AI practitioners increasingly use large language model (LLM) agents in compound AI systems to solve complex reasoning tasks, these agent executions often fail to meet human standards, leading to errors that compromise the system's overall performance. Addressing these failures through human intervention is challenging due to the agents' opaque reasoning processes, misalignment with human expectations, the complexity of agent dependencies, and the high cost of manual inspection. This paper thus introduces a human-centered evaluation framework for Verifying LLM Agent failures (VeriLA), which systematically assesses agent failures to reduce human effort and make these agent failures interpretable to humans. The framework first defines clear expectations of each agent by curating human-designed agent criteria. Then, it develops a human-aligned agent verifier module, trained with human gold standards, to assess each agent's execution output. This approach enables granular evaluation of each agent's performance by revealing failures from a human standard, offering clear guidelines for revision, and reducing human cognitive load. Our case study results show that VeriLA is both interpretable and efficient in helping practitioners interact more effectively with the system. By upholding accountability in human-agent collaboration, VeriLA paves the way for more trustworthy and human-aligned compound AI systems.

In multi-turn dialogues, large language models (LLM) face a critical challenge of ensuring coherence while adapting to user-specific information. This study introduces the persona knowledge gap, the discrepancy between a model's internal understanding and the knowledge required for coherent, personalized conversations. While prior research has recognized these gaps, computational methods for their identification and resolution remain underexplored. We propose Conversation Preference Elicitation and Recommendation (CPER), a novel framework that dynamically detects and resolves persona knowledge gaps using intrinsic uncertainty quantification and feedback-driven refinement. CPER consists of three key modules: a Contextual Understanding Module for preference extraction, a Dynamic Feedback Module for measuring uncertainty and refining persona alignment, and a Persona-Driven Response Generation module for adapting responses based on accumulated user context. We evaluate CPER on two real-world datasets: CCPE-M for preferential movie recommendations and ESConv for mental health support. Using A/B testing, human evaluators preferred CPER's responses 42% more often than baseline models in CCPE-M and 27% more often in ESConv. A qualitative human evaluation confirms that CPER's responses are preferred for maintaining contextual relevance and coherence, particularly in longer (12+ turn) conversations.

Image generative models, particularly diffusion-based models, have surged in popularity due to their remarkable ability to synthesize highly realistic images. However, since these models are data-driven, they inherit biases from the training datasets, frequently leading to disproportionate group representations that exacerbate societal inequities. Traditionally, efforts to debiase these models have relied on predefined sensitive attributes, classifiers trained on such attributes, or large language models to steer outputs toward fairness. However, these approaches face notable drawbacks: predefined attributes do not adequately capture complex and continuous variations among groups. To address these issues, we introduce the Debiasing Diffusion Model (DDM), which leverages an indicator to learn latent representations during training, promoting fairness through balanced representations without requiring predefined sensitive attributes. This approach not only demonstrates its effectiveness in scenarios previously addressed by conventional techniques but also enhances fairness without relying on predefined sensitive attributes as conditions. In this paper, we discuss the limitations of prior bias mitigation techniques in diffusion-based models, elaborate on the architecture of the DDM, and validate the effectiveness of our approach through experiments.

Privacy policies are widely used by digital services and often required for legal purposes. Many machine learning based classifiers have been developed to automate detection of different concepts in a given privacy policy, which can help facilitate other automated tasks such as producing a more reader-friendly summary and detecting legal compliance issues. Despite the successful applications of large language models (LLMs) to many NLP tasks in various domains, there is very little work studying the use of LLMs for automated privacy policy analysis, therefore, if and how LLMs can help automate privacy policy analysis remains under-explored. To fill this research gap, we conducted a comprehensive evaluation of LLM-based privacy policy concept classifiers, employing both prompt engineering and LoRA (low-rank adaptation) fine-tuning, on four state-of-the-art (SOTA) privacy policy corpora and taxonomies. Our experimental results demonstrated that combining prompt engineering and fine-tuning can make LLM-based classifiers outperform other SOTA methods, \emph{significantly} and \emph{consistently} across privacy policy corpora/taxonomies and concepts. Furthermore, we evaluated the explainability of the LLM-based classifiers using three metrics: completeness, logicality, and comprehensibility. For all three metrics, a score exceeding 91.1\% was observed in our evaluation, indicating that LLMs are not only useful to improve the classification performance, but also to enhance the explainability of detection results.

The analysis of LLM biases is an active research field. As the application of LLMs in decision-making activities is increasing, their study is critical to better understand their implications on decisional processes. The coherence and the structural preferences that these models are acquiring over several topics could challenge their applications in several fields [4]. The origin of these biases is complicated to study and could come from different steps in LLM training. For example, these biases could be acquired during the pre-training phase, supervised fine-tuning phase, or even during the final alignment phase. This article focuses on understanding the extent of the political biases of LLM through two experiments. The first experiment has the objective of showing the left leaning of multiple LLMs in the context of several virtual European elections, section 4.1. The second experiment shows that LLMs consider "stupidity" and "ignorance" as human characteristics that make voting for the right wing more probable, section 4.2. As different models could exhibit different leans, we tested four of the most used LLMs in both our experiments, Table 1.

Generative machine learning models for small molecule drug discovery have shown immense promise, but many molecules generated by this approach are too difficult to synthesize to be worth further investigation or further development. We present a novel approach by fine-tuning Meta's Llama3 large language models (LLMs) to create SynLlama, which generates full synthetic pathways made of commonly accessible Enamine building blocks and robust organic reaction templates. SynLlama explores a large synthesizable space using significantly less data compared to other state-of-the-art methods, and offers strong performance in bottom-up synthesis, synthesizable analog generation, and hit expansion, offering medicinal chemists a valuable tool for drug discovery developments. We find that SynLlama can effectively generalize to unseen yet purchasable building blocks, meaning that its reconstruction capabilities extend to a broader synthesizable chemical space than the training data.

Large multimodal models (LMMs) are increasingly integrated into autonomous driving systems for user interaction. However, their limitations in fine-grained spatial reasoning pose challenges for system interpretability and user trust. We introduce Logic-RAG, a novel Retrieval-Augmented Generation (RAG) framework that improves LMMs' spatial understanding in driving scenarios. Logic-RAG constructs a dynamic knowledge base (KB) about object-object relationships in first-order logic (FOL) using a perception module, a query-to-logic embedder, and a logical inference engine. We evaluated Logic-RAG on visual-spatial queries using both synthetic and real-world driving videos. When using popular LMMs (GPT-4V, Claude 3.5) as proxies for an autonomous driving system, these models achieved only 55% accuracy on synthetic driving scenes and under 75% on real-world driving scenes. Augmenting them with Logic-RAG increased their accuracies to over 80% and 90%, respectively. An ablation study showed that even without logical inference, the fact-based context constructed by Logic-RAG alone improved accuracy by 15%. Logic-RAG is extensible: it allows seamless replacement of individual components with improved versions and enables domain experts to compose new knowledge in both FOL and natural language. In sum, Logic-RAG addresses critical spatial reasoning deficiencies in LMMs for autonomous driving applications. Code and data are available at https://github.com/Imran2205/LogicRAG.

Encoder-free multimodal large language models(MLLMs) eliminate the need for a well-trained vision encoder by directly processing image tokens before the language model. While this approach reduces computational overhead and model complexity, it often requires large amounts of training data to effectively capture the visual knowledge typically encoded by vision models like CLIP. The absence of a vision encoder implies that the model is likely to rely on substantial data to learn the necessary visual-semantic alignments. In this work, we present BREEN, a data-efficient encoder-free multimodal architecture that mitigates this issue. BREEN leverages a learnable query and image experts to achieve comparable performance with significantly less training data. The learnable query, positioned between image and text tokens, is supervised by the output of a pretrained CLIP model to distill visual knowledge, bridging the gap between visual and textual modalities. Additionally, the image expert processes image tokens and learnable queries independently, improving efficiency and reducing interference with the LLM's textual capabilities. BREEN achieves comparable performance to prior encoder-free state-of-the-art models like Mono-InternVL, using only 13 million text-image pairs in training about one percent of the data required by existing methods. Our work highlights a promising direction for data-efficient encoder-free multimodal learning, offering an alternative to traditional encoder-based approaches.

Cross-dataset Human Activity Recognition (HAR) suffers from limited model generalization, hindering its practical deployment. To address this critical challenge, inspired by the success of DoReMi in Large Language Models (LLMs), we introduce a data mixture optimization strategy for pre-training HAR models, aiming to improve the recognition performance across heterogeneous datasets. However, directly applying DoReMi to the HAR field encounters new challenges due to the continuous, multi-channel and intrinsic heterogeneous characteristics of IMU sensor data. To overcome these limitations, we propose a novel framework HAR-DoReMi, which introduces a masked reconstruction task based on Mean Squared Error (MSE) loss. By raplacing the discrete language sequence prediction task, which relies on the Negative Log-Likelihood (NLL) loss, in the original DoReMi framework, the proposed framework is inherently more appropriate for handling the continuous and multi-channel characteristics of IMU data. In addition, HAR-DoReMi integrates the Mahony fusion algorithm into the self-supervised HAR pre-training, aiming to mitigate the heterogeneity of varying sensor orientation. This is achieved by estimating the sensor orientation within each dataset and facilitating alignment with a unified coordinate system, thereby improving the cross-dataset generalization ability of the HAR model. Experimental evaluation on multiple cross-dataset HAR transfer tasks demonstrates that HAR-DoReMi improves the accuracy by an average of 6.51%, compared to the current state-of-the-art method with only approximately 30% to 50% of the data usage. These results confirm the effectiveness of HAR-DoReMi in improving the generalization and data efficiency of pre-training HAR models, underscoring its significant potential to facilitate the practical deployment of HAR technology.

Incorporating personas into conversational AI models is crucial for achieving authentic and engaging interactions. However, the cultural diversity and adaptability of existing persona datasets is often overlooked, reducing their efficacy in building culturally aware AI systems. To address this issue, we propose a two-step pipeline for generating culture-specific personas and introduce KoPersona, a dataset comprising 200,000 personas designed to capture Korean cultural values, behaviors, and social nuances. A comprehensive evaluation through various metrics validates the quality of KoPersona and its relevance to Korean culture. This work not only contributes to persona-based research, but also establishes a scalable approach for creating culturally relevant personas adaptable to various languages and cultural contexts.