Associate Professor · EEE · NTU Singapore
নমস্কার I am Soujanya Poria সৌজন্য পড়িয়া
I lead DeCLaRe Lab at NTU Singapore.
Research themes
Our lab works across six research themes: Safety, Trustworthiness, Multimodality, AI for Science, Efficiency, and Embodied AI. We pursue this work at the DeCLaRe Lab, NTU Singapore.
Work with our lab
Prospective students and collaborators can explore our research slides and recent publications to find questions that connect with their interests.
Research slidesRecent updates
Highly Cited Researcher
Recognized by Web of Science.
VentureBeat features δ-mem
An interview on working memory for long-running AI agents.
Healthcare AI Symposium
Invited master class in Singapore.
Joined EEE, NTU
Moved to the School of Electrical and Electronic Engineering.
DeCLaRe Lab
Much of this work grows out of the DeCLaRe Lab. Founded at SUTD in 2019, the lab has been based at NTU since 2025.
Selected contributions
Selected recent and foundational work.
Complete publication recordDelta-Mem
Online memory for language-model agents, enabling experience to shape future behavior without retraining the base model.
Data-Agent
Reframes data selection as an adaptive process in which models learn which training examples they need.
Epistemic Context Learning
Gives language-model agents explicit epistemic context for deciding what to trust in multi-agent communication.
TangoFlux
Fast, high-fidelity text-to-audio generation through flow matching and diffusion-transformer modeling.
OffTopicEval
Tests whether language models can recognize when a request falls outside the active conversational context.
Trustworthy RAG
Studies when retrieval strengthens language models and when conflicting evidence makes their answers less reliable.
MOOSE-Chem
Evaluates whether language models can rediscover previously unseen chemistry hypotheses from scientific evidence.
Error-Free Linear Attention
Derives exact linear-attention dynamics that improve efficiency without introducing approximation error.
NORA-1.5
Trains a generalist vision-language-action model with preference rewards grounded in predicted worlds and executed actions.
Chain-of-Knowledge
Grounds language models by adapting knowledge dynamically across heterogeneous sources during reasoning.
Chain of Utterances
Introduces multi-turn adversarial conversations for red-teaming the safety alignment of language models.
Gender Bias in BERT
Provides a systematic account of gender bias encoded by contextual language representations.
MISA
Separates modality-specific and modality-invariant representations for robust multimodal learning.
COSMIC
Integrates commonsense knowledge into contextual emotion recognition for natural conversations.
DialogueGCN
Represents speakers and conversational dependencies as a graph for emotion recognition in dialogue.
DialogueRNN
Models conversational context and speaker state for emotion recognition in multi-party dialogue.
MELD
Established a multimodal, multi-party benchmark for emotion recognition in natural conversations.
Tensor Fusion Network
Introduced explicit tensor-based interactions among language, vision, and acoustic representations.