On Domain-Specific Post-Training for Multimodal Large Language Models

Recent years have witnessed the rapid development of general multimodal large language models (MLLMs). However, adapting general MLLMs to specific domains, such as scientific fields and industrial applications, remains less explored. This paper systematically investigates domain adaptation of MLLMs through post-training, focusing on data synthesis, training pipelines, and task evaluation. (1) Data Synthesis: Using open-source models, we develop a visual instruction synthesizer that effectively generates diverse visual instruction tasks from domain-specific image-caption pairs. Our synthetic tasks surpass those generated by manual rules, GPT-4, and GPT-4V in enhancing the domain-specific performance of MLLMs. (2) Training Pipeline: While the two-stage training--initially on image-caption pairs followed by visual instruction tasks--is commonly adopted for developing general MLLMs, we apply a single-stage training pipeline to enhance task diversity for domain-specific post-training. (3) Task Evaluation: We conduct experiments in two domains, biomedicine and food, by post-training MLLMs of different sources and scales (e.g., Qwen2-VL-2B, LLaVA-v1.6-8B, Llama-3.2-11B), and then evaluating MLLM performance on various domain-specific tasks. To support further research in MLLM domain adaptation, we will open-source our implementations.

HabitAction: A Video Dataset for Human Habitual Behavior Recognition

Human Action Recognition (HAR) is a very crucial task in computer vision. It helps to carry out a series of downstream tasks, like understanding human behaviors. Due to the complexity of human behaviors, many highly valuable behaviors are not yet encompassed within the available datasets for HAR, e.g., human habitual behaviors (HHBs). HHBs hold significant importance for analyzing a person's personality, habits, and psychological changes. To solve these problems, in this work, we build a novel video dataset to demonstrate various HHBs. These behaviors in the proposed dataset are able to reflect internal mental states and specific emotions of the characters, e.g., crossing arms suggests to shield oneself from perceived threats. The dataset contains 30 categories of habitual behaviors including more than 300,000 frames and 6,899 action instances. Since these behaviors usually appear at small local parts of human action videos, it is difficult for existing action recognition methods to handle these local features. Therefore, we also propose a two-stream model using both human skeletons and RGB appearances. Experimental results demonstrate that our proposed method has much better performance in action recognition than the existing methods on the proposed dataset.

On the Emergence of Symmetrical Reality

Artificial intelligence (AI) has revolutionized human cognitive abilities and facilitated the development of new AI entities capable of interacting with humans in both physical and virtual environments. Despite the existence of virtual reality, mixed reality, and augmented reality for several years, integrating these technical fields remains a formidable challenge due to their disparate application directions. The advent of AI agents, capable of autonomous perception and action, further compounds this issue by exposing the limitations of traditional human-centered research approaches. It is imperative to establish a comprehensive framework that accommodates the dual perceptual centers of humans and AI agents in both physical and virtual worlds. In this paper, we introduce the symmetrical reality framework, which offers a unified representation encompassing various forms of physical-virtual amalgamations. This framework enables researchers to better comprehend how AI agents can collaborate with humans and how distinct technical pathways of physical-virtual integration can be consolidated from a broader perspective. We then delve into the coexistence of humans and AI, demonstrating a prototype system that exemplifies the operation of symmetrical reality systems for specific tasks, such as pouring water. Subsequently, we propose an instance of an AI-driven active assistance service that illustrates the potential applications of symmetrical reality. This paper aims to offer beneficial perspectives and guidance for researchers and practitioners in different fields, thus contributing to the ongoing research about human-AI coexistence in both physical and virtual environments.

Bongard-OpenWorld: Few-Shot Reasoning for Free-form Visual Concepts in the Real World

We introduce Bongard-OpenWorld, a new benchmark for evaluating real-world few-shot reasoning for machine vision. It originates from the classical Bongard Problems (BPs): Given two sets of images (positive and negative), the model needs to identify the set that query images belong to by inducing the visual concepts, which is exclusively depicted by images from the positive set. Our benchmark inherits the few-shot concept induction of the original BPs while adding the two novel layers of challenge: 1) open-world free-form concepts, as the visual concepts in Bongard-OpenWorld are unique compositions of terms from an open vocabulary, ranging from object categories to abstract visual attributes and commonsense factual knowledge; 2) real-world images, as opposed to the synthetic diagrams used by many counterparts. In our exploration, Bongard-OpenWorld already imposes a significant challenge to current few-shot reasoning algorithms. We further investigate to which extent the recently introduced Large Language Models (LLMs) and Vision-Language Models (VLMs) can solve our task, by directly probing VLMs, and combining VLMs and LLMs in an interactive reasoning scheme. We even designed a neuro-symbolic reasoning approach that reconciles LLMs & VLMs with logical reasoning to emulate the human problem-solving process for Bongard Problems. However, none of these approaches manage to close the human-machine gap, as the best learner achieves 64% accuracy while human participants easily reach 91%. We hope Bongard-OpenWorld can help us better understand the limitations of current visual intelligence and facilitate future research on visual agents with stronger few-shot visual reasoning capabilities.

Reconfigurable Data Glove for Reconstructing Physical and Virtual Grasps

We present a reconfigurable data glove design to capture different modes of human hand-object interactions, critical for training embodied AI agents for fine manipulation tasks. Sharing a unified backbone design that reconstructs hand gestures in real-time, our reconfigurable data glove operates in three modes for various downstream tasks with distinct features. In the tactile-sensing mode, the glove system aggregates manipulation force via customized force sensors made from a soft and thin piezoresistive material; this design is to minimize interference during complex hand movements. The Virtual Reality (VR) mode enables real-time interaction in a physically plausible fashion; a caging-based approach is devised to determine stable grasps by detecting collision events. Leveraging a state-of-the-art Finite Element Method (FEM) simulator, the simulation mode collects a fine-grained 4D manipulation event: hand and object motions in 3D space and how the object's physical properties (e.g., stress, energy) change in accord with the manipulation in time. Of note, this glove system is the first to look into, through high-fidelity simulation, the unobservable physical and causal factors behind manipulation actions. In a series of experiments, we characterize our data glove in terms of individual sensors and the overall system. Specifically, we evaluate the system's three modes by (i) recording hand gestures and associated forces, (ii) improving manipulation fluency in VR, and (iii) producing realistic simulation effects of various tool uses, respectively. Together, our reconfigurable data glove collects and reconstructs fine-grained human grasp data in both the physical and virtual environments, opening up new avenues to learning manipulation skills for embodied AI agents.

SCOBO: Sparsity-Aware Comparison Oracle Based Optimization

We study derivative-free optimization for convex functions where we further assume that function evaluations are unavailable. Instead, one only has access to a comparison oracle, which, given two points $x$ and $y$, and returns a single bit of information indicating which point has larger function value, $f(x)$ or $f(y)$, with some probability of being incorrect. This probability may be constant or it may depend on $|f(x)-f(y)|$. Previous algorithms for this problem have been hampered by a query complexity which is polynomially dependent on the problem dimension, $d$. We propose a novel algorithm that breaks this dependence: it has query complexity only logarithmically dependent on $d$ if the function in addition has low dimensional structure that can be exploited. Numerical experiments on synthetic data and the MuJoCo dataset show that our algorithm outperforms state-of-the-art methods for comparison based optimization, and is even competitive with other derivative-free algorithms that require explicit function evaluations.

Zeroth-Order Regularized Optimization (ZORO): Approximately Sparse Gradients and Adaptive Sampling

We consider the problem of minimizing a high-dimensional objective function, which may include a regularization term, using (possibly noisy) evaluations of the function. Such optimization is also called derivative-free, zeroth-order, or black-box optimization. We propose a new $\textbf{Z}$eroth-$\textbf{O}$rder $\textbf{R}$egularized $\textbf{O}$ptimization method, dubbed ZORO. When the underlying gradient is approximately sparse at an iterate, ZORO needs very few objective function evaluations to obtain a new iterate that decreases the objective function. We achieve this with an adaptive, randomized gradient estimator, followed by an inexact proximal-gradient scheme. Under a novel approximately sparse gradient assumption and various different convex settings, we show the (theoretical and empirical) convergence rate of ZORO is only logarithmically dependent on the problem dimension. Numerical experiments show that ZORO outperforms the existing methods with similar assumptions, on both synthetic and real datasets.

VRGym: A Virtual Testbed for Physical and Interactive AI

We propose VRGym, a virtual reality testbed for realistic human-robot interaction. Different from existing toolkits and virtual reality environments, the VRGym emphasizes on building and training both physical and interactive agents for robotics, machine learning, and cognitive science. VRGym leverages mechanisms that can generate diverse 3D scenes with high realism through physics-based simulation. We demonstrate that VRGym is able to (i) collect human interactions and fine manipulations, (ii) accommodate various robots with a ROS bridge, (iii) support experiments for human-robot interaction, and (iv) provide toolkits for training the state-of-the-art machine learning algorithms. We hope VRGym can help to advance general-purpose robotics and machine learning agents, as well as assisting human studies in the field of cognitive science.

Influential Node Detection in Implicit Social Networks using Multi-task Gaussian Copula Models

Influential node detection is a central research topic in social network analysis. Many existing methods rely on the assumption that the network structure is completely known \textit{a priori}. However, in many applications, network structure is unavailable to explain the underlying information diffusion phenomenon. To address the challenge of information diffusion analysis with incomplete knowledge of network structure, we develop a multi-task low rank linear influence model. By exploiting the relationships between contagions, our approach can simultaneously predict the volume (i.e. time series prediction) for each contagion (or topic) and automatically identify the most influential nodes for each contagion. The proposed model is validated using synthetic data and an ISIS twitter dataset. In addition to improving the volume prediction performance significantly, we show that the proposed approach can reliably infer the most influential users for specific contagions.

Hypothesis Testing in Feedforward Networks with Broadcast Failures

Consider a countably infinite set of nodes, which sequentially make decisions between two given hypotheses. Each node takes a measurement of the underlying truth, observes the decisions from some immediate predecessors, and makes a decision between the given hypotheses. We consider two classes of broadcast failures: 1) each node broadcasts a decision to the other nodes, subject to random erasure in the form of a binary erasure channel; 2) each node broadcasts a randomly flipped decision to the other nodes in the form of a binary symmetric channel. We are interested in whether there exists a decision strategy consisting of a sequence of likelihood ratio tests such that the node decisions converge in probability to the underlying truth. In both cases, we show that if each node only learns from a bounded number of immediate predecessors, then there does not exist a decision strategy such that the decisions converge in probability to the underlying truth. However, in case 1, we show that if each node learns from an unboundedly growing number of predecessors, then the decisions converge in probability to the underlying truth, even when the erasure probabilities converge to 1. We also derive the convergence rate of the error probability. In case 2, we show that if each node learns from all of its previous predecessors, then the decisions converge in probability to the underlying truth when the flipping probabilities of the binary symmetric channels are bounded away from 1/2. In the case where the flipping probabilities converge to 1/2, we derive a necessary condition on the convergence rate of the flipping probabilities such that the decisions still converge to the underlying truth. We also explicitly characterize the relationship between the convergence rate of the error probability and the convergence rate of the flipping probabilities.