line with human intentions and values. As AI systems grow more capable, so do risks from misalignment. To provide a comprehensive and up-to-date overview of the alignment field, in this survey, we delve into the core concepts, methodology, and practice of alignment. First, we identify four principles as the key objectives of AI alignment: Robustness, Interpretability, Controllability, and Ethicality (RICE). Guided by these four principles, we outline the landscape of current alignment research and decompose them into two key components: forward alignment and backward alignment. The former aims to make AI systems aligned via alignment training, while the latter aims to gain evidence about the systems' alignment and govern them appropriately to avoid exacerbating misalignment risks. On forward alignment, we discuss techniques for learning from feedback and learning under distribution shift. On backward alignment, we discuss assurance techniques and governance practices. We also release and continually update the website (www.alignmentsurvey.com) which features tutorials, collections of papers, blog posts, and other resources.

When reasoning about actions, e.g., by means of task planning or agent programming with Golog, the robot's actions are typically modeled on an abstract level, where complex actions such as picking up an object are treated as atomic primitives with deterministic effects and preconditions that only depend on the current state. However, when executing such an action on a robot it can no longer be seen as a primitive. Instead, action execution is a complex task involving multiple steps with additional temporal preconditions and timing constraints. Furthermore, the action may be noisy, e.g., producing erroneous sensing results and not always having the desired effects. While these aspects are typically ignored in reasoning tasks, they need to be dealt with during execution. In this thesis, we propose several approaches towards closing this gap.

In the intricate architecture of the mammalian central nervous system, neurons form populations. Axonal bundles communicate between these clusters using spike trains. However, these neuron populations' precise encoding and operations have yet to be discovered. In our analysis, the starting point is a state-of-the-art mechanistic model of a generic neuron endowed with plasticity. From this simple framework emerges a subtle mathematical construct: The representation and manipulation of information can be precisely characterized by an algebra of convex cones. Furthermore, these neuron populations are not merely passive transmitters. They act as operators within this algebraic structure, mirroring the functionality of a low-level programming language. When these populations interconnect, they embody succinct yet potent algebraic expressions. These networks allow them to implement many operations, such as specialization, generalization, novelty detection, dimensionality reduction, inverse modeling, prediction, and associative memory. In broader terms, this work illuminates the potential of matrix embeddings in advancing our understanding in fields like cognitive science and AI. These embeddings enhance the capacity for concept processing and hierarchical description over their vector counterparts.

Model compression plays a vital role in the practical deployment of deep neural networks (DNNs), and evolutionary multi-objective (EMO) pruning is an essential tool in balancing the compression rate and performance of the DNNs. However, due to its population-based nature, EMO pruning suffers from the complex optimization space and the resource-intensive structure verification process, especially in complex networks. To this end, a multi-objective complex network pruning framework based on divide-and-conquer and global performance impairment ranking (EMO-DIR) is proposed in this paper. Firstly, a divide-and-conquer EMO network pruning method is proposed, which decomposes the complex task of EMO pruning on the entire network into easier sub-tasks on multiple sub-networks. On the one hand, this decomposition narrows the pruning optimization space and decreases the optimization difficulty; on the other hand, the smaller network structure converges faster, so the proposed algorithm consumes lower computational resources. Secondly, a sub-network training method based on cross-network constraints is designed, which could bridge independent EMO pruning sub-tasks, allowing them to collaborate better and improving the overall performance of the pruned network. Finally, a multiple sub-networks joint pruning method based on EMO is proposed. This method combines the Pareto Fronts from EMO pruning results on multiple sub-networks through global performance impairment ranking to design a joint pruning scheme. The rich experiments on CIFAR-10/100 and ImageNet-100/1k are conducted. The proposed algorithm achieves a comparable performance with the state-of-the-art pruning methods.

Recent work on object-centric world models aim to factorize representations in terms of objects in a completely unsupervised or self-supervised manner. Such world models are hypothesized to be a key component to address the generalization problem. While self-supervision has shown improved performance however, OOD generalization has not been systematically and explicitly tested. In this paper, we conduct an extensive study on the generalization properties of contrastive world model. We systematically test the model under a number of different OOD generalization scenarios such as extrapolation to new object attributes, introducing new conjunctions or new attributes. Our experiments show that the contrastive world model fails to generalize under the different OOD tests and the drop in performance depends on the extent to which the samples are OOD. When visualizing the transition updates and convolutional feature maps, we observe that any changes in object attributes (such as previously unseen colors, shapes, or conjunctions of color and shape) breaks down the factorization of object representations. Overall, our work highlights the importance of object-centric representations for generalization and current models are limited in their capacity to learn such representations required for human-level generalization.

Large-scale pre-trained language models (PLMs) bring new opportunities to challenging problems, especially those that need high-level intelligence, such as the math word problem (MWPs). However, directly applying existing PLMs to MWPs can fail as the generation process lacks sufficient supervision and thus lacks fast adaptivity as humans. We notice that human reasoning has a dual reasoning framework that consists of an immediate reaction system (system 1) and a delicate reasoning system (system 2), where the entire reasoning is determined by their interaction. This inspires us to develop a cooperative reasoning-induced PLM for solving MWPs, called Cooperative Reasoning (CoRe), resulting in a human-like reasoning architecture with system 1 as the generator and system 2 as the verifier. In our approach, the generator is responsible for generating reasoning paths, and the verifiers are used to supervise the evaluation in order to obtain reliable feedback for the generator. We evaluate our CoRe framework on several mathematical reasoning datasets and achieve decent improvement over state-of-the-art methods, up to 9.6% increase over best baselines. Our codes are available at https://github.com/TianHongZXY/CoRe

The enduring challenge in the field of artificial intelligence has been the control of systems to achieve desired behaviours. While for systems governed by straightforward dynamics equations, methods like Linear Quadratic Regulation (LQR) have historically proven highly effective, most real-world tasks, which require a general problem-solver, demand world models with dynamics that cannot be easily described by simple equations. Consequently, these models must be learned from data using neural networks. Most model predictive control (MPC) algorithms designed for visual world models have traditionally explored gradient-free population-based optimization methods, such as Cross Entropy and Model Predictive Path Integral (MPPI) for planning. However, we present an exploration of a gradient-based alternative that fully leverages the differentiability of the world model. In our study, we conduct a comparative analysis between our method and other MPC-based alternatives, as well as policy-based algorithms. In a sample-efficient setting, our method achieves on par or superior performance compared to the alternative approaches in most tasks. Additionally, we introduce a hybrid model that combines policy networks and gradient-based MPC, which outperforms pure policy based methods thereby holding promise for Gradient-based planning with world models in complex real-world tasks.

In Weighted Model Counting (WMC), we assign weights to literals and compute the sum of the weights of the models of a given propositional formula where the weight of an assignment is the product of the weights of its literals. The current WMC solvers work on Conjunctive Normal Form (CNF) formulas. However, CNF is not a natural representation for human-being in many applications. Motivated by the stronger expressive power of pseudo-Boolean (PB) formulas than CNF, we propose to perform WMC on PB formulas. Based on a recent dynamic programming algorithm framework called ADDMC for WMC, we implement a weighted PB counting tool PBCounter. We compare PBCounter with the state-of-the-art weighted model counters SharpSAT-TD, ExactMC, D4, and ADDMC, where the latter tools work on CNF with encoding methods that convert PB constraints into a CNF formula. The experiments on three domains of benchmarks show that PBCounter is superior to the model counters on CNF formulas.

Emotion Recognition in Conversation (ERC) has attracted growing attention in recent years as a result of the advancement and implementation of human-computer interface technologies. In this paper, we propose a novel approach to dependency modeling driven by Emotional Inertia and Contagion (EmotionIC) for ERC task. Our EmotionIC consists of three main components, i.e., Identity Masked Multi-Head Attention (IMMHA), Dialogue-based Gated Recurrent Unit (DiaGRU), and Skip-chain Conditional Random Field (SkipCRF). Compared to previous ERC models, EmotionIC can model a conversation more thoroughly at both the feature-extraction and classification levels. The proposed model attempts to integrate the advantages of attention- and recurrence-based methods at the feature-extraction level. Specifically, IMMHA is applied to capture identity-based global contextual dependencies, while DiaGRU is utilized to extract speaker- and temporal-aware local contextual information. At the classification level, SkipCRF can explicitly mine complex emotional flows from higher-order neighboring utterances in the conversation. Experimental results show that our method can significantly outperform the state-of-the-art models on four benchmark datasets. The ablation studies confirm that our modules can effectively model emotional inertia and contagion.

Since the key to artificial general intelligence (AGI) is commonly believed to be commonsense reasoning (CSR) or, roughly equivalently, discovery of a knowledge representation method (KRM) that is particularly suitable for CSR, the author developed a custom KRM for CSR. This novel KRM called Tumbug was designed to be pictorial in nature because there exists increasing evidence that the human brain uses some pictorial type of KRM, and no well-known prior research in AGI has researched this KRM possibility. Tumbug is somewhat similar to Roger Schank's Conceptual Dependency (CD) theory, but Tumbug is pictorial and uses about 30 components based on fundamental concepts from the sciences and human life, in contrast to CD theory, which is textual and uses about 17 components (= 6 Primitive Conceptual Categories + 11 Primitive Acts) based mainly on human-oriented activities. All the Building Blocks of Tumbug were found to generalize to only five Basic Building Blocks that exactly correspond to the three components {O, A, V} of traditional Object-Attribute-Value representation plus two new components {C, S}, which are Change and System. Collectively this set of five components, called "SCOVA," seems to be a universal foundation for all knowledge representation.