Mixture-of-Experts large language models (MoE-LLMs) marks a significant step forward of language models, however, they encounter two critical challenges in practice: 1) expert parameters lead to considerable memory consumption and loading latency; and 2) the current activated experts are redundant, as many tokens may only require a single expert. Motivated by these issues, we investigate the MoE-LLMs and make two key observations: a) different experts exhibit varying behaviors on activation reconstruction error, routing scores, and activated frequencies, highlighting their differing importance, and b) not all tokens are equally important-only a small subset is critical. Building on these insights, we propose MC-MoE, a training-free Mixture-Compressor for MoE-LLMs, which leverages the significance of both experts and tokens to achieve an extreme compression. First, to mitigate storage and loading overheads, we introduce Pre-Loading Mixed-Precision Quantization (PMQ), which formulates the adaptive bit-width allocation as a Linear Programming (LP) problem, where the objective function balances multi-factors reflecting the importance of each expert. Additionally, we develop Online Dynamic Pruning (ODP), which identifies important tokens to retain and dynamically select activated experts for other tokens during inference to optimize efficiency while maintaining performance. Our MC-MoE integrates static quantization and dynamic pruning to collaboratively achieve extreme compression for MoE-LLMs with less accuracy loss, ensuring an optimal trade-off between performance and efficiency. Extensive experiments confirm the effectiveness of our approach. During dynamic inference, we further reduce activated parameters by 15%, with a performance drop of less than 0.6%. Remarkably, MC-MoE even surpasses floating-point 13b dense LLMs with significantly smaller parameter sizes, suggesting that mixture compression in MoE-LLMs has the potential to outperform both comparable and larger dense LLMs.

Ensuring model explainability and robustness is essential for reliable deployment of deep vision systems. Current methods for evaluating robustness rely on collecting and annotating extensive test sets. While this is common practice, the process is labor-intensive and expensive with no guarantee of sufficient coverage across attributes of interest. Recently, model diagnosis frameworks have emerged leveraging user inputs (e.g., text) to assess the vulnerability of the model. However, such dependence on human can introduce bias and limitation given the domain knowledge of particular users. This paper proposes Unsupervised Model Diagnosis (UMO), that leverages generative models to produce semantic counterfactual explanations without any user guidance. Given a differentiable computer vision model (i.e., the target model), UMO optimizes for the most counterfactual directions in a generative latent space. Our approach identifies and visualizes changes in semantics, and then matches these changes to attributes from wide-ranging text sources, such as dictionaries or language models. We validate the framework on multiple vision tasks (e.g., classification, segmentation, keypoint detection). Extensive experiments show that our unsupervised discovery of semantic directions can correctly highlight spurious correlations and visualize the failure mode of target models without any human intervention.

Despite their success in many domains, large language models (LLMs) remain under-studied in scenarios requiring optimal decision-making under uncertainty. This is crucial as many real-world applications, ranging from personalized recommendations to healthcare interventions, demand that LLMs not only predict but also actively learn to make optimal decisions through exploration. In this work, we measure LLMs' (in)ability to make optimal decisions in bandits, a state-less reinforcement learning setting relevant to many applications. We develop a comprehensive suite of environments, including both context-free and contextual bandits with varying task difficulties, to benchmark LLMs' performance. Motivated by the existence of optimal exploration algorithms, we propose efficient ways to integrate this algorithmic knowledge into LLMs: by providing explicit algorithm-guided support during inference; and through algorithm distillation via in-context demonstrations and fine-tuning, using synthetic data generated from these algorithms. Impressively, these techniques allow us to achieve superior exploration performance with smaller models, surpassing larger models on various tasks. We conducted an extensive ablation study to shed light on various factors, such as task difficulty and data representation, that influence the efficiency of LLM exploration. Additionally, we conduct a rigorous analysis of the LLM's exploration efficiency using the concept of regret, linking its ability to explore to the model size and underlying algorithm.

To operate at a building scale, service robots must perform very long-horizon mobile manipulation tasks by navigating to different rooms, accessing different floors, and interacting with a wide and unseen range of everyday objects. We refer to these tasks as Building-wide Mobile Manipulation. To tackle these inherently long-horizon tasks, we introduce BUMBLE, a unified Vision-Language Model (VLM)-based framework integrating open-world RGBD perception, a wide spectrum of gross-to-fine motor skills, and dual-layered memory. Our extensive evaluation (90+ hours) indicates that BUMBLE outperforms multiple baselines in long-horizon building-wide tasks that require sequencing up to 12 ground truth skills spanning 15 minutes per trial. BUMBLE achieves 47.1% success rate averaged over 70 trials in different buildings, tasks, and scene layouts from different starting rooms and floors. Our user study demonstrates 22% higher satisfaction with our method than state-of-the-art mobile manipulation methods. Finally, we demonstrate the potential of using increasingly-capable foundation models to push performance further. For more information, see https://robin-lab.cs.utexas.edu/BUMBLE/

As artificial intelligence (AI) systems advance, we move towards broad AI: systems capable of performing well on diverse tasks, understanding context, and adapting rapidly to new scenarios. A central challenge for broad AI systems is to generalize over tasks in related domains and being robust to distribution shifts. Neuro-symbolic (NeSy) AI bridges the gap between symbolic and sub-symbolic paradigms to address these challenges, enabling adaptable, generalizable, and more interpretable systems. The development of broad AI requires advancements in domain adaptation (DA), enabling models trained on source domains to effectively generalize to unseen target domains. Traditional approaches often rely on parameter optimization and fine-tuning, which can be impractical due to high costs and risks of catastrophic forgetting. NeSy AI systems use multiple models and methods to generalize to unseen domains and maintain performance across varying conditions. We analyze common DA and NeSy approaches with a focus on deep domain-invariant learning, extending to real-world challenges such as adapting to continuously changing domains and handling large domain gaps. We showcase state-of-the-art model-selection methods for scenarios with limited samples and introduce domain-specific adaptations without gradient-based updates for cases where model tuning is infeasible. This work establishes a framework for scalable and generalizable broad AI systems applicable across various problem settings, demonstrating how symbolic reasoning and large language models can build universal computational graphs that generalize across domains and problems, contributing to more adaptable AI approaches for real-world applications.

Large vision and language assistants have enabled new capabilities for interpreting natural images. These approaches have recently been adapted to earth observation data, but they are only able to handle single image inputs, limiting their use for many real-world tasks. In this work, we develop a new vision and language assistant called TEOChat that can engage in conversations about temporal sequences of earth observation data. To train TEOChat, we curate an instructionfollowing dataset composed of many single image and temporal tasks including building change and damage assessment, semantic change detection, and temporal scene classification. We show that TEOChat can perform a wide variety of spatial and temporal reasoning tasks, substantially outperforming previous vision and language assistants, and even achieving comparable or better performance than specialist models trained to perform these specific tasks. Furthermore, TEOChat achieves impressive zero-shot performance on a change detection and change question answering dataset, outperforms GPT-4o and Gemini 1.5 Pro on multiple temporal tasks, and exhibits stronger single image capabilities than a comparable single EO image instruction-following model. Many earth observation (EO) tasks require the ability to reason over time. For example, change detection is a widely studied task where the goal is to identify salient changes in a region using multiple EO images capturing the region at different times (Chughtai et al., 2021; Bai et al., 2023; Cheng et al., 2023). Previous methods to automatically detect change in EO imagery have been specialist models, constraining their use to a single task or small set of tasks that they were explicitly trained to perform (Bai et al., 2023; Cheng et al., 2023). Advancements in the modeling of multimodal data have enabled generalist vision-language models (VLMs) that can perform a variety of natural image interpretation tasks specified flexibly through natural language (Achiam et al., 2023; Team et al., 2023; Liu et al., 2023). However, no prior VLMs can model temporal EO data (left of Figure 1), notably including change detection tasks. We investigate the performance of Video-LLaVA (Lin et al., 2023), a strong natural image pre-trained VLM that can receive images and videos as input, and GeoChat (Kuckreja et al., 2023), a strong VLM fine-tuned on single EO image tasks (right of Figure 1). We find that Video-LLaVA generates inaccurate information, likely because it has primarily been trained on natural images and videos, whereas GeoChat can only input single images and cannot process information across time. TEOChat is the first VLM to model temporal earth observation (EO) data. We compare a temporal VLM (Video-LLaVA (Lin et al., 2023)) and an EO VLM (GeoChat (Kuckreja et al., 2023)) with TEOChat.

The ability to steer AI behavior is crucial to preventing its long term dangerous and catastrophic potential. Representation Engineering (RepE) has emerged as a novel, powerful method to steer internal model behaviors, such as "honesty", at a top-down level. Understanding the steering of representations should thus be placed at the forefront of alignment initiatives. Unfortunately, current efforts to understand plasticity at this level are highly neglected. This paper aims to bridge the knowledge gap and understand how LLM representation stability, specifically for the concept of "honesty", and model plasticity evolve by applying steering vectors extracted at different fine-tuning stages, revealing differing magnitudes of shifts in model behavior. The findings are pivotal, showing that while early steering exhibits high plasticity, later stages have a surprisingly responsive critical window. This pattern is observed across different model architectures, signaling that there is a general pattern of model plasticity that can be used for effective intervention. These insights greatly contribute to the field of AI transparency, addressing a pressing lack of efficiency limiting our ability to effectively steer model behavior.

In reinforcement learning, if the agent's reward differs from the designers' true utility, even only rarely, the state distribution resulting from the agent's policy can be very bad, in theory and in practice. When RL policies would devolve into undesired behavior, a common countermeasure is KL regularization to a trusted policy ("Don't do anything I wouldn't do"). All current cutting-edge language models are RL agents that are KL-regularized to a "base policy" that is purely predictive. Unfortunately, we demonstrate that when this base policy is a Bayesian predictive model of a trusted policy, the KL constraint is no longer reliable for controlling the behavior of an advanced RL agent. We demonstrate this theoretically using algorithmic information theory, and while systems today are too weak to exhibit this theorized failure precisely, we RL-finetune a language model and find evidence that our formal results are plausibly relevant in practice. We also propose a theoretical alternative that avoids this problem by replacing the "Don't do anything I wouldn't do" principle with "Don't do anything I mightn't do".

Positional Encodings (PEs) are a critical component of Transformer-based Large Language Models (LLMs), providing the attention mechanism with important sequence-position information. One of the most popular types of encoding used today in LLMs are Rotary Positional Encodings (RoPE), that rotate the queries and keys based on their relative distance. A common belief is that RoPE is useful because it helps to decay token dependency as relative distance increases. In this work, we argue that this is unlikely to be the core reason. We study the internals of a trained Gemma 7B model to understand how RoPE is being used at a mechanical level. We find that Gemma learns to use RoPE to construct robust'positional' attention patterns by exploiting the highest frequencies. We also find that, in general, Gemma greatly prefers to use the lowest frequencies of RoPE, which we suspect are used to carry semantic information. We mathematically prove interesting behaviours of RoPE and conduct experiments to verify our findings, proposing a modification of RoPE that fixes some highlighted issues and improves performance. We believe that this work represents an interesting step in better understanding PEs in LLMs, which we believe holds crucial value for scaling LLMs to large sizes and context lengths. It is common to provide positional information to the attention mechanism in Transformers through the use of absolute positional encodings (Vaswani et al., 2017), relative positional encodings (Su et al., 2024), or by introducing a bias directly to the activations (Press et al., 2021). One of the currently most widely adopted encodings, especially in Large Language Models (LLMs), are Rotary Positional Encodings (RoPE) (Su et al., 2024), being used in popular models such as LLama 3 (Dubey et al., 2024) and Gemma (Gemma Team et al., 2024). The method can be implemented efficiently and provides an interesting geometric approach to positional encodings. Despite the significant adoption of RoPE, the specific reasons why this method is useful to Transformer models remains poorly understood. One of the main arguments in favour of RoPE made by Su et al. (2024) is that the method helps to decay attention coefficients as the relative distance grows. Most such claims, however, rely on the queries and keys being constant vectors - which is uncommon in practice. In fact, in this work we find that there are many situations in which this decay does not occur and that this is exploited at times by attention heads in Gemma 7B (Gemma Team et al., 2024).

Generating high-quality, in-depth textual documents, such as academic papers, news articles, Wikipedia entries, and books, remains a significant challenge for Large Language Models (LLMs). In this paper, we propose to use planning to generate long form content. To achieve our goal, we generate intermediate steps via an auxiliary task that teaches the LLM to plan, reason and structure before generating the final text. Our main novelty lies in a single auxiliary task that does not require multiple rounds of prompting or planning. To overcome the scarcity of training data for these intermediate steps, we leverage LLMs to generate synthetic intermediate writing data such as outlines, key information and summaries from existing full articles. Our experiments demonstrate on two datasets from different domains, namely the scientific news dataset SciNews and Wikipedia datasets in KILT-Wiki and FreshWiki, that LLMs fine-tuned with the auxiliary task generate higher quality documents. We observed +2.5% improvement in ROUGE-Lsum, and a strong 3.60 overall win/loss ratio via human SxS evaluation, with clear wins in organization, relevance, and verifiability.