Nava Tintarev

Group Recommender Systems (GRS) face the challenge of combining conflicting individual preferences. While Large Language Models (LLMs) are increasingly integrated into GRS, they may introduce risks regarding output inconsistency or sensitivity to group characteristics. In this paper, we analyze LLMs acting as both decision-makers and explanation generators for group recommendations. We contribute a novel dataset featuring fictitious group preferences, top-10 recommendations, and natural language explanations. Explanations are generated across multiple LLMs, testing differences between model families ( GPT-OSS vs. Mistral ) and model sizes. Our methodology evaluates recommendations and explanations against traditional social choice-based aggregation strategies across domains (abstract, low-stakes, high-stakes) and group configurations (uniform, divergent, coalitional, or minority). Findings reveal that the choice of LLM backbone is the primary driver of variance in recommendation quality and recommendation procedures detailed in the explanations. We found a clear divergence across model families. GPT-OSS provided consistent rankings, while Mistral exhibited more sensitivity to domain and group configurations. While main effects for domain and group configuration were not found, significant higher-order interactions suggest performance effects persist in specific combinations of configuration, domain, and aggregation strategy. We further discuss LLM roles in mitigating cold-start issues and group moderation, alongside drawbacks like transparency and privacy.

In this special issue, we bring together contributions that frame recommender systems for good as a broader reorientation of recommender systems research toward societal impact, responsibility, inclusion, and beneficial intervention. The papers collected here help define the emerging RS4Good agenda, examine its risks and normative constraints, study high-stakes and vulnerable contexts, and demonstrate constructive applications in domains such as healthcare, education, tourism, and sustainability. Taken together, they show that recommender systems for good are defined not only by the domains in which they are deployed, but also by the values, design processes, and evaluation criteria that guide their development.

The use of Artificial Intelligence (AI) in high-risk, decision-making scenarios presents technical, safety, and normative challenges; problems that may only be ameliorated by human oversight. However, notions of human oversight lack a common foundational understanding: oversight architectures are not well defined, the roles involved remain unclear, and implementation steps are opaque. Hence, researchers and practitioners struggle to determine how to design, implement, and evaluate systems that enable effective human oversight. This paper advances a practical framework for effective human oversight of AI systems, based on a cross-disciplinary perspective that draws on insights from computer science, human-computer interaction, psychology, philosophy, and law. The core contributions are: (1) a foundational framework, with a working definition, architecture and processes for effective human oversight of AI systems; (2) an initial template for documenting oversight architectures and processes, applied to diverse domains; and (3) a synthesis of open research challenges that need to be considered in the emerging field of effective human oversight of AI systems.

Large vision-language models (VLMs) enable intuitive visual search using natural language queries. However, improving their performance often requires fine-tuning and scaling to larger model variants. In this work, we propose a mechanism inspired by traditional text-based search to improve retrieval performance at inference time: relevance feedback. While relevance feedback can serve as an alternative to fine-tuning, its model-agnostic design also enables use with fine-tuned VLMs. Specifically, we introduce and evaluate four feedback strategies for VLM-based retrieval. First, we revise classical pseudo-relevance feedback (PRF), which refines query embeddings based on top-ranked results. To address its limitations, we propose generative relevance feedback (GRF), which uses synthetic captions for query refinement. Furthermore, we introduce an attentive feedback summarizer (AFS), a custom transformer-based model that integrates multimodal fine-grained features from relevant items. Finally, we simulate explicit feedback using ground-truth captions as an upper-bound baseline. Experiments on Flickr30k and COCO with the VLM backbones show that GRF, AFS, and explicit feedback improve retrieval performance by 3-5% in MRR@5 for smaller VLMs, and 1-3% for larger ones, compared to retrieval with no feedback. Moreover, AFS, similarly to explicit feedback, mitigates query drift and is more robust than GRF in iterative, multi-turn retrieval settings. Our findings demonstrate that relevance feedback can consistently enhance retrieval across VLMs and open up opportunities for interactive and adaptive visual search.

In this paper, I describe lessons learned in nearly 20 years of evaluating explanation interfaces with people. In my work, I have observed that tailoring the explanation to the user (e.g., a domain expert or a layperson) and task (e.g., decision support or model improvement), or context (e.g., under time pressure) is necessary for meaningful assessment of explanation quality (e.g., correct decisions or better understanding). Learning from trends in empirical research methods (natural language processing, information retrieval, and machine learning), I further argue that this is an issue for both human-computer interaction and machine learning. Real-world factors influencing the performance of systems are at best implicitly encoded when inferring probabilities from observations, and at worst, no longer applicable to the settings for which they are used. In seeking a balanced approach between this reality and a pragmatic, data-driven and (by necessity) reductionist approach, I make constructive suggestions for evaluating the quality of explainable artificial intelligence.