Research
Sycophancy Measurement
📏
SPINE: Measuring LLM Sycophancy under Sustained Multi-Turn Pressure
preprint on arXiv · under review at ACL Rolling Review
Do LLMs abandon correct positions under sustained pressure? An adaptive LLM user-proxy challenges models for up to 25 turns — and collapse rates keep climbing long after short benchmarks stop counting. With collaborators.
arXiv
preprint on arXiv · under review at ACL Rolling Review
Do LLMs abandon correct positions under sustained pressure? An adaptive LLM user-proxy challenges models for up to 25 turns — and collapse rates keep climbing long after short benchmarks stop counting. With collaborators.
arXiv
Sycophancy Mitigation
🛡️
Mitigating Sycophancy and Improving Reasoning in LLMs via Exploratory Thinking and RL
ongoing research
We're exploring different methods to keep models from caving under sustained pushback. A simple persona prompt already halves the SPINE collapse rate at turn 25. Ongoing.
ongoing research
We're exploring different methods to keep models from caving under sustained pushback. A simple persona prompt already halves the SPINE collapse rate at turn 25. Ongoing.
Multi-Agent Simulation & Synthetic Data
👥
Multi-Agent Systems with LLMs for Cyberbullying Intervention Training and Safe Social Simulation
ongoing research
A multi-agent LLM simulator where adolescents practice responding to cyberbullying alongside AI agents, with message-level safety review and severity-capped agent behavior keeping the chat safe.
ongoing research
A multi-agent LLM simulator where adolescents practice responding to cyberbullying alongside AI agents, with message-level safety review and severity-capped agent behavior keeping the chat safe.