Large language models exhibit systematic vulnerabilities to adversarial attacks despite extensive safety alignment through supervised fine-tuning and reinforcement learning from human feedback. These vulnerabilities manifest as differential safety behavior across token positions, with safety modifications concentrating in early positions while later positions show minimal distributional changes from base models. We provide a mechanistic analysis of safety alignment training dynamics, revealing that gradient concentration during autoregressive training creates signal decay across token positions. This leads to incomplete distributional learning where safety training fails to fully transform model preferences in later response regions. We introduce base-favored tokens as computational indicators of incomplete safety learning. Analysis reveals that while early positions undergo substantial distributional changes, later positions retain concerning base model preferences in safety-critical contexts, indicating systematic incomplete learning due to insufficient training signals. We develop a targeted completion method that addresses these undertrained regions through adaptive penalties and hybrid teacher distillation. Experimental evaluation across Llama and Qwen model families demonstrates remarkable improvements in adversarial robustness, with dramatic reductions in attack success rates across multiple attack types while fully preserving general capabilities.

How can Large Language Models (LLMs) be aligned with human intentions and values? A typical solution is to gather human preference on model outputs and finetune the LLMs accordingly while ensuring that updates do not deviate too far from a reference model. Recent approaches, such as direct preference optimization (DPO), have eliminated the need for unstable and sluggish reinforcement learning optimization by introducing close-formed supervised losses. However, a significant limitation of the current approach is its design for a single reference model only, neglecting to leverage the collective power of numerous pretrained LLMs. To overcome this limitation, we introduce a novel closed-form formulation for direct preference optimization using multiple reference models. The resulting algorithm, Multi-Reference Preference Optimization (MRPO), leverages broader prior knowledge from diverse reference models, substantially enhancing preference learning capabilities compared to the single-reference DPO. Our experiments demonstrate that LLMs finetuned with MRPO generalize better in various preference data, regardless of data scarcity or abundance. Furthermore, MRPO effectively finetunes LLMs to exhibit superior performance in several downstream natural language processing tasks such as HH-RLHF, GSM8K and TruthfulQA.

Efficient molecular dynamics (MD) simulation is vital for understanding atomic-scale processes in materials science and biophysics. Traditional density functional theory (DFT) methods are computationally expensive, which limits the feasibility of long-term simulations. We propose a novel approach that formulates MD simulation as a time-series forecasting problem, enabling advanced forecasting models to predict atomic trajectories via displacements rather than absolute positions. We incorporate a physics-informed loss and inference mechanism based on DFT-parametrised pair-wise Morse potential functions that penalize unphysical atomic proximity to enforce physical plausibility. Our method consistently surpasses standard baselines in simulation accuracy across diverse materials. The results highlight the importance of incorporating physics knowledge to enhance the reliability and precision of atomic trajectory forecasting. Remarkably, it enables stable modeling of thousands of MD steps in minutes, offering a scalable alternative to costly DFT simulations.

Effective decision-making in partially observable environments demands robust memory management. Despite their success in supervised learning, current deep-learning memory models struggle in reinforcement learning environments that are partially observable and long-term. They fail to efficiently capture relevant past information, adapt flexibly to changing observations, and maintain stable updates over long episodes. We theoretically analyze the limitations of existing memory models within a unified framework and introduce the Stable Hadamard Memory, a novel memory model for reinforcement learning agents. Our model dynamically adjusts memory by erasing no longer needed experiences and reinforcing crucial ones computationally efficiently. To this end, we leverage the Hadamard product for calibrating and updating memory, specifically designed to enhance memory capacity while mitigating numerical and learning challenges. Our approach significantly outperforms state-of-the-art memory-based methods on challenging partially observable benchmarks, such as meta-reinforcement learning, long-horizon credit assignment, and POPGym, demonstrating superior performance in handling long-term and evolving contexts.

Artificial agents with the aid of large language models (LLMs) are effective in various real-world scenarios but struggle to cooperate in social dilemmas. When making decisions under the strain of selecting between long-term consequences and short-term benefits in commonly shared resources, LLM-based agents often exploit the environment, leading to early depletion. Inspired by the concept of consideration of future consequences (CFC), which is well-known in social psychology, we propose a framework to enable the ability to consider future consequences for LLM-based agents, which results in a new kind of agent that we term the CFC-Agent. We enable the CFC-Agent to act toward different levels of consideration for future consequences. Our first set of experiments, where LLM is directly asked to make decisions, shows that agents considering future consequences exhibit sustainable behaviour and achieve high common rewards for the population. Extensive experiments in complex environments showed that the CFC-Agent can manage a sequence of calls to LLM for reasoning and engaging in communication to cooperate with others to resolve the common dilemma better. Finally, our analysis showed that considering future consequences not only affects the final decision but also improves the conversations between LLM-based agents toward a better resolution of social dilemmas.

Enabling multiple autonomous machines to perform reliably requires the development of efficient cooperative control algorithms. This paper presents a survey of algorithms that have been developed for controlling and coordinating autonomous machines in complex environments. We especially focus on task allocation methods using computational intelligence (CI) and deep reinforcement learning (RL). The advantages and disadvantages of the surveyed methods are analysed thoroughly. We also propose and discuss in detail various future research directions that shed light on how to improve existing algorithms or create new methods to enhance the employability and performance of autonomous machines in real-world applications. The findings indicate that CI and deep RL methods provide viable approaches to addressing complex task allocation problems in dynamic and uncertain environments. The recent development of deep RL has greatly contributed to the literature on controlling and coordinating autonomous machines, and it has become a growing trend in this area. It is envisaged that this paper will provide researchers and engineers with a comprehensive overview of progress in machine learning research related to autonomous machines. It also highlights underexplored areas, identifies emerging methodologies, and suggests new avenues for exploration in future research within this domain.

The increasing demand for autonomous systems in complex and dynamic environments has driven significant research into intelligent path planning methodologies. For decades, graph-based search algorithms, linear programming techniques, and evolutionary computation methods have served as foundational approaches in this domain. Recently, deep reinforcement learning (DRL) has emerged as a powerful method for enabling autonomous agents to learn optimal navigation strategies through interaction with their environments. This survey provides a comprehensive overview of traditional approaches as well as the recent advancements in DRL applied to path planning tasks, focusing on autonomous vehicles, drones, and robotic platforms. Key algorithms across both conventional and learning-based paradigms are categorized, with their innovations and practical implementations highlighted. This is followed by a thorough discussion of their respective strengths and limitations in terms of computational efficiency, scalability, adaptability, and robustness. The survey concludes by identifying key open challenges and outlining promising avenues for future research. Special attention is given to hybrid approaches that integrate DRL with classical planning techniques to leverage the benefits of both learning-based adaptability and deterministic reliability, offering promising directions for robust and resilient autonomous navigation.

In this paper, we study the “dataset bias” problem from a statistical standpoint, and identify the main cause of the problem as the strong correlation between a class attribute u and a non-class attribute b in the input x, represented by p(u|b) differing significantly from p(u). Since p(u|b) appears as part of the sampling distributions in the standard maximum log-likelihood (MLL) objective, a model trained on a biased dataset via MLL inherently incorporates such correlation into its parameters, leading to poor generalization to unbiased test data. From this observation, we propose to mitigate dataset bias via either weighting the objective of each sample n by 1 / p(un|bn) or sampling that sample with a weight proportional to 1 / p(un|bn). While both methods are statistically equivalent, the former proves more stable and effective in practice. Additionally, we establish a connection between our debiasing approach and causal reasoning, reinforcing our method’s theoretical foundation. However, when the bias label is unavailable, computing p(u|b) exactly is difficult. To overcome this challenge, we propose to approximate 1 / p(u|b) using a biased classifier trained with “bias amplification” losses. Extensive experiments on various biased datasets demonstrate the superiority of our method over existing debiasing techniques in most settings, validating our theoretical analysis.

Abstract Deep learning has been applied to achieve significant progress in emotion recognition from multimedia data. Despite such substantial progress, existing approaches are hindered by insufficient training data, leading to weak generalisation under mismatched conditions. To address these challenges, we propose a learning strategy which jointly transfers emotional knowledge learnt from rich datasets to source-poor datasets. Our method is also able to learn cross-domain features, leading to improved recognition performance. To demonstrate the robustness of the proposed learning strategy, we conducted extensive experiments on several benchmark datasets including eNTERFACE, SAVEE, EMODB, and RAVDESS. Experimental results show that the proposed method surpassed existing transfer learning schemes by a significant margin.

Social reasoning necessitates the capacity of theory of mind (ToM), the ability to contextualise and attribute mental states to others without having access to their internal cognitive structure. Recent machine learning approaches to ToM have demonstrated that we can train the observer to read the past and present behaviours of other agents and infer their beliefs (including false beliefs about things that no longer exist), goals, intentions and future actions. The challenges arise when the behavioural space is complex, demanding skilful space navigation for rapidly changing contexts for an extended period. We tackle the challenges by equipping the observer with novel neural memory mechanisms to encode, and hierarchical attention to selectively retrieve information about others. The memories allow rapid, selective querying of distal related past behaviours of others to deliberatively reason about their current mental state, beliefs and future behaviours. This results in ToMMY, a theory of mind model that learns to reason while making little assumptions about the underlying mental processes. We also construct a new suite of experiments to demonstrate that memories facilitate the learning process and achieve better theory of mind performance, especially for high-demand false-belief tasks that require inferring through multiple steps of changes.

Understanding other agents is a key challenge in constructing artificial social agents. Current works focus on centralised training, wherein agents are allowed to know all the information about others and the environmental state during training. In contrast, this work studies decentralised training, wherein agents must learn the model of other agents in order to cooperate with them under partially-observable conditions, even during training, i.e. learning agents are myopic. The intrinsic motivation for artificial agents is modelled on the concept of human social motivation that entices humans to meet and understand each other, especially when experiencing a utility loss. Our intrinsic motivation encourages agents to stay near each other to obtain better observations and construct a model of others. They do so when their model of other agents is poor, or the overall task performance is bad during the learning phase. This simple but effective method facilitates the processes of modelling others, resulting in an improvement of the performance in cooperative tasks significantly. Our experiments demonstrate that the socially-motivated agent can model others better and promote cooperation across different tasks.

Abstract Deep learning has been widely adopted in automatic emotion recognition and has lead to significant progress in the field. However, due to insufficient training data, pre-trained models are limited in their generalisation ability, leading to poor performance on novel test sets. To mitigate this challenge, transfer learning performed by fine-tuning pr-etrained models on novel domains has been applied. However, the fine-tuned knowledge may overwrite and/or discard important knowledge learnt in pre-trained models. In this paper, we address this issue by proposing a PathNet-based meta-transfer learning method that is able to (i) transfer emotional knowledge learnt from one visual/audio emotion domain to another domain and (ii) transfer emotional knowledge learnt from multiple audio emotion domains to one another to improve overall emotion recognition accuracy. To show the robustness of our proposed method, extensive experiments on facial expression-based emotion recognition and speech emotion recognition are carried out on three bench-marking data sets: SAVEE, EMODB, and eNTERFACE. Experimental results show that our proposed method achieves superior performance compared with existing transfer learning methods.

We introduce a novel training procedure for policy gradient methods wherein episodic memory is used to optimize the hyperparameters of reinforcement learning algorithms on-the-fly. Unlike other hyperparameter searches, we formulate hyperparameter scheduling as a standard Markov Decision Process and use episodic memory to store the outcome of used hyperparameters and their training contexts. At any policy update step, the policy learner refers to the stored experiences, and adaptively reconfigures its learning algorithm with the new hyperparameters determined by the memory. This mechanism, dubbed as Episodic Policy Gradient Training (EPGT), enables an episodic learning process, and jointly learns the policy and the learning algorithm's hyperparameters within a single run. Experimental results on both continuous and discrete environments demonstrate the advantage of using the proposed method in boosting the performance of various policy gradient algorithms.

Data-free Knowledge Distillation (DFKD) has attracted attention recently thanks to its appealing capability of transferring knowledge from a teacher network to a student network without using training data. The main idea is to use a generator to synthesize data for training the student. As the generator gets updated, the distribution of synthetic data will change. Such distribution shift could be large if the generator and the student are trained adversarially, causing the student to forget the knowledge it acquired at the previous steps. To alleviate this problem, we propose a simple yet effective method called Momentum Adversarial Distillation (MAD) which maintains an exponential moving average (EMA) copy of the generator and uses synthetic samples from both the generator and the EMA generator to train the student. Since the EMA generator can be considered as an ensemble of the generator's old versions and often undergoes a smaller change in updates compared to the generator, training on its synthetic samples can help the student recall the past knowledge and prevent the student from adapting too quickly to the new updates of the generator. Our experiments on six benchmark datasets including big datasets like ImageNet and Places365 demonstrate the superior performance of MAD over competing methods for handling the large distribution shift problem. Our method also compares favorably to existing DFKD methods and even achieves state-of-the-art results in some cases.

We introduce a constrained optimization method for policy gradient reinforcement learning, which uses two trust regions to regulate each policy update. In addition to using the proximity of one single old policy as the first trust region as done by prior works, we propose forming a second trust region by constructing another virtual policy that represents a wide range of past policies. We then enforce the new policy to stay closer to the virtual policy, which is beneficial if the old policy performs poorly. We propose a mechanism to automatically build the virtual policy from a memory buffer of past policies, providing a new capability for dynamically selecting appropriate trust regions during the optimization process. Our proposed method, dubbed Memory-Constrained Policy Optimization (MCPO), is examined in diverse environments, including robotic locomotion control, navigation with sparse rewards and Atari games, consistently demonstrating competitive performance against recent on-policy constrained policy gradient methods.

Multimodal dimensional emotion recognition has drawn a great attention from the affective computing community and numerous schemes have been extensively investigated, making a significant progress in this area. However, several questions still remain unanswered for most of existing approaches including: (i) how to simultaneously learn compact yet representative features from multimodal data, (ii) how to effectively capture complementary features from multimodal streams, and (iii) how to perform all the tasks in an end-to-end manner. To address these challenges, in this paper, we propose a novel deep neural network architecture consisting of a two-stream auto-encoder and a long short term memory for effectively integrating visual and audio signal streams for emotion recognition. To validate the robustness of our proposed architecture, we carry out extensive experiments on the multimodal emotion in the wild dataset: RECOLA. Experimental results show that the proposed method achieves state-of-the-art recognition performance.

Guilt aversion induces experience of a utility loss in people if they believe they have disappointed others, and this promotes cooperative behaviour in human. In psychological game theory, guilt aversion necessitates modelling of agents that have theory about what other agents think, also known as Theory of Mind (ToM). We aim to build a new kind of affective reinforcement learning agents, called Theory of Mind Agents with Guilt Aversion (ToMAGA), which are equipped with an ability to think about the wellbeing of others instead of just self-interest. To validate the agent design, we use a general-sum game known as Stag Hunt as a test bed. As standard reinforcement learning agents could learn suboptimal policies in social dilemmas like Stag Hunt, we propose to use belief-based guilt aversion as a reward shaping mechanism. We show that our belief-based guilt averse agents can efficiently learn cooperative behaviours in Stag Hunt Games.