Deep Cognition and Language Research Lab
Research in Safe, Trustworthy, Multimodal AI
DeCLaRe studies how intelligent systems reason, perceive, communicate, and act across language, vision, audio, scientific knowledge, and embodied environments.
About DeCLaRe
DeCLaRe, short for Deep Cognition and Language Research, was established by Soujanya Poria at the Singapore University of Technology and Design in 2019 with founding members Navonil Majumder, Devamanyu Hazarika, and Deepanway Ghosal. The lab moved to Nanyang Technological University in 2025.
Our research develops grounded, interpretable, and robust AI systems for settings in which benchmark accuracy alone is insufficient.
The robot recalls early computing while forming 宣 (xuān), “to declare”: an intelligent social agent connecting computing’s history with collaborative AI.
Research Themes
What each direction covers, and the work that anchors it.
Safety
Operational safety, red-teaming, refusal behavior, and alignment interventions.
Trustworthiness
Grounded attribution, reliable retrieval, uncertainty, and citation-aware responses.
Multimodality
Language, vision, audio, and video models for reasoning, generation, and interaction.
AI for Science
Scientific hypothesis discovery, chemistry, and literature-grounded reasoning.
Efficiency
Data selection, adapters, model merging, memory, and compact model training.
Embodied AI
Vision-language-action systems, action grounding, and interactive evaluation.
Hot Papers 🔥
Four current anchors across the lab's research program.
δ-mem: Efficient Online Memory for Large Language Models
A compact online state gives language models working memory without extending the visible context.
OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always!
An operational-safety benchmark for whether task-specific agents respect their intended boundaries.
Data Agent: Learning to Select Data via End-to-End Dynamic Optimization
A training-aware policy selects useful data as the target model evolves.
Measuring and Enhancing Trustworthiness of LLMs in RAG
Trust-Score and Trust-Align evaluate and improve grounded attribution, citations, and refusal.
Funded Research Directions
Active support spans embodied foundation models, trustworthy language generation, efficient architectures, and multimodal AI.
Explore active and completed projects, funding periods, agencies, and the research directions they support.