When Is Enough Not Enough? Illusory Completion in Search Agents
Search agents often fail multi-constraint problems by engaging in illusory completion—believing tasks are done despite unresolved constraints. Researchers introduced the EPISTEMIC LEDGER to track evidential support and agent beliefs for each constraint during reasoning. This revealed four failure patterns, and a mitigation strategy, LIVELEDGER (an inference-time tracker), significantly improved performance by explicitly checking constraint states.