Research Themes
Each theme combines recent work with long-running foundations from the group, from multimodal emotion and sentiment analysis to retrieval-grounded generation, scientific discovery, efficient learning, and vision-language-action models.
Safety
We study whether AI systems behave safely inside the situations they are built for, especially when LLM agents face off-topic, adversarial, or underspecified requests.
- Operational safety for task-specific LLM agents
- Red-teaming, jailbreak analysis, and refusal behavior
- Test-time alignment and model steering for safer responses
OffTopicEval
An evaluation suite for measuring whether LLM agents accept valid in-domain requests and refuse out-of-domain ones.
Team: Jingdi Lei, Varun Gumma, Rishabh Bhardwaj, Seok Min Lim, Chuan Li, Amir Zadeh, Soujanya Poria
Safety Arithmetic
A test-time framework for steering language models toward safer behavior through parameters and activations.
Team: Safety and alignment group
Chain of Utterances
RED-EVAL and Chain-of-Utterances prompting for probing harmful behavior and studying safety alignment in LLMs.
Team: Safety and dialogue systems group
Gender Bias in BERT
A highly cited analysis of gender bias encoded in contextualized language representations.
Team: Bias and responsible AI collaborators
Trustworthiness
We develop methods for AI systems that know when to answer, when to cite, when to refuse, and how to communicate uncertainty in grounded settings.
- Trustworthy retrieval-augmented generation
- Grounded attribution and citation-aware generation
- Trust calibration in multi-agent and human-facing systems
Epistemic Context Learning
Studying trust formation and calibrated reliance in LLM-based multi-agent systems.
Team: Trustworthy and interactive AI group
Multimodality
Multimodality is a long-running foundation of DeCLaRe Lab: we build models and benchmarks that integrate language, vision, audio, video, and social context.
- Multimodal fusion and representation learning
- Emotion, sentiment, sarcasm, and social signal understanding
- Vision-language and audio-language reasoning benchmarks
MELD and DialogueRNN
Highly cited resources and models for multimodal, multi-party emotion recognition in conversations.
Team: Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, collaborators
AI for Science
We explore how language models and agentic systems can help scientific reasoning: retrieving inspirations, forming hypotheses, ranking ideas, and synthesizing evidence.
- Scientific hypothesis discovery and rediscovery
- Chemistry-focused benchmarks and multi-agent discovery pipelines
- Open-domain scientific literature reasoning
MOOSE-Chem
An ICLR 2025 benchmark and framework for testing whether LLMs can rediscover valid chemistry hypotheses from background questions and literature.
Team: Zonglin Yang, Wanhao Liu, Ben Gao, Tong Xie, Yuqiang Li, Wanli Ouyang, Soujanya Poria, Erik Cambria, Dongzhan Zhou
Open-Domain Scientific Hypotheses
LLM methods for automated scientific hypothesis discovery across broad literature collections.
Team: Scientific reasoning group
Efficiency
We build techniques that make training, adaptation, and inference cheaper without losing reliability or downstream performance.
- Dynamic data selection and data-efficient training
- Online memory, adapters, distillation, token retention, and long-context methods
- Compact multimodal and embodied models for practical deployment
δ-mem
A lightweight online memory mechanism that compresses history into a compact state and uses it to modulate frozen Transformer attention.
Team: Jingdi Lei, Di Zhang, Junxian Li, Weida Wang, Kaixuan Fan, Xiang Liu, Qihan Liu, Xiaoteng Ma, Baian Chen, Soujanya Poria
DELLA-Merging
Magnitude-based sampling for reducing interference when merging task-specialized language models.
Team: Prateek Yadav Deep, Rishabh Bhardwaj, Soujanya Poria
EFLA
Error-Free Linear Attention derives an exact continuous-time solution for robust long-context computation.
Team: Jingdi Lei, Dong Zhang, Soujanya Poria
PromptDistill, LLM-Adapters, UDApter
Efficient inference, parameter-efficient fine-tuning, and adapter-based transfer for language and speech models.
Team: Efficient learning group
Embodied AI
We develop embodied agents that perceive, reason, and act, with emphasis on compact VLA models, action grounding, and interactive benchmarks.
- Vision-language-action models for robotic tasks
- World-model and action-based preference rewards
- Interactive reasoning benchmarks from perception to action
