AAAI 2025 Papers

Multimodal vision language models (VLMs) have made significant progress with the support of continuously increasing model sizes and data volumes. Running VLMs on edge devices has become a challenge for their widespread application. There are several efficient VLM efforts, but they often sacrifice linguistic capabilities to enhance multimodal abilities, or require extensive training. To address this quandary, we introduce the innovative framework of Efficient Vision Language Models with Elastic Visual Experts (Eve). By strategically incorporating adaptable visual expertise at multiple stages of training, Eve strikes a balance between preserving linguistic abilities and augmenting multimodal capabilities. This balanced approach results in a versatile model with only 1.8B parameters that delivers significant improvements in both multimodal and linguistic tasks. Notably, in configurations below 3B parameters, Eve distinctly outperforms in language benchmarks and achieves state-of-the-art results in VLM Benchmarks. Additionally, its multimodal accuracy outstrips that of the larger 7B LLaVA-1.5 model.

Vision-language retrieval (VLR) has attracted significant attention in both academia and industry, which involves using text (or images) as queries to retrieve corresponding images (or text). However, existing methods often neglect the rich visual semantics knowledge of entities, thus leading to incorrect retrieval results. To address this problem, we propose the Entity Visual Description enhanced CLIP (EvdCLIP), designed to leverage the visual knowledge of entities to enrich queries. Specifically, since humans recognize entities through visual cues, we employ a large language model (LLM) to generate Entity Visual Descriptions (EVDs) as alignment cues to complement textual data. These EVDs are then integrated into raw queries to create visually-rich, EVD-enhanced queries. Furthermore, recognizing that EVD-enhanced queries may introduce noise or low-quality expansions, we develop a novel, trainable EVD-aware Rewriter (EaRW) for vision-language retrieval tasks. EaRW utilizes EVD knowledge and the generative capabilities of the language model to effectively rewrite queries. With our specialized training strategy, EaRW can generate high-quality and low-noise EVD-enhanced queries. Extensive quantitative and qualitative experiments on image-text retrieval benchmarks validate the superiority of EvdCLIP on vision-language retrieval tasks.

Enterprise AI Assistants are increasingly deployed in domains where accuracy is paramount, making each erroneous output a potentially significant incident. This paper presents a comprehensive framework for monitoring, benchmarking, and continuously improving such complex, multi-component systems under active development by multiple teams. Our approach encompasses three key elements: (1) a hierarchical ``severity'' framework for incident detection that identifies and categorizes errors while attributing component-specific error rates, facilitating targeted improvements; (2) a scalable and principled methodology for benchmark construction, evaluation, and deployment, designed to accommodate multiple development teams, mitigate overfitting risks, and assess the downstream impact of system modifications; and (3) a continual improvement strategy leveraging multidimensional evaluation, enabling the identification and implementation of diverse enhancement opportunities. By adopting this holistic framework, organizations can systematically enhance the reliability and performance of their AI Assistants, ensuring their efficacy in critical enterprise environments. We conclude by discussing how this multifaceted evaluation approach opens avenues for various classes of enhancements, paving the way for more robust and trustworthy AI systems.

LLMs-as-a-judge is a recently popularized method which replaces human judgements in task evaluation with automatic evaluation using LLMs. Due to widespread use of RLHF (Reinforcement Learning from Human Feedback), state-of-the-art LLMs like GPT4 and Llama3 are expected to have strong alignment with human preferences when prompted for a quality judgement, such as the coherence of a text. While this seems beneficial, it is not clear whether the assessments by an LLM-as-a-judge constitute only an evaluation based on the instructions in the prompts, or reflect its preference for high-quality data similar to its fine-tune data. To investigate how much influence prompting the LLMs-as-a-judge has on the alignment of AI judgements to human judgements, we analyze prompts with increasing levels of instructions about the target quality of an evaluation, for several LLMs-as-a-judge. Further, we compare to a prompt-free method using model perplexity as a quality measure instead. We aggregate a taxonomy of quality criteria commonly used across state-of-the-art evaluations with LLMs and provide this as a rigorous benchmark of models as judges. Overall, we show that the LLMs-as-a-judge benefit only little from highly detailed instructions in prompts and that perplexity can sometimes align better with human judgements than prompting, especially on textual quality.

The leaderboard of Large Language Models (LLMs) in mathematical tasks has been continuously updated. However, the majority of evaluations focus solely on the final results, neglecting the quality of the intermediate steps. This oversight can mask underlying problems, such as logical errors or unnecessary steps in the reasoning process. To measure reasoning beyond final-answer accuracy, we introduce ReasonEval, a new methodology for evaluating the quality of reasoning steps. ReasonEval employs validity and redundancy to characterize the reasoning quality, as well as accompanying LLMs to assess them automatically. We explore different design options for the LLM-based evaluators and empirically demonstrate that ReasonEval, when instantiated with base models possessing strong mathematical knowledge and trained with high-quality labeled data, consistently outperforms baseline methods in the meta-evaluation datasets. We also highlight the strong generalization capabilities of ReasonEval. By utilizing ReasonEval to evaluate LLMs specialized in math, we find that an increase in final-answer accuracy does not necessarily guarantee an improvement in the overall quality of the reasoning steps for challenging mathematical problems. Additionally, we observe that ReasonEval can play a significant role in data selection. We open-source the best-performing model, meta-evaluation script, and all evaluation results to facilitate future research.

In this paper, we introduce and apply Operations Research Question Answering (ORQA), a new benchmark, to assess the generalization capabilities of Large Language Models (LLMs) in the specialized technical domain of Operations Research (OR). This benchmark is designed to evaluate whether LLMs can emulate the knowledge and reasoning skills of OR experts when given diverse and complex optimization problems. The dataset, crafted by OR experts, presents real-world optimization problems that require multistep reasoning to build their mathematical models. Our evaluations of various open-source LLMs, such as LLaMA 3.1, DeepSeek, and Mixtral reveal their modest performance, indicating a gap in their aptitude to generalize to specialized technical domains. This work contributes to the ongoing discourse on LLMs’ generalization capabilities, providing insights for future research in this area. The dataset and evaluation code are publicly available.

In many applications of AI for Social Impact (e.g., when allocating spots in support programs for underserved communities), resources are scarce and an allocation policy is needed to decide who receives a resource. Before being deployed at scale, a rigorous evaluation of an AI-powered allocation policy is vital. In this paper, we introduce the methods necessary to evaluate index-based allocation policies, which allocate a limited number of resources to those who need them the most. Such policies create dependencies between agents, rendering standard statistical tests invalid and ineffective. Addressing the arising practical and technical challenges, we describe an efficient estimator and methods for drawing valid statistical conclusions. Our extensive experiments validate our methodology in practical settings while also showcasing its statistical power. We conclude by proposing and empirically verifying extensions of our methodology that enable us to reevaluate a past randomized control trial conducted with 10000 beneficiaries for a mHealth program for pregnant women. Our new methodology allows us to draw previously invisible conclusions when comparing two different ML allocation policies.

Despite the impressive success of text-to-image (TTI) models, existing studies overlook the issue of whether these models accurately convey factual information. In this paper, we focus on the problem of image hallucination, where images created by TTI models fail to faithfully depict factual content. To address this, we introduce I-HallA (Image Hallucination evaluation with Question Answering), a novel automated evaluation metric that measures the factuality of generated images through visual question answering (VQA). We also introduce I-HallA v1.0, a curated benchmark dataset for this purpose. As part of this process, we develop a pipeline that generates high-quality question-answer pairs using multiple GPT-4 Omni-based agents, with human judgments to ensure accuracy. Our evaluation protocols measure image hallucination by testing if images from existing TTI models can correctly respond to these questions. The I-HallA v1.0 dataset comprises 1.2K diverse image-text pairs across nine categories with 1,000 rigorously curated questions covering various compositional challenges. We evaluate five TTI models using I-HallA and reveal that these state-of-the-art models often fail to accurately convey factual information. Moreover, we validate the reliability of our metric by demonstrating a strong Spearman correlation (ρ=0.95) with human judgments. Our benchmark dataset and metric can serve as a foundation for developing factually accurate TTI models.

Obtaining high-quality explanations of a model's output enables developers to identify and correct biases, align the system's behavior with human values, and ensure ethical compliance. Explainable Artificial Intelligence (XAI) practitioners rely on specific measures to gauge the quality of such explanations. These measures assess key attributes, such as how closely an explanation aligns with a model's decision process (faithfulness), how accurately it pinpoints the relevant input features (localization), and its consistency across different cases (robustness). Despite providing valuable information, these measures do not fully address a critical practitioner's concern: how does the quality of a given explanation compare to other potential explanations? Traditionally, the quality of an explanation has been assessed by comparing it to a randomly generated counterpart. This paper introduces an alternative: the Quality Gap Estimate (QGE). The QGE method offers a direct comparison to what can be viewed as the `inverse' explanation, one that conceptually represents the antithesis of the original explanation. Our extensive testing across multiple model architectures, datasets, and established quality metrics demonstrates that the QGE method is superior to the traditional approach. Furthermore, we show that QGE enhances the statistical reliability of these quality assessments. This advance represents a significant step toward a more insightful evaluation of explanations that enables a more effective inspection of a model's behavior.

We present EvalAssist, a framework that simplifies the LLM- as-a-judge workflow. The system provides an online criteria development environment, where users can interactively build, test, and share custom evaluation criteria in a structured and portable format. A library of LLM based evaluators is made available that incorporates various algorithmic innovations such as token-probability based judgement, positional bias checking, and certainty estimation that help to engender trust in the evaluation process. We have computed extensive benchmarks and also deployed the system internally in our organization with several hundreds of users.

In the domain of space-time video super-resolution, it is typically challenging to handle complex motions (including large and nonlinear motions) and varying illumination scenes due to the lack of inter-frame information. Leveraging the dense temporal information provided by event signals offers a promising solution. Traditional event-based methods typically rely on multiple images, using motion estimation and compensation, which can introduce errors. Accumulated errors from multiple frames often lead to artifacts and blurriness in the output. To mitigate these issues, we propose EvSTVSR, a method that uses fewer adjacent frames and integrates dense temporal information from events to guide alignment. Additionally, we introduce a coordinate-based feature fusion upsampling module to achieve spatial super-resolution. Experimental results demonstrate that our method not only outperforms existing RGB-based approaches but also excels in handling large motion scenarios.

We present EvHDR-NeRF to recover a High Dynamic Range (HDR) radiance field from event streams and a set of Low Dynamic Range (LDR) views with single exposures. Using the EvHDR-NeRF, we can generate both novel HDR views and novel LDR views under different exposures. The key to our method is to model the new relationship between events streams and LDR images, which considers both the Camera Response Function (CRF) and exposure time. Based on this relationship, we categorize events into inter-frame events and intra-exposure. The former is utilized for building HDR radiance field and the latter is used to deblur potentially blurred images. Compared to existing methods, this method can effectively reconstruct the HDR radiance field even when the input images are degraded. Experimental results demonstrate that our method achieves state-of-the-art HDR reconstruction, providing a more adaptable and accurate solution for complex imaging applications.

High Dynamic Range (HDR) video reconstruction seeks to accurately restore the extensive dynamic range present in real-world scenes and is widely employed in downstream applications. Existing methods typically operate on one or a small number of consecutive frames, which often leads to inconsistent brightness across the video due to their limited perspective on the video sequence. Moreover, supervised learning-based approaches are susceptible to data bias, resulting in reduced effectiveness when confronted with test inputs exhibiting a domain gap relative to the training data. To address these limitations, we present an event-guided HDR video reconstruction method through building 3D Gaussian Splatting (3DGS), to ensure consistent brightness imposed by 3D consistency. We introduce HDR 3D Gaussians capable of simultaneously representing HDR and low-dynamic-range (LDR) colors. Furthermore, we incorporate a learnable HDR-to-LDR transformation optimized by input event streams and LDR frames to eliminate the data bias. Experimental results on both synthetic and real-world datasets demonstrate that the proposed method achieves state-of-the-art performance.

We address unsupervised dependency parsing by building an ensemble of diverse existing models through post hoc aggregation of their output dependency parse structures. We observe that these ensembles often suffer from low robustness against weak ensemble components due to error accumulation. To tackle this problem, we propose an efficient ensemble-selection approach that considers error diversity and avoids error accumulation. Results demonstrate that our approach outperforms each individual model as well as previous ensemble techniques. Additionally, our experiments show that the proposed ensemble-selection method significantly enhances the performance and robustness of our ensemble, surpassing previously proposed strategies, which have not accounted for error diversity.

Gaussian Process Regression (GPR) is a powerful and elegant method for learning complex functions from noisy data with a wide range of applications, including in safety-critical domains. Such applications have two key features: (i) they require rigorous error quantification, and (ii) the noise is often bounded and non-Gaussian due to, e.g., physical constraints. While error bounds for applying GPR in the presence of non-Gaussian noise exist, they tend to be overly restrictive and conservative in practice. In this paper, we provide novel error bounds for GPR under bounded support noise. Specifically, by relying on concentration inequalities and assuming that the latent function has low complexity in the reproducing kernel Hilbert space (RKHS) corresponding to the GP kernel, we derive both probabilistic and deterministic bounds on the error of the GPR. We show that these errors are substantially tighter than existing state-of-the-art bounds and are particularly well-suited for GPR with neural network kernels, i.e., Deep Kernel Learning (DKL). Furthermore, motivated by applications in safety-critical domains, we illustrate how these bounds can be combined with stochastic barrier functions to successfully quantify the safety probability of an unknown dynamical system from finite data. We validate the efficacy of our approach through several benchmarks and comparisons against existing bounds. The results show that our bounds are consistently smaller, and that DKLs can produce error bounds tighter than sample noise, significantly improving the safety probability of control systems.

In recent years, there have been a growing number of works studying the generalization properties of stochastic gradient descent (SGD) from the perspective of algorithmic stability. However, few of them devote to simultaneously studying the generalization and optimization for the non-convex setting, especially pairwise SGD with heavy-tailed gradient noise. This paper considers the impact of the heavy-tailed gradient noise obeying sub-Weibull distribution on the stability-based learning guarantees for non-convex pairwise SGD by investigating its generalization and optimization jointly. Specifically, based on two novel pairwise uniform model stability tools, we firstly bound the generalization error of pairwise SGD in the general non-convex setting after bridging the quantitative relationships between stability and generalization error. Then, we further consider the practical heavy-tailed sub-Weibull gradient noise condition to establish a refined generalization bound without the bounded gradient condition. Finally, sharper error bounds for generalization and optimization are built by introducing the gradient dominance condition. Comparing these results reveals that sub-Weibull gradient noise brings some positive dependencies on the heavy-tailed strength for generalization and optimization. Furthermore, we extend our analysis to the corresponding pairwise minibatch SGD and derive the first stability-based near-optimal generalization and optimization bounds which are consistent with many empirical observations.

Graph unlearning, which aims to eliminate the influence of specific nodes, edges, or attributes from a trained Graph Neural Network (GNN), is essential in applications where privacy, bias, or data obsolescence is a concern. However, existing graph unlearning techniques often necessitate additional training on the remaining data, leading to significant computational costs, particularly with large-scale graphs. To address these challenges, we propose a two-stage training-free approach, Erase then Rectify (ETR), designed for efficient and scalable graph unlearning while preserving the model utility. Specifically, we first build a theoretical foundation showing that masking parameters critical for unlearned samples enables effective unlearning. Building on this insight, the Erase stage strategically edits model parameters to eliminate the impact of unlearned samples and their propagated influence on intercorrelated nodes. To further ensure the GNN's utility, the Rectify stage devises a gradient approximation method to estimate the model's gradient on the remaining dataset, which is then used to enhance model performance. Overall, ETR achieves graph unlearning without additional training or full training data access, significantly reducing computational overhead and preserving data privacy. Extensive experiments on seven public datasets demonstrate the consistent superiority of ETR in model utility, unlearning efficiency, and unlearning effectiveness, establishing it as a promising solution for real-world graph unlearning challenges.

A multimodal fusion technique using LiDAR-camera has been developed for precise 3D object detection in autonomous driving and provides acceptable detection performance in ideal conditions with clear weather. However, the existing multimodal methods are still vulnerable to adverse weather conditions, such as snow, rain, and fog. These factors increase the point cloud sparsity due to occlusion and attenuation of the laser signal. A point cloud becomes sparser with increased distance, posing a challenge for object detection. To address these problems, we propose a point reconstruction network using equirectangular projection for multimodal 3D object detection. This network consists of distance-constrained denoising to remove adverse weather noise and an object-centric ray generator to generate distant object points flexibly. We propose a domain adaptation method that injects feature perturbations to improve detection performance by reducing the domain gap between different datasets. Furthermore, we propose a multimodal weather noise matching method for realistic data synthesis-based training to align the adverse weather noise between synthetic point clouds and images. The experimental results on adverse weather datasets confirm that the proposed approach outperforms the existing methods.

This paper studies a generalized variant of the Colonel Blotto game, referred to as the Colonel Blotto game with costs. Unlike the classic Colonel Blotto game, which imposes the use-it-or-lose-it budget assumption, the Colonel Blotto game with costs captures the strategic importance of costs related both to obtaining resources and assigning them across battlefields. We show that every instance of the Colonel Blotto game with costs is strategically equivalent to an instance of the zero-sum Colonel Blotto game with one additional battlefield. This enables the computation of Nash equilibria of the Colonel Blotto game with costs in polynomial time with respect to the game parameters: the number of battlefields plus the number of resources available to the players.