Research Themes
Safety
A model that behaves safely in evaluation can stop behaving safely once it is fine-tuned, narrowed to one job, or pushed by an adversary. We study whether safety survives those conditions: whether a task-specific agent refuses what falls outside its remit, whether alignment persists after adaptation, and how it can be restored or steered at test time without giving up capability.
OffTopicEval
An evaluation suite measuring whether LLM agents accept valid in-domain requests and refuse out-of-domain ones.
Safety Arithmetic
A test-time framework for steering language models toward safer behavior through parameters and activations.
Chain of Utterances
RED-EVAL and Chain-of-Utterances prompting for probing harmful behavior and studying safety alignment.
Gender Bias in BERT
A highly cited analysis of gender bias encoded in contextualized language representations.
Trustworthiness
A fluent answer and a trustworthy one are not the same thing. We work on systems that ground what they say in retrieved evidence, attribute it honestly, decline when the evidence will not support an answer, and — where several agents cooperate — calibrate how much they should rely on one another.
Epistemic Context Learning
Trust formation and calibrated reliance in LLM-based multi-agent systems.
Multimodality
Meaning arrives through words, voice, face, and scene at once, and a model that reads only one channel misses most of it. We study how to fuse those channels — separating what they share from what belongs to each — and how to generate in them, from emotion in multi-party conversation to text-to-audio synthesis.
MELD and DialogueRNN
Highly cited resources and models for multimodal, multi-party emotion recognition in conversations.
TFN, MISA, and Multimodal-Infomax
Foundational work on modality interaction, invariant and specific representations, and robust multimodal sentiment analysis.
AI for Science
We ask whether language models can do more than summarize the literature — whether they can propose hypotheses a scientist would think worth testing. Measuring that claim honestly is half the problem, so we build benchmarks that check whether a model rediscovers findings deliberately held out of its training data.
MOOSE-Chem
A benchmark and framework testing whether LLMs can rediscover valid chemistry hypotheses from background literature.
Open-Domain Scientific Hypotheses
Methods for automated scientific hypothesis discovery across broad literature collections.
Efficiency
Capability that is too expensive to train, serve, or remember does not reach anyone. We work on getting more from a fixed budget: compact online memory instead of ever-longer context, learned data selection instead of simply more data, merging instead of retraining, and exact linear-time attention for long inputs.
DELLA-Merging
Magnitude-based sampling to reduce interference when merging task-specialized language models.
EFLA
Error-Free Linear Attention derives an exact continuous-time solution for robust long-context computation.
PromptDistill, LLM-Adapters, UDApter
Efficient inference, parameter-efficient fine-tuning, and adapter-based transfer for language and speech models.
Embodied AI
Language and vision only matter to a robot if they end in the right action. We build compact vision-language-action models that ground perception in behavior, and evaluations that test the whole path — from what the agent sees, through what it decides, to what it actually does.