Abstract:Pointing serves as a fundamental and intuitive mechanism for grounding language within visual contexts, with applications spanning robotics, assistive technologies, and interactive AI systems. While recent multimodal models have started to support pointing capabilities, existing benchmarks typically focus only on referential object localization tasks. We introduce PointArena, a comprehensive platform for evaluating multimodal pointing across diverse reasoning scenarios. PointArena comprises three components: (1) Point-Bench, a curated dataset containing approximately 1,000 pointing tasks across five reasoning categories; (2) Point-Battle, an interactive, web-based arena facilitating blind, pairwise model comparisons, which has already gathered over 4,500 anonymized votes; and (3) Point-Act, a real-world robotic manipulation system allowing users to directly evaluate multimodal model pointing capabilities in practical settings. We conducted extensive evaluations of both state-of-the-art open-source and proprietary multimodal models. Results indicate that Molmo-72B consistently outperforms other models, though proprietary models increasingly demonstrate comparable performance. Additionally, we find that supervised training specifically targeting pointing tasks significantly enhances model performance. Across our multi-stage evaluation pipeline, we also observe strong correlations, underscoring the critical role of precise pointing capabilities in enabling multimodal models to effectively bridge abstract reasoning with concrete, real-world actions. Project page: https://pointarena.github.io/
Abstract:Hamilton-Jacobi Reachability (HJR) analysis has been successfully used in many robotics and control tasks, and is especially effective in computing reach-avoid sets and control laws that enable an agent to reach a goal while satisfying state constraints. However, the original HJR formulation provides no guarantees of safety after a) the prescribed time horizon, or b) goal satisfaction. The reach-avoid-stabilize (RAS) problem has therefore gained a lot of focus: find the set of initial states (the RAS set), such that the trajectory can reach the target, and stabilize to some point of interest (POI) while avoiding obstacles. Solving RAS problems using HJR usually requires defining a new value function, whose zero sub-level set is the RAS set. The existing methods do not consider the problem when there are a series of targets to reach and/or obstacles to avoid. We propose a method that uses the idea of admissible control sets; we guarantee that the system will reach each target while avoiding obstacles as prescribed by the given time series. Moreover, we guarantee that the trajectory ultimately stabilizes to the POI. The proposed method provides an under-approximation of the RAS set, guaranteeing safety. Numerical examples are provided to validate the theory.
Abstract:In this article, we consider the infinite-horizon reach-avoid (RA) and stabilize-avoid (SA) zero-sum game problems for general nonlinear continuous-time systems, where the goal is to find the set of states that can be controlled to reach or stabilize to a target set, without violating constraints even under the worst-case disturbance. Based on the Hamilton-Jacobi reachability method, we address the RA problem by designing a new Lipschitz continuous RA value function, whose zero sublevel set exactly characterizes the RA set. We establish that the associated Bellman backup operator is contractive and that the RA value function is the unique viscosity solution of a Hamilton-Jacobi variational inequality. Finally, we develop a two-step framework for the SA problem by integrating our RA strategies with a recently proposed Robust Control Lyapunov-Value Function, thereby ensuring both target reachability and long-term stability. We numerically verify our RA and SA frameworks on a 3D Dubins car system to demonstrate the efficacy of the proposed approach.
Abstract:Multi-task model merging offers a promising paradigm for integrating multiple expert models into a unified model without additional training. Existing state-of-the-art techniques, such as Task Arithmetic and its variants, merge models by accumulating task vectors -- the parameter differences between pretrained and finetuned models. However, task vector accumulation is often hindered by knowledge conflicts, leading to performance degradation. To address this challenge, we propose Conflict-Aware Task Merging (CAT Merging), a novel training-free framework that selectively trims conflict-prone components from the task vectors. CAT Merging introduces several parameter-specific strategies, including projection for linear weights and masking for scaling and shifting parameters in normalization layers. Extensive experiments on vision, language, and vision-language tasks demonstrate that CAT Merging effectively suppresses knowledge conflicts, achieving average accuracy improvements of up to 2.5% (ViT-B/32) and 2.0% (ViT-L/14) over state-of-the-art methods.
Abstract:Despite impressive advancements in various multimodal tasks, vision-language models (VLMs) still struggle with compositional understanding due to limited exposure to training samples that contain subtle variations within paired examples. With advances in multimodal generative models, a natural solution is to generate synthetic samples with subtle variations for training VLMs. However, generating and training on synthetic samples with subtle variations presents two challenges: difficulty in accurately creating precise variations and inconsistency in cross-modal alignment quality. To address these challenges, we propose SVD-GT (Subtle Variation Data Generation and Training), which integrates image feature injection into a text-to-image generative model to enhance the quality of synthetic variations and employs an adaptive margin loss to differentiate samples using adaptive margins, which help filter out potentially incorrect synthetic samples and focus the learning on informative hard samples. Evaluations on four compositional understanding benchmarks demonstrate that SVD-GT significantly improves the compositionality of VLMs, boosting the average accuracy of CLIP by over 8% across all benchmarks and outperforming state-of-the-art methods by 2% on three benchmarks.
Abstract:Multi-task model merging offers an efficient solution for integrating knowledge from multiple fine-tuned models, mitigating the significant computational and storage demands associated with multi-task training. As a key technique in this field, Task Arithmetic (TA) defines task vectors by subtracting the pre-trained model ($\theta_{\text{pre}}$) from the fine-tuned task models in parameter space, then adjusting the weight between these task vectors and $\theta_{\text{pre}}$ to balance task-generalized and task-specific knowledge. Despite the promising performance of TA, conflicts can arise among the task vectors, particularly when different tasks require distinct model adaptations. In this paper, we formally define this issue as knowledge conflicts, characterized by the performance degradation of one task after merging with a model fine-tuned for another task. Through in-depth analysis, we show that these conflicts stem primarily from the components of task vectors that align with the gradient of task-specific losses at $\theta_{\text{pre}}$. To address this, we propose Task Arithmetic in Trust Region (TATR), which defines the trust region as dimensions in the model parameter space that cause only small changes (corresponding to the task vector components with gradient orthogonal direction) in the task-specific losses. Restricting parameter merging within this trust region, TATR can effectively alleviate knowledge conflicts. Moreover, TATR serves as both an independent approach and a plug-and-play module compatible with a wide range of TA-based methods. Extensive empirical evaluations on eight distinct datasets robustly demonstrate that TATR improves the multi-task performance of several TA-based model merging methods by an observable margin.
Abstract:Current vision-language models may incorporate single-dimensional spatial cues, such as depth, object boundary, and basic spatial directions (e.g. left, right, front, back), yet often lack the multi-dimensional spatial reasoning necessary for human-like understanding and real-world applications. To address this gap, we develop SPHERE (Spatial Perception and Hierarchical Evaluation of REasoning), a hierarchical evaluation framework with a new human-annotated dataset to pinpoint model strengths and weaknesses, advancing from single-skill tasks to multi-skill tasks, and ultimately to complex reasoning tasks that require the integration of multiple spatial and visual cues with logical reasoning. Benchmark evaluation of state-of-the-art open-source models reveal significant shortcomings, especially in the abilities to understand distance and proximity, to reason from both allocentric and egocentric viewpoints, and to perform complex reasoning in a physical context. This work underscores the need for more advanced approaches to spatial understanding and reasoning, paving the way for improvements in vision-language models and their alignment with human-like spatial capabilities. The dataset will be open-sourced upon publication.
Abstract:Vision-Language Models (VLMs) achieved strong performance on a variety of tasks (e.g., image-text retrieval, visual question answering). However, most VLMs rely on coarse-grained image-caption pairs for alignment, relying on data volume to resolve ambiguities and ground linguistic concepts in images. The richer semantic and syntactic structure within text is largely overlooked. To address this, we propose HIerarchically STructured Learning (HIST) that enhances VLM training without any additional supervision, by hierarchically decomposing captions into the constituent Subject, Noun Phrases, and Composite Phrases. Entailment between these constituent components allows us to formulate additional regularization constraints on the VLM attention maps. Specifically, we introduce two novel loss functions: (1) Subject Loss, which aligns image content with the subject of corresponding phrase, acting as an entailment of standard contrastive/matching losses at the Phrase level; (2) Addition Loss, to balance attention across multiple objects. HIST is general, and can be applied to any VLM for which attention between vision and language can be computed; we illustrate its efficacy on BLIP and ALBEF. HIST outperforms baseline VLMs, achieving up to +9.8% improvement in visual grounding, +6.3% in multi-object referring segmentation, +1.1% in image-text retrieval, and +0.2% in visual question answering, underscoring the value of structuring learning in VLMs.
Abstract:The commonsense reasoning capabilities of vision-language models (VLMs), especially in abductive reasoning and defeasible reasoning, remain poorly understood. Most benchmarks focus on typical visual scenarios, making it difficult to discern whether model performance stems from keen perception and reasoning skills, or reliance on pure statistical recall. We argue that by focusing on atypical events in videos, clearer insights can be gained on the core capabilities of VLMs. Explaining and understanding such out-of-distribution events requires models to extend beyond basic pattern recognition and regurgitation of their prior knowledge. To this end, we introduce BlackSwanSuite, a benchmark for evaluating VLMs' ability to reason about unexpected events through abductive and defeasible tasks. Our tasks artificially limit the amount of visual information provided to models while questioning them about hidden unexpected events, or provide new visual information that could change an existing hypothesis about the event. We curate a comprehensive benchmark suite comprising over 3,800 MCQ, 4,900 generative and 6,700 yes/no tasks, spanning 1,655 videos. After extensively evaluating various state-of-the-art VLMs, including GPT-4o and Gemini 1.5 Pro, as well as open-source VLMs such as LLaVA-Video, we find significant performance gaps of up to 32% from humans on these tasks. Our findings reveal key limitations in current VLMs, emphasizing the need for enhanced model architectures and training strategies.
Abstract:Visual grounding aims to localize the image regions based on a textual query. Given the difficulty of large-scale data curation, we investigate how to effectively learn visual grounding under data-scarce settings in this paper. To address data scarcity, we propose a novel framework, POBF (Paint Outside the Box, then Filter). POBF synthesizes images by inpainting outside the box, tackling a label misalignment issue encountered in previous works. Furthermore, POBF leverages an innovative filtering scheme to identify the most effective training data. This scheme combines a hardness score and an overfitting score, balanced by a penalty term. Experimental results show that POBF achieves superior performance across four datasets, delivering an average improvement of 5.83% and outperforming leading baselines by 2.29% to 3.85% in accuracy. Additionally, we validate the robustness and generalizability of POBF across various generative models, data ratios, and model architectures.