Interactive retrieval under partial evidence is a sequential information‑acquisition problem: an agent must decide which question will generate the most useful evidence for the next retrieval update. Existing systems train this decision by imitating an offline ordering of candidate QA pairs, even though a question’s value is determined by the response it elicits and its downstream effect on retrieval. We first show that candidate discriminativeness and perceived usefulness serve as weak supervision for this objective. Then we introduce RAVEL (Retrieval‑Aware online reinforcement Learning), an online RL framework for interactive person re‑identification. RAVEL initializes from supervised question generation, directly observes the current Top‑4 candidates, and optimizes the question policy using rank feedback from the full question‑answer‑retrieval loop. Experiments on the Interactive‑PEDES benchmark demonstrate that RAVEL delivers progressively stronger retrieval performance across five interaction rounds. Further analysis reveals that RAVEL reallocates the questioning budget toward localized open‑ended attributes, which provide more useful retrieval evidence and yield the largest gains on initially difficult queries.
Review: By feeding retrieval feedback straight into the reinforcement‑learning loop, RAVEL overcomes the limitations of offline imitation and yields a more adaptive questioning strategy. Its focus on open‑ended attributes especially boosts performance on hard queries, opening new avenues for interactive retrieval research.