About this explorer

Browse the questions, generations and evaluations from Clarify, Abstain or Answer? Strategising in Conversation with Belief-Augmented Generation (INLG 2026).

Each record starts from an AmbigQA question. The assistant samples K answers (its belief state), then picks a strategy: answer, ask a clarification question, or abstain. SAG picks from the question alone; BAG1–3 also see the belief state. A simulated user, who has one hidden AmbigQA interpretation in mind, answers any clarification question, and the assistant answers again in turn 4 (forced answer, or SAG+/BAG+, which may still abstain). An LLM judge scores answers against the reference for the user's intended interpretation.

The coloured claim clusters and per-sample ✓/✗ marks in the belief state are an evaluation overlay: the judge model groups the samples after the fact for the faithfulness analysis, and BAG never sees them. Turn them off with the Judge clusters switch. BAG3 also clusters its own samples as part of its output; you can see that grouping in its raw output in turn 2.

Use the list on the left to filter records, the method row to compare strategies on the same question, and ← → (or j/k) to step through records.