Trung Bui

Trung Bui

Adobe Systems (United States)
5 previous affiliations
80Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Trung Bui focuses on natural language processing, conversational AI, and multimodal machine learning. His trajectory spans early work in dialogue systems and question answering to recent advancements in the efficiency, personalization, and evaluation of large language models. Currently, he investigates retrieval-augmented generation and the internal steering of mixture-of-experts architectures.

Research Trajectory

Bui’s early research established a foundation in multimodal search and natural language understanding. In 2017, he collaborated on AMC: Attention Guided Multi-modal Correlation Learning for Image Search, which utilized attention mechanisms to balance visual and textual modalities based on user query intent. He subsequently synthesized developments in text-based question answering in A Review on Deep Learning Techniques Applied to Answer Selection (2018).

A significant portion of his subsequent work focused on bridging natural language and image editing. He guided the development of A System for Automated Image Editing from Natural Language Commands (2018), which translated user requests into discrete executable actions. He later expanded this paradigm into a robust interactive framework with A Multimodal Dialogue System for Conversational Image Editing (2020). By modeling the interaction as a Partially Observed Markov Decision Process trained via reinforcement learning, the system successfully managed noisy speech and gesture inputs. Concurrently, he addressed challenges in e-commerce question answering with Supervised Transfer Learning for Product Information Question Answering (2019), demonstrating that noisy community data could be leveraged to improve performance on specialized product platforms.

Moving into the early 2020s, Bui’s focus broadened to the evaluation and interpretability of neural models. He investigated the research direction of Word Order Invariance in Neural Natural Language Understanding in the paper Out of Order: How important is the sequential order of words in a sentence in Natural Language Understanding tasks? (2021). This work revealed that models like BERT often rely on superficial cues rather than deep semantic comprehension, maintaining high performance even when sentence word order is randomized. He also proposed a streamlined approach to entity tracking in End-To-End Neural Coreference Resolution Revisited: A Simple Yet Effective Baseline (2022). In the domain of model explainability, he examined the limitations of input marginalization in Double Trouble: How to not Explain a Text Classifier's Decisions Using Counterfactuals Synthesized by Masked Language Models? (2022), showing that generating multiple counterfactuals is necessary to improve explanation reliability. This interest in interpretability extended to computer vision; he participated in research contributing to the broader area of Interpretable Deep Learning via Concept Bottleneck Models with PEEB: Part-based Image Classifiers with an Explainable and Editable Language Bottleneck (2024), which replaced opaque class labels with editable natural language part descriptors.

In recent years, Bui has directed research toward complex video understanding and specialized generative models. He oversaw work on Moment detection in long tutorial videos (2024), addressing the challenge of localizing specific events in hours-long educational content. He also led the development of Retrieval Augmented Generation for Domain-specific Question Answering (2024), designing a specialized retrieval and fine-tuning pipeline to reduce hallucinations and ground large language models in domain-specific contexts.

His most recent collaborative work addresses the frontiers of multimodal personalization, model efficiency, and continual learning. He explored the research area of Instance Level Personalization for Vision Language Models through Yo'Chameleon: Personalized Vision and Language Generation (2025), adapting soft-prompt tuning to embed subject-specific visual knowledge into large multimodal models. To address computational bottlenecks, he contributed to Lizard: An Efficient Linearization Framework for Large Language Models (2025), a framework that converts heavy Transformer architectures into sub-quadratic models for constant-memory inference. Finally, he co-authored Steering MoE LLMs via Expert (De)Activation (2026), which demonstrated how controlling expert activation at inference time can bypass or enforce safety guardrails, and helped introduce the OAKS benchmark in Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams (2026) to evaluate how well models track dynamically changing real-world facts.

Current Directions

The most recent papers in the evidence indicate a strong focus on the operational realities of deploying large language and multimodal models. Recent work investigates the efficiency of these architectures, specifically through linearization techniques that enable constant-memory inference without sacrificing reasoning capabilities. Additionally, Bui is actively exploring the dynamic behavior of Mixture-of-Experts (MoE) models, focusing on how inference-time expert activation can be used to steer model safety and factuality. Another major theme in his current research is continual adaptation and personalization, addressing how models can integrate subject-specific visual concepts and keep pace with real-time, evolving knowledge streams.

Profile updated

See something wrong? Report a mistake