In this work, we study the problem of data retrieval for few-shot imitation learning: selecting data from a large dataset to train a performant policy for a specific task, given only a few target demonstrations. Prior methods retrieve data using a single-feature distance heuristic, assuming that the best demonstrations are those that most closely resemble the target examples in visual, semantic, or motion space. However, this approach captures only a subset of the relevant information and can introduce detrimental demonstrations, e.g., retrieving data from unrelated tasks due to similar scene layouts, or selecting similar motions from tasks with divergent goals. We present COLLAGE, a method for COLLective data AGgrEgation in few-shot imitation learning that uses an adaptive late fusion mechanism to guide the selection of relevant demonstrations based on a task-specific combination of multiple cues. COLLAGE follows a simple, flexible, and efficient recipe: it assigns weights to subsets of the dataset that are pre-selected using a single feature (e.g., appearance, shape, or language similarity), based on how well a policy trained on each subset predicts actions in the target demonstrations. These weights are then used to perform importance sampling during policy training, sampling data more densely or sparsely according to estimated relevance. COLLAGE is general and feature-agnostic, allowing it to combine any number of subsets selected by any retrieval heuristic, and to identify which subsets provide the greatest benefit for the target task. In extensive experiments, COLLAGE outperforms state-of-the-art retrieval and multi-task learning approaches by 5.1% in simulation across 10 tasks, and by 16.6% in the real world across 6 tasks, where we perform retrieval from the large-scale DROID dataset. More information at https://robin-lab.cs.utexas.edu/COLLAGE .

Multimodal Large Language Models (MLLMs) already achieve state-of-the-art results across a wide range of tasks and modalities. T o push their reasoning ability further, recent studies explore advanced prompting schemes and post-training fine-tuning. Although these techniques improve logical accuracy, they frequently leave the models' outputs burdened with pronounced social biases. Clarifying how reasoning gains interact with bias mitigation--and whether the two objectives inherently trade off--therefore remains an open and pressing research problem. Our study begins by benchmarking three bias-mitigation strategies--supervised fine-tuning (SFT), knowledge distillation (KD), and rule-based reinforcement learning (RL)--under identical conditions, establishing their baseline strengths and weaknesses. Building on these results, we vary the proportion of debias-focused and reasoning-centric samples within each paradigm to chart the reasoning-versus-bias trade-off. Our sweeps reveal a consistent sweet spot: a roughly 1:4 mix trained with reinforcement learning cuts stereotype scores by 10% while retaining 88% of the model's original reasoning accuracy, offering concrete guidance for balancing fairness and capability in MLLMs.

Code-mixing, the practice of switching between languages within a conversation, poses unique challenges for traditional NLP. Existing benchmarks are limited by their narrow language pairs and tasks, failing to adequately assess large language models' (LLMs) code-mixing abilities. Despite the recognized importance of code-mixing for multilingual users, research on LLMs in this context remains sparse. Additionally, current techniques for synthesizing code-mixed data are underdeveloped to generate code-mixing. In response, we introduce CodeMixBench, a comprehensive benchmark covering eight tasks, including three specific to LLMs and five traditional NLP tasks, and 18 languages across seven language families. We also propose a new method for generating large-scale synthetic code-mixed texts by combining word substitution with GPT-4 prompting. Our evaluation reveals consistent underperformance of LLMs on code-mixed datasets involving different language families. Enhancements in training data size, model scale, and few-shot learning could improve their performance. The code and dataset are available at https://github.com/Jeromeyluck/CodeMixBench.

Large Language Models (LLMs) have shown strong potential for tabular data generation by modeling textualized feature-value pairs. However, tabular data inherently exhibits sparse feature-level dependencies, where many feature interactions are structurally insignificant. This creates a fundamental mismatch as LLMs' self-attention mechanism inevitably distributes focus across all pairs, diluting attention on critical relationships, particularly in datasets with complex dependencies or semantically ambiguous features. To address this limitation, we propose GraDe (Graph-Guided Dependency Learning), a novel method that explicitly integrates sparse dependency graphs into LLMs' attention mechanism. GraDe employs a lightweight dynamic graph learning module guided by externally extracted functional dependencies, prioritizing key feature interactions while suppressing irrelevant ones. Our experiments across diverse real-world datasets demonstrate that GraDe outperforms existing LLM-based approaches by up to 12% on complex datasets while achieving competitive results with state-of-the-art approaches in synthetic data quality. Our method is minimally intrusive yet effective, offering a practical solution for structure-aware tabular data modeling with LLMs.

Behavior Cloning (BC) on curated (or filtered) data is the predominant paradigm for supervised fine-tuning (SFT) of large language models; as well as for imitation learning of control policies. Here, we draw on a connection between this successful strategy and the theory and practice of finding optimal policies via Reinforcement Learning (RL). Building on existing literature, we clarify that SFT can be understood as maximizing a lower bound on the RL objective in a sparse reward setting. Giving support to its often observed good performance. From this viewpoint, we realize that a small modification to SFT leads to an importance weighted variant that behaves closer to training with RL as it: i) optimizes a tighter bound to the RL objective and, ii) can improve performance compared to SFT on curated data. We refer to this variant as importance weighted supervised fine-tuning (iw-SFT). We show that it is easy to implement and can be further generalized to training with quality scored data. The resulting SFT variants are competitive with more advanced RL algorithms for large language models and for training policies in continuous control tasks. For example achieving 66.7% on the AIME 2024 dataset.

Counterfactual explanations (CFs) offer human-centric insights into machine learning predictions by highlighting minimal changes required to alter an outcome. Therefore, CFs can be used as (i) interventions for abnormality prevention and (ii) augmented data for training robust models. In this work, we explore large language models (LLMs), specifically GPT-4o-mini, for generating CFs in a zero-shot and three-shot setting. We evaluate our approach on two datasets: the AI-Readi flagship dataset for stress prediction and a public dataset for heart disease detection. Compared to traditional methods such as DiCE, CFNOW, and NICE, our few-shot LLM-based approach achieves high plausibility (up to 99%), strong validity (up to 0.99), and competitive sparsity. Moreover, using LLM-generated CFs as augmented samples improves downstream classifier performance (an average accuracy gain of 5%), especially in low-data regimes. This demonstrates the potential of prompt-based generative techniques to enhance explainability and robustness in clinical and physiological prediction tasks. Code base: github.com/shovito66/SenseCF.

Identifying associations between imaging phenotypes, disease risk factors, and clinical outcomes is essential for understanding disease mechanisms. However, traditional approaches rely on human-driven hypothesis testing and selection of association factors, often overlooking complex, non-linear dependencies among imaging phenotypes and other multi-modal data. To address this, we introduce Multi-agent Exploratory Synergy for the Heart (MESHAgents): a framework that leverages large language models as agents to dynamically elicit, surface, and decide confounders and phenotypes in association studies. Specifically, we orchestrate a multi-disciplinary team of AI agents, which spontaneously generate and converge on insights through iterative, self-organizing reasoning. The framework dynamically synthesizes statistical correlations with multi-expert consensus, providing an automated pipeline for phenome-wide association studies (PheWAS). We demonstrate the system's capabilities through a population-based study of imaging phenotypes of the heart and aorta. MESHAgents autonomously uncovered correlations between imaging phenotypes and a wide range of non-imaging factors, identifying additional confounder variables beyond standard demographic factors. Validation on diagnosis tasks reveals that MESHAgents-discovered phenotypes achieve performance comparable to expert-selected phenotypes, with mean AUC differences as small as $-0.004_{\pm0.010}$ on disease classification tasks. Notably, the recall score improves for 6 out of 9 disease types. Our framework provides clinically relevant imaging phenotypes with transparent reasoning, offering a scalable alternative to expert-driven methods.

As Large Language Models (LLMs) continue to advance, Retrieval-Augmented Generation (RAG) has emerged as a vital technique to enhance factual accuracy by integrating external knowledge into the generation process. However, LLMs often fail to faithfully integrate retrieved evidence into their generated responses, leading to factual inconsistencies. To quantify this gap, we introduce Entity-Context Divergence (ECD), a metric that measures the extent to which retrieved information is accurately reflected in model outputs. We systematically evaluate contemporary LLMs on their ability to preserve factual consistency in retrieval-augmented settings, a capability we define as RAG-ability. Our empirical analysis reveals that RAG-ability remains low across most LLMs, highlighting significant challenges in entity retention and context fidelity. This paper introduces Radiant (Retrieval AugmenteD entIty-context AligNmenT), a novel framework that merges RAG with alignment designed to optimize the interplay between retrieved evidence and generated content. Radiant extends Direct Preference Optimization (DPO) to teach LLMs how to integrate provided additional information into subsequent generations. As a behavior correction mechanism, Radiant boosts RAG performance across varied retrieval scenarios, such as noisy web contexts, knowledge conflicts, and hallucination reduction. This enables more reliable, contextually grounded, and factually coherent content generation.

The explosive growth of Large Language Models (LLMs), such as GPT-4 with 1.8 trillion parameters, demands a fundamental rethinking of data center architecture to ensure scalability, efficiency, and cost-effectiveness. Our work provides a comprehensive co-design framework that jointly explores FLOPS, HBM bandwidth and capacity, multiple network topologies (two-tier vs. FullFlat optical), the size of the scale-out domain, and popular parallelism/optimization strategies used in LLMs. We introduce and evaluate FullFlat network architectures, which provide uniform high-bandwidth, low-latency connectivity between all nodes, and demonstrate their transformative impact on performance and scalability. Through detailed sensitivity analyses, we quantify the benefits of overlapping compute and communication, leveraging hardware-accelerated collectives, widening the scale-out domain, and increasing memory capacity. Our study spans both sparse (mixture of experts) and dense transformer-based LLMs, revealing how system design choices affect Model FLOPS Utilization (MFU = Model FLOPS per token * Observed tokens per second / Peak FLOPS of the hardware) and overall throughput. For the co-design study, we utilized an analytical performance modeling tool capable of predicting LLM runtime within 10% of real-world measurements. Our findings offer actionable insights and a practical roadmap for designing AI data centers that can efficiently support trillion-parameter models, reduce optimization complexity, and sustain the rapid evolution of AI capabilities.

The integration of Large Language Models (LLMs) with computer vision is profoundly transforming perception tasks like image segmentation. For intelligent transportation systems (ITS), where accurate scene understanding is critical for safety and efficiency, this new paradigm offers unprecedented capabilities. This survey systematically reviews the emerging field of LLM-augmented image segmentation, focusing on its applications, challenges, and future directions within ITS. We provide a taxonomy of current approaches based on their prompting mechanisms and core architectures, and we highlight how these innovations can enhance road scene understanding for autonomous driving, traffic monitoring, and infrastructure maintenance. Finally, we identify key challenges, including real-time performance and safety-critical reliability, and outline a perspective centered on explainable, human-centric AI as a prerequisite for the successful deployment of this technology in next-generation transportation systems.