Sections

Project List

Mechanistic Interpretability for Factuality: Explore how mechanistic interpretability methods can help explain factual consistency and hallucinations in LLM outputs.
LLMs as Workers in Crowdsourcing: Design experiments to assess LLMs acting as crowd workers, focusing on alignment with human judgments and task diversity.
Multidimensional Truthfulness Assessment with LLMs: Design and evaluate methods for assessing truthfulness along multiple dimensions (e.g., factuality, intent, harm) using Large Language Models.
Understanding Tool Invocation in Tool-Augmented LLMs: Investigate how LLMs decide which external tool or agent to invoke using counterfactual perturbations.
RAG vs. Recall in Fact-Checking: Analyze when fact-checking LLMs rely more on retrieved evidence versus pretrained knowledge to assess the role and effect of evidence
LLMs in robotics for autonomous planning, trajectory optimization, efficient locomotion, and generalization using reinforcement learning (RL). Key areas include language-guided decision-making, hierarchical planning, policy generalization, and adaptive control to enhance robotic autonomy and efficiency. See for example https://deepmind.google/discover/blog/shaping-the-future-of-advanced-robotics/
Fine-tune LLMs for autonomous crowdsourcing tasks, enabling AI agents to independently perform data collection and complex task execution on crowdsourcing platforms, with a focus on improving collaboration and efficiency in data acquisition workflows.
Develop real-time fact-checking systems for video content, leveraging Whisper for speech-to-text transcription and LLMs augmented with Retrieval-Augmented Generation (RAG) techniques for automatic fact verification and misinformation detection.
Use OpenAI Gymnasium to design environments for studying LLM behavior and decision-making in planning tasks, analyzing strategies and performance. cfr. https://dl.acm.org/doi/abs/10.1145/3586183.3606763.
Adapt AlphaGo Zero-style models to tackle LLM challenges, focusing on improving complex reasoning paths and automatic prompt optimization using self-play and reinforcement learning. cfr https://arxiv.org/pdf/1712.01815, https://www.nature.com/articles/nature16961, https://www.nature.com/articles/s41586-021-03819-2.