Embedded Files
Hello, I'm Nunu,
Hello, I'm Nunu,
AI Research Scientist working on LLM post-training for faithful, grounded generation. I lead a research line on hallucination detection and mitigation across the training lifecycle: preference optimization, RL post-training, detector-based rewards, and evaluation. My background is in reinforcement learning, uncertainty and evaluation for dialogue systems. I teach graduate courses on reinforcement learning and dialogue systems at Heinrich Heine University Düsseldorf. My goal is to bridge the gap between cutting-edge NLP research and safe, beneficial, and reliable AI technologies.
Current Research
Current Research
Language models often produce fluent text that isn’t supported by their input. I study how post-training causes, detects and reduces these hallucinations. Since 2025 I have led a research line of ten projects that I designed and supervise, using models of up to 8B parameters.
What we work on
Preference optimization: DPO with preference data labelled for faithfulness. It reduced hallucination by about 40% according to an independent LLM judge, outperformed standard ways of building preference pairs by an order of magnitude, and transferred to a second generation task.
RL post-training: GRPO and self-distillation on 7B models, including an analysis of why sparse rewards and fluency collapse cause these methods to fail.
Detection as reward: hallucination detectors developed into reward models for RL.
Evaluation: comparing rule-based, learned and LLM-judge detectors, which can disagree enough to reverse which method looks best.
What we’ve learned so far
The quality of the preference signal matters more than any other factor.
Preference data contains shortcuts, such as length, that DPO exploits. Sometimes this makes hallucination worse.
Pushing for faithfulness often makes models say less, so hallucination rates need to be reported together with coverage.
A result that only one evaluator supports isn’t reliable.
What's next
Most hallucination research, including the results above, measures models after training is finished. I am increasingly interested in how hallucination forms during training: when it emerges, which kinds of hallucination further training resolves and which persist, and whether a model’s own confidence can signal its faithfulness. I am now bringing this diagnostic view into post-training, tracking how preference optimization and RL reshape hallucination and calibration as they happen, and designing rewards that target confident hallucinations, so that models stay faithful to their evidence and say when they don’t know.
News
News
October 2026
I am holding the Implementing Transformers Practical course and co-holding the Reinforcement Learning course for the Master programme AI and Data Science at HHU.
I received an Outstanding Reviewer Award from EMNLP 2026.
April 2026
I am holding the Spoken Dialogue System course for the Master programme AI and Data Science at HHU.
January 2026
I reached over 1000 citations in Google Scholar!
October 2025
Our paper, "Less is More: Local Intrinsic Dimensions of Contextual Language Models", has been accepted for presentation at NeurIPS 2025!
I am co-holding the Reinforcement Learning course for the Master programme AI and Data Science at HHU.
July 2025
I am giving an invited tech talk + demo on ConvLab3 at SIM4IA Workshop.
April 2025
I am holding the Spoken Dialogue System course for the Master programme AI and Data Science at HHU.
December 2024
Our paper "Learning from Noisy Labels via Self-Taught On-the-Fly Meta Loss Rescaling" is accepted at AAAI 2025!
October 2024
I am co-holding the Reinforcement Learning course for the Master programme AI and Data Science at HHU.
July 2024
Our paper "Affect Recognition in Conversations Using Large Language Models" is accepted at SIGDIAL 2024!
Our paper "Infusing Emotions into Task-oriented Dialogue Systems: Understanding, Management, and Generation" is accepted at SIGDIAL 2024!
July 2023
I gave an invited talk titled "Dialogue Evaluation via Offline Reinforcement Learning and Emotion Prediction" at NLP for Conversational AI 5 co-located at ACL 2023 in Toronto, Canada.
We presented two posters at ACL 2023 in Toronto, Canada.
Our paper "From Chatter to Matter: Addressing Critical Steps of Emotion Recognition Learning in Task-oriented Dialogue" is accepted at SIGDIAL 2023!
June 2023
I gave a keynote titled "Dialogue Evaluation via Offline Reinforcement Learning and Emotion Prediction" at the 32nd TaCoS Conference in Düsseldorf, Germany.
May 2023
Our paper "ChatGPT for Zero-shot Dialogue State Tracking: A Solution or an Opportunity?" is accepted at ACL 2023!
December 2022
Online Talks
Online Talks
Nunu_SIGDIAL2022_v2.mp4SIGDIAL 2022
SIGDIAL 2022
Dialogue Evaluation with Offline RL
LAVA.mp4COLING 2020
COLING 2020
LAVA: Latent action spaces via variational auto-encoding
Get in touch at lubis [at] hhu [dot] de
Get in touch at lubis [at] hhu [dot] de
Universitätsstr. 1 Gebäude 25.12.01.34, Düsseldorf, Germany 40225
Page updated
Google Sites
Report abuse