We study an LLM fine-tuning task for designing reward functions for sequential resource allocation problems in public health, guided by human preferences expressed in natural language. This setting presents a challenging testbed for alignment due to complex and ambiguous objectives and limited data availability. We propose DPO-PRO, a robust fine-tuning algorithm based on Direct Preference Optimization (DPO), which accounts for uncertainty in the preference distribution using a lightweight Distributionally Robust Optimization (DRO) formulation. Unlike prior DRO-based variants, DPO-PRO focuses solely on uncertainty in preferences, avoiding unnecessary conservatism and incurring negligible computational overhead. We evaluate DPO-PRO on a real-world maternal mobile health program operated by the nonprofit organization ARMMAN, as well as on standard alignment benchmarks. Experimental results demonstrate that our method consistently improves robustness to noisy preference signals compared to existing DPO variants. Moreover, DPO-PRO achieves comparable performance to prior self-reflection-based baseline for reward function design, while requiring significantly lower inference-time cost.

Fairness in multi-agent reinforcement learning (MARL) is often framed as a workload balance problem, overlooking agent expertise and the structured coordination required in real-world domains. In healthcare, equitable task allocation requires workload balance or expertise alignment to prevent burnout and overuse of highly skilled agents. Workload balance refers to distributing an approximately equal number of subtasks or equalised effort across healthcare workers, regardless of their expertise. We make two contributions to address this problem. First, we propose FairSkillMARL, a framework that defines fairness as the dual objective of workload balance and skill-task alignment. Second, we introduce MARLHospital, a customizable healthcare-inspired environment for modeling team compositions and energy-constrained scheduling impacts on fairness, as no existing simulators are well-suited for this problem. We conducted experiments to compare FairSkillMARL in conjunction with four standard MARL methods, and against two state-of-the-art fairness metrics. Our results suggest that fairness based solely on equal workload might lead to task-skill mismatches and highlight the need for more robust metrics that capture skill-task misalignment. Our work provides tools and a foundation for studying fairness in heterogeneous multi-agent systems where aligning effort with expertise is critical.

Retrieval-Augmented Generation (RAG) is a well-established and rapidly evolving field within AI that enhances the outputs of large language models by integrating relevant information retrieved from external knowledge sources. While industry adoption of RAG is now beginning, there is a significant lack of research on its practical application in industrial contexts. To address this gap, we conducted a semistructured interview study with 13 industry practitioners to explore the current state of RAG adoption in real-world settings. Our study investigates how companies apply RAG in practice, providing (1) an overview of industry use cases, (2) a consolidated list of system requirements, (3) key challenges and lessons learned from practical experiences, and (4) an analysis of current industry evaluation methods. Our main findings show that current RAG applications are mostly limited to domain-specific QA tasks, with systems still in prototype stages; industry requirements focus primarily on data protection, security, and quality, while issues such as ethics, bias, and scalability receive less attention; data preprocessing remains a key challenge, and system evaluation is predominantly conducted by humans rather than automated methods.

Table reasoning, including tabular QA and fact verification, often depends on annotated data or complex data augmentation, limiting flexibility and generalization. LLMs, despite their versatility, often underperform compared to simple supervised models. To approach these issues, we introduce PanelTR, a framework utilizing LLM agent scientists for robust table reasoning through a structured scientific approach. PanelTR's workflow involves agent scientists conducting individual investigations, engaging in self-review, and participating in collaborative peer-review discussions. This process, driven by five scientist personas, enables semantic-level transfer without relying on data augmentation or parametric optimization. Experiments across four benchmarks show that PanelTR outperforms vanilla LLMs and rivals fully supervised models, all while remaining independent of training data. Our findings indicate that structured scientific methodology can effectively handle complex tasks beyond table reasoning with flexible semantic understanding in a zero-shot context.

Language-image pre-training (LIP) enables the development of vision-language models capable of zero-shot classification, localization, multimodal retrieval, and semantic understanding. Various explanation methods have been proposed to visualize the importance of input image-text pairs on the model's similarity outputs. However, popular saliency maps are limited by capturing only first-order attributions, overlooking the complex cross-modal interactions intrinsic to such encoders. We introduce faithful interaction explanations of LIP models (FIxLIP) as a unified approach to decomposing the similarity in vision-language encoders. FIxLIP is rooted in game theory, where we analyze how using the weighted Banzhaf interaction index offers greater flexibility and improves computational efficiency over the Shapley interaction quantification framework. From a practical perspective, we propose how to naturally extend explanation evaluation metrics, such as the pointing game and area between the insertion/deletion curves, to second-order interaction explanations. Experiments on the MS COCO and ImageNet-1k benchmarks validate that second-order methods, such as FIxLIP, outperform first-order attribution methods. Beyond delivering high-quality explanations, we demonstrate the utility of FIxLIP in comparing different models, e.g. CLIP vs. SigLIP-2.

University of Pittsburgh Pittsburgh, P A, USA User-controllable privacy is important in modern sensing systems, as privacy preferences can vary significantly from person to person and may evolve over time. This is especially relevant in devices equipped with Inertial Measurement Unit (IMU) sensors, such as smartphones and wearables, which continuously collect rich time-series data that can inadvertently expose sensitive user behaviors. While prior work has proposed privacy-preserving methods for sensor data, most rely on static, predefined privacy labels or require large quantities of private training data, limiting their adaptability and user agency. In this work, we introduce PrivCLIP, a dynamic, user-controllable, few-shot privacy-preserving sensing framework. PrivCLIP allows users to specify and modify their privacy preferences by categorizing activities as sensitive (blacklisted), non-sensitive (white-listed), or neutral (gray-listed). Leveraging a multimodal contrastive learning approach, Priv-CLIP aligns IMU sensor data with natural language activity descriptions in a shared embedding space, enabling few-shot detection of sensitive activities. When a privacy-sensitive activity is identified, the system uses a language-guided activity sanitizer and a motion generation module (IMU-GPT) to transform the original data into a privacy-compliant version that semantically resembles a non-sensitive activity. We evaluate PrivCLIP on multiple human activity recognition datasets and demonstrate that it significantly outperforms baseline methods in terms of both privacy protection and data utility. A growing number of smart devices, including wearables and smartphones, are equipped with sensors that enable applications in health monitoring, fitness tracking, and human activity recognition (HAR). Among these, inertial measurement units (IMUs) are particularly useful, as they capture fine-grained motion data that can be used to infer user behavior, physical condition, and mobility patterns. Typically, this sensor data is collected and transmitted to third-party cloud services for large-scale sensing and analytics. In many applications, online data transmission is desirable. Online tracking facilitates data sharing with peers, which enhances user engagement by providing timely feedback and positive reinforcement, which can be critical for sustained participation. However, outsourcing data processing to third-party providers raises significant privacy concerns.

Unsupervised video segmentation is a challenging computer vision task, especially due to the lack of supervisory signals coupled with the complexity of visual scenes. To overcome this challenge, state-of-the-art models based on slot attention often have to rely on large and computationally expensive neural architectures. To this end, we propose a simple knowledge distillation framework that effectively transfers object-centric representations to a lightweight student. The proposed framework, called SlotMatch, aligns corresponding teacher and student slots via the cosine similarity, requiring no additional distillation objectives or auxiliary supervision. The simplicity of SlotMatch is confirmed via theoretical and empirical evidence, both indicating that integrating additional losses is redundant. We conduct experiments on three datasets to compare the state-of-the-art teacher model, SlotContrast, with our distilled student. The results show that our student based on SlotMatch matches and even outperforms its teacher, while using 3.6x less parameters and running up to 2.7x faster. Moreover, our student surpasses all other state-of-the-art unsupervised video segmentation models.

As Audio Large Language Models (ALLMs) emerge as powerful tools for speech processing, their safety implications demand urgent attention. While considerable research has explored textual and vision safety, audio's distinct characteristics present significant challenges. This paper first investigates: Is ALLM vulnerable to backdoor attacks exploiting acoustic triggers? In response to this issue, we introduce Hidden in the Noise (HIN), a novel backdoor attack framework designed to exploit subtle, audio-specific features. HIN applies acoustic modifications to raw audio waveforms, such as alterations to temporal dynamics and strategic injection of spectrally tailored noise. These changes introduce consistent patterns that an ALLM's acoustic feature encoder captures, embedding robust triggers within the audio stream. To evaluate ALLM robustness against audio-feature-based triggers, we develop the AudioSafe benchmark, assessing nine distinct risk types. Extensive experiments on AudioSafe and three established safety datasets reveal critical vulnerabilities in existing ALLMs: (I) audio features like environment noise and speech rate variations achieve over 90% average attack success rate. (II) ALLMs exhibit significant sensitivity differences across acoustic features, particularly showing minimal response to volume as a trigger, and (III) poisoned sample inclusion causes only marginal loss curve fluctuations, highlighting the attack's stealth.

Supervised Fine-Tuning (SFT) plays a pivotal role in adapting Large Language Models (LLMs) to specialized domains such as medical reasoning. However, existing SFT practices often rely on unfiltered datasets that contain redundant and low-quality samples, leading to substantial computational costs and suboptimal performance. Although existing methods attempt to alleviate this problem by selecting data based on sample difficulty, defined by knowledge and reasoning complexity, they overlook each sample's optimization utility reflected in its gradient. Interestingly, we find that gradient-based influence alone favors easy-to-optimize samples that cause large parameter shifts but lack deep reasoning chains, while difficulty alone selects noisy or overly complex cases that fail to guide stable optimization. Based on this observation, we propose a data selection strategy, Difficulty-Influence Quadrant (DIQ), which prioritizes samples in the high-difficulty-high-influence quadrant to balance complex clinical reasoning with substantial gradient influence, enabling efficient medical reasoning with minimal fine-tuning data. Furthermore, Human and LLM-as-a-judge evaluations show that DIQ-selected subsets demonstrate higher data quality and generate clinical reasoning that is more aligned with expert practices in differential diagnosis, safety check, and evidence citation, as DIQ emphasizes samples that foster expert-like reasoning patterns. Extensive experiments on medical reasoning benchmarks demonstrate that DIQ enables models fine-tuned on only 1% of selected data to match full-dataset performance, while using 10% consistently outperforms baseline methods, highlighting the superiority of principled data selection over brute-force scaling. The code and data are available at https://github.com/mihara-bot/DIQ.

Multiple Instance Learning (MIL) is the leading approach for whole slide image (WSI) classification, enabling efficient analysis of gigapixel pathology slides. Recent work has introduced vision-language models (VLMs) into MIL pipelines to incorporate medical knowledge through text-based class descriptions rather than simple class names. However, when these methods rely on large language models (LLMs) to generate clinical descriptions or use fixed-length prompts to represent complex pathology concepts, the limited token capacity of VLMs often constrains the expressiveness and richness of the encoded class information. Additionally, descriptions generated solely by LLMs may lack domain grounding and fine-grained medical specificity, leading to suboptimal alignment with visual features. To address these challenges, we propose a vision-language MIL framework with two key contributions: (1) A grounded multi-agent description generation system that leverages curated pathology textbooks and agent specialization (e.g., morphology, spatial context) to produce accurate and diverse clinical descriptions; (2) A text encoding strategy using a list of descriptions rather than a single prompt, capturing fine-grained and complementary clinical signals for better alignment with visual features. Integrated into a VLM-MIL pipeline, our approach shows improved performance over single-prompt class baselines and achieves results comparable to state-of-the-art models, as demonstrated on renal and lung cancer datasets.