Each record shows one question and its reference(s) from AmbigQA (📚), the assistant LLM's (🤖) belief state (K sampled candidate answers), its augmentation output (a strategy, reasoning trace, and response), the resulting potentially multi-turn conversation with a simulated user (👤), reference-based LLM-judge verdicts (⚖️), and direct generation baselines.
Pick an assistant model, augmentation method (SAG is a prompt-only method, BAG1-3 exploit the belief state), optionally again for the second assistant turn (turn 4) after clarification. You can also pick a brevity-inducing prompt to shorten generations (and thus also the belief state)