Reward models (RMs) are at the crux of successfully using RLHF to align pretrained models to human preferences, yet there has been relatively little study that focuses on evaluation of those models. Evaluating reward models presents an opportunity to understand the opaque technologies used for alignment of language models and which values are embedded in them. Resources for reward model training and understanding are sparse in the nascent open-source community around them. To enhance scientific understanding of reward models, we present RewardBench, a benchmark dataset and code-base for evaluation. The RewardBench dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety, to benchmark how reward models perform on challenging, structured and out-of-distribution queries. We create specific comparison datasets for RMs that have subtle, but verifiable reasons (e.g. bugs, incorrect facts) why one answer should be preferred to another. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods, such as the direct MLE training of classifiers and the implicit reward modeling of Direct Preference Optimization (DPO). We present many findings on propensity for refusals, reasoning limitations, and instruction following shortcomings of various reward models towards a better understanding of the RLHF process.

Imitation learning provides an efficient way to teach robots dexterous skills; however, learning complex skills robustly and generalizablely usually consumes large amounts of human demonstrations. To tackle this challenging problem, we present 3D Diffusion Policy (DP3), a novel visual imitation learning approach that incorporates the power of 3D visual representations into diffusion policies, a class of conditional action generative models. The core design of DP3 is the utilization of a compact 3D visual representation, extracted from sparse point clouds with an efficient point encoder. In our experiments involving 72 simulation tasks, DP3 successfully handles most tasks with just 10 demonstrations and surpasses baselines with a 24.2% relative improvement. In 4 real robot tasks, DP3 demonstrates precise control with a high success rate of 85%, given only 40 demonstrations of each task, and shows excellent generalization abilities in diverse aspects, including space, viewpoint, appearance, and instance. Interestingly, in real robot experiments, DP3 rarely violates safety requirements, in contrast to baseline methods which frequently do, necessitating human intervention. Our extensive evaluation highlights the critical importance of 3D representations in real-world robot learning. Videos, code, and data are available on https://3d-diffusion-policy.github.io .

In a practical setting, how to enable robust Federated Learning (FL) systems, both in terms of generalization and personalization abilities, is one important research question. It is a challenging issue due to the consequences of non-i.i.d. properties of client's data, often referred to as statistical heterogeneity, and small local data samples from the various data distributions. Therefore, to develop robust generalized global and personalized models, conventional FL methods need to redesign the knowledge aggregation from biased local models while considering huge divergence of learning parameters due to skewed client data. In this work, we demonstrate that the knowledge transfer mechanism achieves these objectives and develop a novel knowledge distillation-based approach to study the extent of knowledge transfer between the global model and local models. Henceforth, our method considers the suitability of transferring the outcome distribution and (or) the embedding vector of representation from trained models during cross-device knowledge transfer using a small proxy dataset in heterogeneous FL. In doing so, we alternatively perform cross-device knowledge transfer following general formulations as 1) global knowledge transfer and 2) on-device knowledge transfer. Through simulations on three federated datasets, we show the proposed method achieves significant speedups and high personalized performance of local models. Furthermore, the proposed approach offers a more stable algorithm than other baselines during the training, with minimal communication data load when exchanging the trained model's outcomes and representation.

Large language models (LLMs) can be used to generate natural language explanations (NLE) that are adapted to different users' situations. However, there is yet to be a quantitative evaluation of the extent of such adaptation. To bridge this gap, we collect a benchmarking dataset, Situation-Based Explanation. This dataset contains 100 explanandums. Each explanandum is paired with explanations targeted at three distinct audience types-such as educators, students, and professionals-enabling us to assess how well the explanations meet the specific informational needs and contexts of these diverse groups e.g. students, teachers, and parents. For each "explanandum paired with an audience" situation, we include a human-written explanation. These allow us to compute scores that quantify how the LLMs adapt the explanations to the situations. On an array of pretrained language models with varying sizes, we examine three categories of prompting methods: rule-based prompting, meta-prompting, and in-context learning prompting. We find that 1) language models can generate prompts that result in explanations more precisely aligned with the target situations, 2) explicitly modeling an "assistant" persona by prompting "You are a helpful assistant..." is not a necessary prompt technique for situated NLE tasks, and 3) the in-context learning prompts only can help LLMs learn the demonstration template but can't improve their inference performance. SBE and our analysis facilitate future research towards generating situated natural language explanations.

This paper explores the in-context learning capabilities of masked language models, challenging the common view that this ability does not 'emerge' in them. We present an embarrassingly simple inference technique that enables DeBERTa to operate as a generative model without any additional training. Our findings demonstrate that DeBERTa can match and even surpass GPT-3, its contemporary that famously introduced the paradigm of in-context learning. The comparative analysis reveals that the masked and causal language models behave very differently, as they clearly outperform each other on different categories of tasks. This suggests that there is great potential for a hybrid training approach that takes advantage of the strengths of both training objectives.

Catastrophic Forgetting (CF) means models forgetting previously acquired knowledge when learning new data. It compromises the effectiveness of large language models (LLMs) during fine-tuning, yet the underlying causes have not been thoroughly investigated. This paper takes the first step to reveal the direct link between the flatness of the model loss landscape and the extent of CF in the field of LLMs. Based on this, we introduce the sharpness-aware minimization to mitigate CF by flattening the loss landscape. Experiments on three widely-used fine-tuning datasets, spanning different model scales, demonstrate the effectiveness of our method in alleviating CF. Analyses show that we nicely complement the existing anti-forgetting strategies, further enhancing the resistance of LLMs to CF.

The integration of generative artificial intelligence (GenAI) tools into education has been a game-changer for teaching and assessment practices, bringing new opportunities, but also novel challenges which need to be dealt with. This paper presents the Comprehensive AI Assessment Framework (CAIAF), an evolved version of the AI Assessment Scale (AIAS) by Perkins, Furze, Roe, and MacVaugh, targeted toward the ethical integration of AI into educational assessments. This is where the CAIAF differs, as it incorporates stringent ethical guidelines, with clear distinctions based on educational levels, and advanced AI capabilities of real-time interactions and personalized assistance. The framework developed herein has a very intuitive use, mainly through the use of a color gradient that enhances the user-friendliness of the framework. Methodologically, the framework has been developed through the huge support of a thorough literature review and practical insight into the topic, becoming a dynamic tool to be used in different educational settings. The framework will ensure better learning outcomes, uphold academic integrity, and promote responsible use of AI, hence the need for this framework in modern educational practice.

Knowledge tracing (KT) is a crucial task in intelligent education, focusing on predicting students' performance on given questions to trace their evolving knowledge. The advancement of deep learning in this field has led to deep-learning knowledge tracing (DLKT) models that prioritize high predictive accuracy. However, many existing DLKT methods overlook the fundamental goal of tracking students' dynamical knowledge mastery. These models do not explicitly model knowledge mastery tracing processes or yield unreasonable results that educators find difficulty to comprehend and apply in real teaching scenarios. In response, our research conducts a preliminary analysis of mainstream KT approaches to highlight and explain such unreasonableness. We introduce GRKT, a graph-based reasonable knowledge tracing method to address these issues. By leveraging graph neural networks, our approach delves into the mutual influences of knowledge concepts, offering a more accurate representation of how the knowledge mastery evolves throughout the learning process. Additionally, we propose a fine-grained and psychological three-stage modeling process as knowledge retrieval, memory strengthening, and knowledge learning/forgetting, to conduct a more reasonable knowledge tracing process. Comprehensive experiments demonstrate that GRKT outperforms eleven baselines across three datasets, not only enhancing predictive accuracy but also generating more reasonable knowledge tracing results. This makes our model a promising advancement for practical implementation in educational settings. The source code is available at https://github.com/JJCui96/GRKT.

This survey presents an overview of methods for learning from video (LfV) in the context of reinforcement learning (RL) and robotics. We focus on methods capable of scaling to large internet video datasets and, in the process, extracting foundational knowledge about the world's dynamics and physical human behaviour. Such methods hold great promise for developing general-purpose robots. We open with an overview of fundamental concepts relevant to the LfV-for-robotics setting. This includes a discussion of the exciting benefits LfV methods can offer (e.g., improved generalization beyond the available robot data) and commentary on key LfV challenges (e.g., missing information in video and LfV distribution shifts). Our literature review begins with an analysis of video foundation model techniques that can extract knowledge from large, heterogeneous video datasets. Next, we review methods that specifically leverage video data for robot learning. Here, we categorise work according to which RL knowledge modality (KM) benefits from the use of video data. We additionally highlight techniques for mitigating LfV challenges, including reviewing action representations that address missing action labels in video. Finally, we examine LfV datasets and benchmarks, before concluding with a discussion of challenges and opportunities in LfV. Here, we advocate for scalable foundation model approaches that can leverage the full range of internet video data, and that target the learning of the most promising RL KMs: the policy and dynamics model. Overall, we hope this survey will serve as a comprehensive reference for the emerging field of LfV, catalysing further research in the area and facilitating progress towards the development of general-purpose robots.

How can we trust the correctness of a learned model on a particular input of interest? Model accuracy is typically measured *on average* over a distribution of inputs, giving no guarantee for any fixed input. This paper proposes a theoretically-founded solution to this problem: to train *Self-Proving models* that prove the correctness of their output to a verification algorithm $V$ via an Interactive Proof. Self-Proving models satisfy that, with high probability over a random input, the model generates a correct output *and* successfully proves its correctness to $V\!$. The *soundness* property of $V$ guarantees that, for *every* input, no model can convince $V$ of the correctness of an incorrect output. Thus, a Self-Proving model proves correctness of most of its outputs, while *all* incorrect outputs (of any model) are detected by $V$. We devise a generic method for learning Self-Proving models, and we prove convergence bounds under certain assumptions. The theoretical framework and results are complemented by experiments on an arithmetic capability: computing the greatest common divisor (GCD) of two integers. Our learning method is used to train a Self-Proving transformer that computes the GCD *and* proves the correctness of its answer.