Thao Minh Le

Thao Minh Le, Vuong Le, Sunil Gupta et al.

The current success of modern visual reasoning systems is arguably attributed to cross-modality attention mechanisms. However, in deliberative reasoning such as in VQA, attention is unconstrained at each step, and thus may serve as a statistical pooling mechanism rather than a semantic operation intended to select information relevant to inference. This is because at training time, attention is only guided by a very sparse signal (i.e. the answer label) at the end of the inference chain. This causes the cross-modality attention weights to deviate from the desired visual-language bindings. To rectify this deviation, we propose to guide the attention mechanism using explicit linguistic-visual grounding. This grounding is derived by connecting structured linguistic concepts in the query to their referents among the visual objects. Here we learn the grounding from the pairing of questions and images alone, without the need for answer annotation or external grounding supervision. This grounding guides the attention mechanism inside VQA models through a duality of mechanisms: pre-training attention weight calculation and directly guiding the weights at inference time on a case-by-case basis. The resultant algorithm is capable of probing attention-based reasoning models, injecting relevant associative knowledge, and regulating the core reasoning process. This scalable enhancement improves the performance of VQA models, fortifies their robustness to limited access to supervised data, and increases interpretability.

Hoang-Anh Pham, Thao Minh Le, Vuong Le et al.

It would be a technological feat to be able to create a system that can hold a meaningful conversation with humans about what they watch. A setup toward that goal is presented as a video dialog task, where the system is asked to generate natural utterances in response to a question in an ongoing dialog. The task poses great visual, linguistic, and reasoning challenges that cannot be easily overcome without an appropriate representation scheme over video and dialog that supports high-level reasoning. To tackle these challenges we present a new object-centric framework for video dialog that supports neural reasoning dubbed COST - which stands for Conversation about Objects in Space-Time. Here dynamic space-time visual content in videos is first parsed into object trajectories. Given this video abstraction, COST maintains and tracks object-associated dialog states, which are updated upon receiving new questions. Object interactions are dynamically and conditionally inferred for each question, and these serve as the basis for relational reasoning among them. COST also maintains a history of previous answers, and this allows retrieval of relevant object-centric information to enrich the answer forming process. Language production then proceeds in a step-wise manner, taking into the context of the current utterance, the existing dialog, the current question. We evaluate COST on the DSTC7 and DSTC8 benchmarks, demonstrating its competitiveness against state-of-the-arts.

Tuyen Tran, Thao Minh Le, Hung Tran et al.

Recognizing human activities in videos is challenging due to the spatio-temporal complexity and context-dependence of human interactions. Prior studies often rely on single input modalities, such as RGB or skeletal data, limiting their ability to exploit the complementary advantages across modalities. Recent studies focus on combining these two modalities using simple feature fusion techniques. However, due to the inherent disparities in representation between these input modalities, designing a unified neural network architecture to effectively leverage their complementary information remains a significant challenge. To address this, we propose a comprehensive multimodal framework for robust video-based human activity recognition. Our key contribution is the introduction of a novel compositional query machine, called COMPUTER ($\textbf{COMP}ositional h\textbf{U}man-cen\textbf{T}ric qu\textbf{ER}y$ machine), a generic neural architecture that models the interactions between a human of interest and its surroundings in both space and time. Thanks to its versatile design, COMPUTER can be leveraged to distill distinctive representations for various input modalities. Additionally, we introduce a consistency loss that enforces agreement in prediction between modalities, exploiting the complementary information from multimodal inputs for robust human movement recognition. Through extensive experiments on action localization and group activity recognition tasks, our approach demonstrates superior performance when compared with state-of-the-art methods. Our code is available at: https://github.com/tranxuantuyen/COMPUTER.

Long Hoang Dang, Thao Minh Le, Vuong Le et al.

Large vision-language models (LVLMs) offer a novel capability for performing in-context learning (ICL) in Visual QA. When prompted with a few demonstrations of image-question-answer triplets, LVLMs have demonstrated the ability to discern underlying patterns and transfer this latent knowledge to answer new questions about unseen images without the need for expensive supervised fine-tuning. However, designing effective vision-language prompts, especially for compositional questions, remains poorly understood. Adapting language-only ICL techniques may not necessarily work because we need to bridge the visual-linguistic semantic gap: Symbolic concepts must be grounded in visual content, which does not share the syntactic linguistic structures. This paper introduces SADL, a new visual-linguistic prompting framework for the task. SADL revolves around three key components: SAmpling, Deliberation, and Pseudo-Labeling of image-question pairs. Given an image-question query, we sample image-question pairs from the training data that are in semantic proximity to the query. To address the compositional nature of questions, the deliberation step decomposes complex questions into a sequence of subquestions. Finally, the sequence is progressively annotated one subquestion at a time to generate a sequence of pseudo-labels. We investigate the behaviors of SADL under OpenFlamingo on large-scale Visual QA datasets, namely GQA, GQA-OOD, CLEVR, and CRIC. The evaluation demonstrates the critical roles of sampling in the neighborhood of the image, the decomposition of complex questions, and the accurate pairing of the subquestions and labels. These findings do not always align with those found in language-only ICL, suggesting fresh insights in vision-language settings.

Thong Bach, Dung Nguyen, Thao Minh Le et al.

Large language models require continuous adaptation to new tasks while preserving safety alignment. However, fine-tuning on even benign data often compromises safety behaviors, including refusal of harmful requests, truthfulness, and commonsense reasoning. We investigate which training samples cause alignment drift through a data-centric lens. Our empirical analysis shows samples contribute unequally: high-gradient samples cause greater safety degradation and drive models toward pretrained distributions, while moderate-gradient samples enable task learning with minimal alignment loss. We propose gradient-based sample selection that filters high-gradient samples during fine-tuning. Across multiple model families on continual domain tasks, our method substantially improves alignment preservation while maintaining competitive task performance, without requiring curated safe data or architectural modifications. Our method is robust across selection ratios, task orderings, and diverse attack benchmarks.

Thong Bach, Dung Nguyen, Thao Minh Le et al.

Large language models exhibit systematic vulnerabilities to adversarial attacks despite extensive safety alignment. We provide a mechanistic analysis revealing that position-dependent gradient weakening during autoregressive training creates signal decay, leading to incomplete safety learning where safety training fails to transform model preferences in later response regions fully. We introduce base-favored tokens -- vocabulary elements where base models assign higher probability than aligned models -- as computational indicators of incomplete safety learning and develop a targeted completion method that addresses undertrained regions through adaptive penalties and hybrid teacher distillation. Experimental evaluation across Llama and Qwen model families demonstrates dramatic improvements in adversarial robustness, with 48--98% reductions in attack success rates while preserving general capabilities. These results establish both a mechanistic understanding and practical solutions for fundamental limitations in safety alignment methodologies.

Thao Minh Le

Visual perception and language understanding are - fundamental components of human intelligence, enabling them to understand and reason about objects and their interactions. It is crucial for machines to have this capacity to reason using these two modalities to invent new robot-human collaborative systems. Recent advances in deep learning have built separate sophisticated representations of both visual scenes and languages. However, understanding the associations between the two modalities in a shared context for multimodal reasoning remains a challenge. Focusing on language and vision modalities, this thesis advances the understanding of how to exploit and use pivotal aspects of vision-and-language tasks with neural networks to support reasoning. We derive these understandings from a series of works, making a two-fold contribution: (i) effective mechanisms for content selection and construction of temporal relations from dynamic visual scenes in response to a linguistic query and preparing adequate knowledge for the reasoning process (ii) new frameworks to perform reasoning with neural networks by exploiting visual-linguistic associations, deduced either directly from data or guided by external priors.

Quang-Hung Le, Long Hoang Dang, Ngan Le et al.

Existing Large Vision-Language Models (LVLMs) excel at matching concepts across multi-modal inputs but struggle with compositional concepts and high-level relationships between entities. This paper introduces Progressive multi-granular Vision-Language alignments (PromViL), a novel framework to enhance LVLMs' ability in performing grounded compositional visual reasoning tasks. Our approach constructs a hierarchical structure of multi-modal alignments, ranging from simple to complex concepts. By progressively aligning textual descriptions with corresponding visual regions, our model learns to leverage contextual information from lower levels to inform higher-level reasoning. To facilitate this learning process, we introduce a data generation process that creates a novel dataset derived from Visual Genome, providing a wide range of nested compositional vision-language pairs. Experimental results demonstrate that our PromViL framework significantly outperforms baselines on various visual grounding and compositional question answering tasks. The code is available at: https://github.com/lqh52/PromViL.

Thao Minh Le, Vuong Le, Kien Do et al.

This paper introduces a new problem, Causal Abductive Reasoning on Video Events (CARVE), which involves identifying causal relationships between events in a video and generating hypotheses about causal chains that account for the occurrence of a target event. To facilitate research in this direction, we create two new benchmark datasets with both synthetic and realistic videos, accompanied by trigger-target labels generated through a novel counterfactual synthesis approach. To explore the challenge of solving CARVE, we present a Causal Event Relation Network (CERN) that examines the relationships between video events in temporal and semantic spaces to efficiently determine the root-cause trigger events. Through extensive experiments, we demonstrate the critical roles of event relational representation learning and interaction modeling in solving video causal reasoning challenges. The introduction of the CARVE task, along with the accompanying datasets and the CERN framework, will advance future research on video causal reasoning and significantly facilitate various applications, including video surveillance, root-cause analysis and movie content management.

Thong Bach, Thanh Nguyen-Tang, Dung Nguyen et al.

Fine-tuning Large Language Models (LLMs) for downstream tasks often compromises safety alignment, even when using parameter-efficient methods like LoRA. In this work, we uncover a notable property: fine-tuned models preserve the geometric structure of their loss landscapes concerning harmful content, regardless of the fine-tuning method employed. This suggests that safety behaviors are not erased but shifted to less influential regions of the parameter space. Building on this insight, we propose a curvature-aware alignment restoration method that leverages influence functions and second-order optimization to selectively increase loss on harmful inputs while preserving task performance. By navigating the shared geometry between base and fine-tuned models, our method discourages unsafe outputs while preserving task-relevant performance, avoiding full reversion and enabling precise, low-impact updates. Extensive evaluations across multiple model families and adversarial settings show that our approach efficiently reduces harmful responses while maintaining or even improving utility and few-shot learning performance.

Duc Kieu, Kien Do, Tuan Hoang et al.

Multi-domain translation (MDT) aims to learn translations between multiple domains, yet existing approaches either require fully aligned tuples or can only handle domain pairs seen in training, limiting their practicality and excluding many cross-domain mappings. We introduce universal MDT (UMDT), a generalization of MDT that seeks to translate between any pair of $K$ domains using only $K-1$ paired datasets with a central domain. To tackle this problem, we propose Diffusion Router (DR), a unified diffusion-based framework that models all central$\leftrightarrow$non-central translations with a single noise predictor conditioned on the source and target domain labels. DR enables indirect non-central translations by routing through the central domain. We further introduce a novel scalable learning strategy with a variational-bound objective and an efficient Tweedie refinement procedure to support direct non-central mappings. Through evaluation on three large-scale UMDT benchmarks, DR achieves state-of-the-art results for both indirect and direct translations, while lowering sampling cost and unlocking novel tasks such as sketch$\leftrightarrow$segmentation. These results establish DR as a scalable and versatile framework for universal translation across multiple domains.

Tuyen Tran, Thao Minh Le, Truyen Tran

Referring-based Video Object Segmentation is a multimodal problem that requires producing fine-grained segmentation results guided by external cues. Traditional approaches to this task typically involve training specialized models, which come with high computational complexity and manual annotation effort. Recent advances in vision-language foundation models open a promising direction toward training-free approaches. Several studies have explored leveraging these general-purpose models for fine-grained segmentation, achieving performance comparable to that of fully supervised, task-specific models. However, existing methods rely on fixed pipelines that lack the flexibility needed to adapt to the dynamic nature of the task. To address this limitation, we propose Multi-Modal Agent, a novel agentic system designed to solve this task in a more flexible and adaptive manner. Specifically, our method leverages the reasoning capabilities of large language models (LLMs) to generate dynamic workflows tailored to each input. This adaptive procedure iteratively interacts with a set of specialized tools designed for low-level tasks across different modalities to identify the target object described by the multimodal cues. Our agentic approach demonstrates clear improvements over prior methods on two multimodal-conditioned VOS tasks: RVOS and Ref-AVS.

Tuyen Tran, Thao Minh Le, Quang-Hung Le et al.

Vision-language alignment in video must address the complexity of language, evolving interacting entities, their action chains, and semantic gaps between language and vision. This work introduces Planner-Refiner, a framework to overcome these challenges. Planner-Refiner bridges the semantic gap by iteratively refining visual elements' space-time representation, guided by language until semantic gaps are minimal. A Planner module schedules language guidance by decomposing complex linguistic prompts into short sentence chains. The Refiner processes each short sentence, a noun-phrase and verb-phrase pair, to direct visual tokens' self-attention across space then time, achieving efficient single-step refinement. A recurrent system chains these steps, maintaining refined visual token representations. The final representation feeds into task-specific heads for alignment generation. We demonstrate Planner-Refiner's effectiveness on two video-language alignment tasks: Referring Video Object Segmentation and Temporal Grounding with varying language complexity. We further introduce a new MeViS-X benchmark to assess models' capability with long queries. Superior performance versus state-of-the-art methods on these benchmarks shows the approach's potential, especially for complex prompts.

Long Hoang Dang, Thao Minh Le, Vuong Le et al.

Video Question Answering (Video QA) is a powerful testbed to develop new AI capabilities. This task necessitates learning to reason about objects, relations, and events across visual and linguistic domains in space-time. High-level reasoning demands lifting from associative visual pattern recognition to symbol-like manipulation over objects, their behavior and interactions. Toward reaching this goal we propose an object-oriented reasoning approach in that video is abstracted as a dynamic stream of interacting objects. At each stage of the video event flow, these objects interact with each other, and their interactions are reasoned about with respect to the query and under the overall context of a video. This mechanism is materialized into a family of general-purpose neural units and their multi-level architecture called Hierarchical Object-oriented Spatio-Temporal Reasoning (HOSTR) networks. This neural model maintains the objects' consistent lifelines in the form of a hierarchically nested spatio-temporal graph. Within this graph, the dynamic interactive object-oriented representations are built up along the video sequence, hierarchically abstracted in a bottom-up manner, and converge toward the key information for the correct answer. The method is evaluated on multiple major Video QA datasets and establishes new state-of-the-arts in these tasks. Analysis into the model's behavior indicates that object-oriented reasoning is a reliable, interpretable and efficient approach to Video QA.

Long Hoang Dang, Thao Minh Le, Vuong Le et al.

Video question answering (Video QA) presents a powerful testbed for human-like intelligent behaviors. The task demands new capabilities to integrate video processing, language understanding, binding abstract linguistic concepts to concrete visual artifacts, and deliberative reasoning over spacetime. Neural networks offer a promising approach to reach this potential through learning from examples rather than handcrafting features and rules. However, neural networks are predominantly feature-based - they map data to unstructured vectorial representation and thus can fall into the trap of exploiting shortcuts through surface statistics instead of true systematic reasoning seen in symbolic systems. To tackle this issue, we advocate for object-centric representation as a basis for constructing spatio-temporal structures from videos, essentially bridging the semantic gap between low-level pattern recognition and high-level symbolic algebra. To this end, we propose a new query-guided representation framework to turn a video into an evolving relational graph of objects, whose features and interactions are dynamically and conditionally inferred. The object lives are then summarized into resumes, lending naturally for deliberative relational reasoning that produces an answer to the query. The framework is evaluated on major Video QA datasets, demonstrating clear benefits of the object-centric approach to video reasoning.

Thao Minh Le, Vuong Le, Svetha Venkatesh et al.

Video QA challenges modelers in multiple fronts. Modeling video necessitates building not only spatio-temporal models for the dynamic visual channel but also multimodal structures for associated information channels such as subtitles or audio. Video QA adds at least two more layers of complexity - selecting relevant content for each channel in the context of the linguistic query, and composing spatio-temporal concepts and relations in response to the query. To address these requirements, we start with two insights: (a) content selection and relation construction can be jointly encapsulated into a conditional computational structure, and (b) video-length structures can be composed hierarchically. For (a) this paper introduces a general-reusable neural unit dubbed Conditional Relation Network (CRN) taking as input a set of tensorial objects and translating into a new set of objects that encode relations of the inputs. The generic design of CRN helps ease the common complex model building process of Video QA by simple block stacking with flexibility in accommodating input modalities and conditioning features across both different domains. As a result, we realize insight (b) by introducing Hierarchical Conditional Relation Networks (HCRN) for Video QA. The HCRN primarily aims at exploiting intrinsic properties of the visual content of a video and its accompanying channels in terms of compositionality, hierarchy, and near and far-term relation. HCRN is then applied for Video QA in two forms, short-form where answers are reasoned solely from the visual content, and long-form where associated information, such as subtitles, presented. Our rigorous evaluations show consistent improvements over SOTAs on well-studied benchmarks including large-scale real-world datasets such as TGIF-QA and TVQA, demonstrating the strong capabilities of our CRN unit and the HCRN for complex domains such as Video QA.

Tri Minh Nguyen, Thin Nguyen, Thao Minh Le et al.

Predicting the interaction between a compound and a target is crucial for rapid drug repurposing. Deep learning has been successfully applied in drug-target affinity (DTA) problem. However, previous deep learning-based methods ignore modeling the direct interactions between drug and protein residues. This would lead to inaccurate learning of target representation which may change due to the drug binding effects. In addition, previous DTA methods learn protein representation solely based on a small number of protein sequences in DTA datasets while neglecting the use of proteins outside of the DTA datasets. We propose GEFA (Graph Early Fusion Affinity), a novel graph-in-graph neural network with attention mechanism to address the changes in target representation because of the binding effects. Specifically, a drug is modeled as a graph of atoms, which then serves as a node in a larger graph of residues-drug complex. The resulting model is an expressive deep nested graph neural network. We also use pre-trained protein representation powered by the recent effort of learning contextualized protein representation. The experiments are conducted under different settings to evaluate scenarios such as novel drugs or targets. The results demonstrate the effectiveness of the pre-trained protein embedding and the advantages our GEFA in modeling the nested graph for drug-target interaction.

Thao Minh Le, Vuong Le, Svetha Venkatesh et al.

We present Language-binding Object Graph Network, the first neural reasoning method with dynamic relational structures across both visual and textual domains with applications in visual question answering. Relaxing the common assumption made by current models that the object predicates pre-exist and stay static, passive to the reasoning process, we propose that these dynamic predicates expand across the domain borders to include pair-wise visual-linguistic object binding. In our method, these contextualized object links are actively found within each recurrent reasoning step without relying on external predicative priors. These dynamic structures reflect the conditional dual-domain object dependency given the evolving context of the reasoning through co-attention. Such discovered dynamic graphs facilitate multi-step knowledge combination and refinements that iteratively deduce the compact representation of the final answer. The effectiveness of this model is demonstrated on image question answering demonstrating favorable performance on major VQA datasets. Our method outperforms other methods in sophisticated question-answering tasks wherein multiple object relations are involved. The graph structure effectively assists the progress of training, and therefore the network learns efficiently compared to other reasoning models.

Thao Minh Le, Vuong Le, Svetha Venkatesh et al.

Video question answering (VideoQA) is challenging as it requires modeling capacity to distill dynamic visual artifacts and distant relations and to associate them with linguistic concepts. We introduce a general-purpose reusable neural unit called Conditional Relation Network (CRN) that serves as a building block to construct more sophisticated structures for representation and reasoning over video. CRN takes as input an array of tensorial objects and a conditioning feature, and computes an array of encoded output objects. Model building becomes a simple exercise of replication, rearrangement and stacking of these reusable units for diverse modalities and contextual information. This design thus supports high-order relational and multi-step reasoning. The resulting architecture for VideoQA is a CRN hierarchy whose branches represent sub-videos or clips, all sharing the same question as the contextual condition. Our evaluations on well-known datasets achieved new SoTA results, demonstrating the impact of building a general-purpose reasoning unit on complex domains such as VideoQA.

Thao Minh Le, Vuong Le, Svetha Venkatesh et al.

What does it take to design a machine that learns to answer natural questions about a video? A Video QA system must simultaneously understand language, represent visual content over space-time, and iteratively transform these representations in response to lingual content in the query, and finally arriving at a sensible answer. While recent advances in lingual and visual question answering have enabled sophisticated representations and neural reasoning mechanisms, major challenges in Video QA remain on dynamic grounding of concepts, relations and actions to support the reasoning process. Inspired by the dual-process account of human reasoning, we design a dual process neural architecture, which is composed of a question-guided video processing module (System 1, fast and reactive) followed by a generic reasoning module (System 2, slow and deliberative). System 1 is a hierarchical model that encodes visual patterns about objects, actions and relations in space-time given the textual cues from the question. The encoded representation is a set of high-level visual features, which are then passed to System 2. Here multi-step inference follows to iteratively chain visual elements as instructed by the textual elements. The system is evaluated on the SVQA (synthetic) and TGIF-QA datasets (real), demonstrating competitive results, with a large margin in the case of multi-step reasoning.

Thao Minh Le, Nobuyuki Shimizu, Takashi Miyazaki et al.

With the widespread use of intelligent systems, such as smart speakers, addressee recognition has become a concern in human-computer interaction, as more and more people expect such systems to understand complicated social scenes, including those outdoors, in cafeterias, and hospitals. Because previous studies typically focused only on pre-specified tasks with limited conversational situations such as controlling smart homes, we created a mock dataset called Addressee Recognition in Visual Scenes with Utterances (ARVSU) that contains a vast body of image variations in visual scenes with an annotated utterance and a corresponding addressee for each scenario. We also propose a multi-modal deep-learning-based model that takes different human cues, specifically eye gazes and transcripts of an utterance corpus, into account to predict the conversational addressee from a specific speaker's view in various real-life conversational scenarios. To the best of our knowledge, we are the first to introduce an end-to-end deep learning model that combines vision and transcripts of utterance for addressee recognition. As a result, our study suggests that future addressee recognition can reach the ability to understand human intention in many social situations previously unexplored, and our modality dataset is a first step in promoting research in this field.

Thao Minh Le, Nakamasa Inoue, Koichi Shinoda

This paper presents a new framework for human action recognition from a 3D skeleton sequence. Previous studies do not fully utilize the temporal relationships between video segments in a human action. Some studies successfully used very deep Convolutional Neural Network (CNN) models but often suffer from the data insufficiency problem. In this study, we first segment a skeleton sequence into distinct temporal segments in order to exploit the correlations between them. The temporal and spatial features of a skeleton sequence are then extracted simultaneously by utilizing a fine-to-coarse (F2C) CNN architecture optimized for human skeleton sequences. We evaluate our proposed method on NTU RGB+D and SBU Kinect Interaction dataset. It achieves 79.6% and 84.6% of accuracies on NTU RGB+D with cross-object and cross-view protocol, respectively, which are almost identical with the state-of-the-art performance. In addition, our method significantly improves the accuracy of the actions in two-person interactions.