ACL 2026 paper creates AI that questions like Supreme Court justices
New AI learns to ask probing legal questions, outperforming baselines on Supreme Court data.
Most conversational AI systems are passive: they wait for user queries and respond politely. But in high-stakes settings like a courtroom, a medical diagnosis, or an investigative interview, the AI must actively probe and steer the conversation to uncover hidden facts. A new paper from researchers at Georgetown University and collaborating institutions introduces the concept of Inquisitive Conversational Agents (ICAs) and demonstrates a working prototype fine-tuned on U.S. Supreme Court oral arguments.
The core innovation is a Dual Hierarchical Reinforcement Learning framework. Two reinforcement learning agents operate at different levels: one agent decides the high-level dialogue strategy (e.g., whether to challenge a statement, ask for clarification, or move to a new topic), while a second agent generates the specific utterances to execute that strategy. This hierarchy allows the system to mimic the structured yet flexible questioning patterns of Supreme Court justices. Evaluated on a dataset of real oral arguments, the model outperformed standard baselines (including vanilla RL and supervised approaches) on measures of fact coverage, coherence, and strategic goal achievement. Accepted as a Findings paper at ACL 2026, this work opens the door for AI that doesn’t just answer questions—it asks the right ones.
- Introduces Inquisitive Conversational Agents (ICAs) that proactively extract information, unlike passive chatbots.
- Uses a Dual Hierarchical Reinforcement Learning framework with two cooperating RL agents for strategy and utterance generation.
- Trained and evaluated on U.S. Supreme Court oral arguments dataset, outperforming multiple baselines across several metrics.
Why It Matters
Proactive AI interrogators could revolutionize legal, medical, and investigative domains by systematically extracting critical hidden information.