State-of-the-art cross-encoders can be fine-tuned to be highly effective in passage re-ranking. The typical fine-tuning process of cross-encoders as re-rankers requires large amounts of manually labelled data, a contrastive learning objective, and a set of heuristically sampled negatives. An alternative recent approach for fine-tuning instead involves teaching the model to mimic the rankings of a highly effective large language model using a distillation objective. These fine-tuning strategies can be applied either individually, or in sequence. In this work, we systematically investigate the effectiveness of point-wise cross-encoders when fine-tuned independently in a single stage, or sequentially in two stages. Our experiments show that the effectiveness of point-wise cross-encoders fine-tuned using contrastive learning is indeed on par with that of models fine-tuned with multi-stage approaches. Code is available for reproduction at https://github.com/fpezzuti/multistage-finetuning.

While XAI focuses on providing AI explanations to humans, can the reverse - humans explaining their judgments to AI - foster richer, synergistic human-AI systems? This paper explores various forms of human inputs to AI and examines how human explanations can guide machine learning models toward automated judgments and explanations that align more closely with human concepts.

The attribution of the author of an art piece is typically a laborious manual process, usually relying on subjective evaluations of expert figures. However, there are some situations in which quantitative features of the artwork can support these evaluations. The extraction of these features can sometimes be automated, for instance, with the use of Machine Learning (ML) techniques. An example of these features is represented by repeated, mechanically impressed patterns, called punches, present chiefly in 13th and 14th-century panel paintings from Tuscany. Previous research in art history showcased a strong connection between the shapes of punches and specific artists or workshops, suggesting the possibility of using these quantitative cues to support the attribution. In the present work, we first collect a dataset of large-scale images of these panel paintings. Then, using YOLOv10, a recent and popular object detection model, we train a ML pipeline to perform object detection on the punches contained in the images. Due to the large size of the images, the detection procedure is split across multiple frames by adopting a sliding-window approach with overlaps, after which the predictions are combined for the whole image using a custom non-maximal suppression routine. Our results indicate how art historians working in the field can reliably use our method for the identification and extraction of punches.

Given points from an arbitrary metric space and a sequence of point updates sent by an adversary, what is the minimum recourse per update (i.e., the minimum number of changes needed to the set of centers after an update), in order to maintain a constant-factor approximation to a $k$-clustering problem? This question has received attention in recent years under the name consistent clustering. Previous works by Lattanzi and Vassilvitskii [ICLM '17] and Fichtenberger, Lattanzi, Norouzi-Fard, and Svensson [SODA '21] studied $k$-clustering objectives, including the $k$-center and the $k$-median objectives, under only point insertions. In this paper we study the $k$-center objective in the fully dynamic setting, where the update is either a point insertion or a point deletion. Before our work, {\L}\k{a}cki, Haeupler, Grunau, Rozho\v{n}, and Jayaram [SODA '24] gave a deterministic fully dynamic constant-factor approximation algorithm for the $k$-center objective with worst-case recourse of $2$ per update. In this work, we prove that the $k$-center clustering problem admits optimal recourse bounds by developing a deterministic fully dynamic constant-factor approximation algorithm with worst-case recourse of $1$ per update. Moreover our algorithm performs simple choices based on light data structures, and thus is arguably more direct and faster than the previous one which uses a sophisticated combinatorial structure. Additionally, we develop a new deterministic decremental algorithm and a new deterministic incremental algorithm, both of which maintain a $6$-approximate $k$-center solution with worst-case recourse of $1$ per update. Our incremental algorithm improves over the $8$-approximation algorithm by Charikar, Chekuri, Feder, and Motwani [STOC '97]. Finally, we remark that since all three of our algorithms are deterministic, they work against an adaptive adversary.

Adapting models to a language that was only partially present in the pre-training data requires fine-tuning, which is expensive in terms of both data and computational resources. As an alternative to fine-tuning, we explore the potential of activation steering-based techniques to enhance model performance on Italian tasks. Through our experiments we show that Italian steering (i) can be successfully applied to different models, (ii) achieves performances comparable to, or even better than, fine-tuned models for Italian, and (iii) yields higher quality and consistency in Italian generations. We also discuss the utility of steering and fine-tuning in the contemporary LLM landscape where models are anyway getting high Italian performances even if not explicitly trained in this language.

The use of Natural Language Processing (NLP) for helping decision-makers with Climate Change action has recently been highlighted as a use case aligning with a broader drive towards NLP technologies for social good. In this context, Aspect-Based Summarization (ABS) systems that extract and summarize relevant information are particularly useful as they provide stakeholders with a convenient way of finding relevant information in expert-curated reports. In this work, we release a new dataset for ABS of Climate Change reports and we employ different Large Language Models (LLMs) and so-called Small Language Models (SLMs) to tackle this problem in an unsupervised way. Considering the problem at hand, we also show how SLMs are not significantly worse for the problem while leading to reduced carbon footprint; we do so by applying for the first time an existing framework considering both energy efficiency and task performance to the evaluation of zero-shot generative models for ABS. Overall, our results show that modern language models, both big and small, can effectively tackle ABS for Climate Change reports but more research is needed when we frame the problem as a Retrieval Augmented Generation (RAG) problem and our work and dataset will help foster efforts in this direction.

Modular soft robot arms (MSRAs) are composed of multiple independent modules connected in a sequence. Due to their modular structure and high degrees of freedom (DOFs), these modules can simultaneously bend at different angles in various directions, enabling complex deformation. This capability allows MSRAs to perform more intricate tasks than single module robots. However, the modular structure also induces challenges in accurate planning, modeling, and control. Nonlinearity, hysteresis, and gravity complicate the physical model, while the modular structure and increased DOFs further lead to accumulative errors along the sequence. To address these challenges, we propose a flexible task space planning and control strategy for MSRAs, named S2C2A (State to Configuration to Action). Our approach formulates an optimization problem, S2C (State to Configuration planning), which integrates various loss functions and a forward MSRA model to generate configuration trajectories based on target MSRA states. Given the model complexity, we leverage a biLSTM network as the forward model. Subsequently, a configuration controller C2A (Configuration to Action control) is implemented to follow the planned configuration trajectories, leveraging only inaccurate internal sensing feedback. Both a biLSTM network and a physical model are utilized for configuration control. We validated our strategy using a cable-driven MSRA, demonstrating its ability to perform diverse offline tasks such as position control, orientation control, and obstacle avoidance. Furthermore, our strategy endows MSRA with online interaction capability with targets and obstacles. Future work will focus on addressing MSRA challenges, such as developing more accurate physical models and reducing configuration estimation errors along the module sequence.

Dolphins seem to make open-mouthed facial expressions most often while they are visible to a playmate, suggesting such displays may be similar to human smiles. While we often perceive these as a smile, there has been little research on facial communication in dolphins. We're finally realising that many species are To find out more, Elisabetta Palagi at the University of Pisa, Italy, and her colleagues analysed the behaviour of 22 captive bottlenose dolphins (Tursiops truncatus) at two wildlife parks: Zoomarine Rome in Italy and Planète Sauvage in Port-Saint-Père, France. In 80 hours of footage, the team observed a total of 1288 open-mouth expressions during social play sessions. More than 90 per cent of these events occurred during play between dolphins, with the rest happening during interactions between the dolphins and people.

Quantum Machine Learning (QML) is considered one of the most promising applications of Quantum Computing in the Noisy Intermediate Scale Quantum (NISQ) era for the impact it is thought to have in the near future. Although promising theoretical assumptions, the exploration of how QML could foster new discoveries in Medicine and Biology fields is still in its infancy with few examples. In this study, we aimed to assess whether Quantum Kernels (QK) could effectively classify subtypes of Breast Cancer (BC) patients on the basis of molecular characteristics. We performed an heuristic exploration of encoding configurations with different entanglement levels to determine a trade-off between kernel expressivity and performances. Our results show that QKs yield comparable clustering results with classical methods while using fewer data points, and are able to fit the data with a higher number of clusters. Additionally, we conducted the experiments on the Quantum Processing Unit (QPU) to evaluate the effect of noise on the outcome. We found that less expressive encodings showed a higher resilience to noise, indicating that the computational pipeline can be reliably implemented on the NISQ devices. Our findings suggest that QK methods show promises for application in Precision Oncology, especially in scenarios where the dataset is limited in size and a granular non-trivial stratification of complex molecular data cannot be achieved classically.

Vision-Language Models (VLMs) demonstrate remarkable zero-shot generalization to unseen tasks, but fall short of the performance of supervised methods in generalizing to downstream tasks with limited data. Prompt learning is emerging as a parameter-efficient method for adapting VLMs, but state-of-the-art approaches require annotated samples. In this paper we propose a novel approach to prompt learning based on unsupervised knowledge distillation from more powerful models. Our approach, which we call Knowledge Distillation Prompt Learning (KDPL), can be integrated into existing prompt learning techniques and eliminates the need for labeled examples during adaptation. Our experiments on more than ten standard benchmark datasets demonstrate that KDPL is very effective at improving generalization of learned prompts for zero-shot domain generalization, zero-shot cross-dataset generalization, and zero-shot base-to-novel class generalization problems. KDPL requires no ground-truth labels for adaptation, and moreover we show that even in the absence of any knowledge of training class names it can be used to effectively transfer knowledge. The code is publicly available at https://github.com/miccunifi/KDPL.