Evaluating large language models (LLMs) in long-form, knowledge-grounded role-play dialogues remains challenging. This study compares LLM-generated and human-authored responses in multi-turn professional training simulations through human evaluation ($N=38$) and automated LLM-as-a-judge assessment. Human evaluation revealed significant degradation in LLM-generated response quality across turns, particularly in naturalness, context maintenance and overall quality, while human-authored responses progressively improved. In line with this finding, participants also indicated a consistent preference for human-authored dialogue. These human judgements were validated by our automated LLM-as-a-judge evaluation, where Gemini 2.0 Flash achieved strong alignment with human evaluators on both zero-shot pairwise preference and stochastic 6-shot construct ratings, confirming the widening quality gap between LLM and human responses over time. Our work contributes a multi-turn benchmark exposing LLM degradation in knowledge-grounded role-play dialogues and provides a validated hybrid evaluation framework to guide the reliable integration of LLMs in training simulations.

With the technological advancements of the last couple of decades, machine learning (ML) and artificial intelligence (AI) play an important part in automated decision-making pipelines [1-3]. Even though these tools are generally created by optimising with respect to their accuracy and performance, there are other important aspects that should be considered, such as their fairness, robustness, and privacy [4]. One of these aspects, fairness, becomes even more crucial when AI-based tools are used for decision-making tasks such as checking whether accepting a credit application is profitable and risk-free, if an applicant is worthy of a job position, or if a defendant has a higher risk of committing a crime again.

In the field of weather forecasting, traditional models often grapple with discretization errors and time-dependent source discrepancies, which limit their predictive performance. In this paper, we present WeatherODE, a novel one-stage, physics-driven ordinary differential equation (ODE) model designed to enhance weather forecasting accuracy. By leveraging wave equation theory and integrating a time-dependent source model, WeatherODE effectively addresses the challenges associated with time-discretization error and dynamic atmospheric processes. Moreover, we design a CNN-ViT-CNN sandwich structure, facilitating efficient learning dynamics tailored for distinct yet interrelated tasks with varying optimization biases in advection equation estimation. Through rigorous experiments, WeatherODE demonstrates superior performance in both global and regional weather forecasting tasks, outperforming recent state-of-the-art approaches by significant margins of over 40.0\% and 31.8\% in root mean square error (RMSE), respectively. The source code is available at \url{https://github.com/DAMO-DI-ML/WeatherODE}.

Neural Language Models of Code, or Neural Code Models (NCMs), are rapidly progressing from research prototypes to commercial developer tools. As such, understanding the capabilities and limitations of such models is becoming critical. However, the abilities of these models are typically measured using automated metrics that often only reveal a portion of their real-world performance. While, in general, the performance of NCMs appears promising, currently much is unknown about how such models arrive at decisions. To this end, this paper introduces $do_{code}$, a post-hoc interpretability methodology specific to NCMs that is capable of explaining model predictions. $do_{code}$ is based upon causal inference to enable programming language-oriented explanations. While the theoretical underpinnings of $do_{code}$ are extensible to exploring different model properties, we provide a concrete instantiation that aims to mitigate the impact of spurious correlations by grounding explanations of model behavior in properties of programming languages. To demonstrate the practical benefit of $do_{code}$, we illustrate the insights that our framework can provide by performing a case study on two popular deep learning architectures and nine NCMs. The results of this case study illustrate that our studied NCMs are sensitive to changes in code syntax and statistically learn to predict tokens related to blocks of code (e.g., brackets, parenthesis, semicolon) with less confounding bias as compared to other programming language constructs. These insights demonstrate the potential of $do_{code}$ as a useful model debugging mechanism that may aid in discovering biases and limitations in NCMs.

Recovering global rankings from pairwise comparisons is an important problem with many applications, ranging from time synchronization to sports team ranking. Pairwise comparisons corresponding to matches in a competition can naturally be construed as edges in a directed graph (digraph), whose nodes represent competitors with an unknown rank or skill strength. However, existing methods addressing the rank estimation problem have thus far not utilized powerful neural network architectures to optimize ranking objectives. Hence, we propose to augment an algorithm with neural network, in particular graph neural network (GNN) for its coherence to the problem at hand. In this paper, we introduce GNNRank, a modeling framework that is compatible with any GNN capable of learning digraph embeddings, and we devise trainable objectives to encode ranking upsets/violations. This framework includes a ranking score estimation approach, and adds a useful inductive bias by unfolding the Fiedler vector computation of the graph constructed from a learnable similarity matrix. Experimental results on a wide range of data sets show that our methods attain competitive and often superior performance compared with existing approaches. It also shows promising transfer ability to new data based on the trained GNN model.

Probabilistic models often use neural networks to control their predictive uncertainty. However, when making out-of-distribution (OOD)} predictions, the often-uncontrollable extrapolation properties of neural networks yield poor uncertainty predictions. Such models then don't know what they don't know, which directly limits their robustness w.r.t unexpected inputs. To counter this, we propose to explicitly train the uncertainty predictor where we are not given data to make it reliable. As one cannot train without data, we provide mechanisms for generating pseudo-inputs in informative low-density regions of the input space, and show how to leverage these in a practical Bayesian framework that casts a prior distribution over the model uncertainty. With a holistic evaluation, we demonstrate that this yields robust and interpretable predictions of uncertainty while retaining state-of-the-art performance on diverse tasks such as regression and generative modelling

Machine learning (ML) techniques are increasingly applied to decision-making and control problems in Cyber-Physical Systems among which many are safety-critical, e.g., chemical plants, robotics, autonomous vehicles. Despite the significant benefits brought by ML techniques, they also raise additional safety issues because 1) most expressive and powerful ML models are not transparent and behave as a black box and 2) the training data which plays a crucial role in ML safety is usually incomplete. An important technique to achieve safety for ML models is "Safe Fail", i.e., a model selects a reject option and applies the backup solution, a traditional controller or a human operator for example, when it has low confidence in a prediction. Data-driven models produced by ML algorithms learn from training data, and hence they are only as good as the examples they have learnt. As pointed in [17], ML models work well in the "training space" (i.e., feature space with sufficient training data), but they could not extrapolate beyond the training space. As observed in many previous studies, a feature space that lacks training data generally has a much higher error rate than the one that contains sufficient training samples [31]. Therefore, it is essential to identify the training space and avoid extrapolating beyond the training space. In this paper, we propose an efficient Feature Space Partitioning Tree (FSPT) to address this problem. Using experiments, we also show that, a strong relationship exists between model performance and FSPT score.