Runner-Up Veto Replay for Evidence-Based Stopping in Web Search Agents
Year: 2026
Type: article
Abstract: For a search agent asked when a building opened, removing a citation may leave the favorite year in first place while taking away the only reason to dismiss a rival year. The missing page might record that construction ended that year and the doors were still closed. Without that page the ranking alone tells us little about whether the remaining passages distinguish the two dates. Runner-Up Veto Replay uses this distinction to decide whether a search agent should retrieve again or release its answer. Each retrieval updates the ordering of the answers under consideration. For each challenger to the current favorite, the agent records a claim on which their order could turn and quotes the strongest passages for and against that claim from a compact reading set. The challenger stays in contention unless a passage contradicts something its answer entails and the opposing quotation fails to reconcile the conflict. Once every challenger has a documented objection and the favorite has direct textual support, the reader repeats its judgments with each source hidden in turn. If a challenger lacks a documented objection after a source is hidden, the agent continues searching. On a fixed collection and on the open web, we would hold the reader and reading-set size constant and compare questions with similar leading margins. Would the veto prevent premature stopping when a margin test or a check on the winner after masking permits the agent to answer? We would measure answer correctness, retrieval calls and the total cost of reading and search.
Source: Figshare
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7226023091
