New framework forces AI vision agents to show their work before an answer counts as correct

8:00 PM ET · July 31, 2026

A team of seven researchers, Enjun Du, Hange Zhou, Chenxu Du, Siyi Liu, Zirong Chen, Ziyu Zheng, and Yongqi Zhang, published LedgerMind to address a gap in how multimodal AI agents get evaluated: standard benchmarks score only whether a visual question answering agent’s final answer is correct, which cannot tell apart an agent that reasoned correctly from one that guessed right, absorbed a language bias, or had two separate errors cancel each other out. LedgerMind instead normalizes every tool output, such as an image crop, a retrieved document, or a calculation, into a Structured Evidence Ledger that serves as the agent’s working state, and requires that any downstream reasoning claim cite only entries actually present in that ledger, with a Three-Layer Grounding Protocol checking claims at the entity and numeric level. An Adaptive Dual-Path Dispatcher routes simple questions through a lighter reasoning path and harder ones through deeper verification, while an Event-Triggered Verification-and-Repair engine catches and fixes ungrounded claims through typed state transitions that come with a formal guarantee against introducing new, unverified content during repair. The authors say the design specifically targets four recurring failure patterns, including what they call Phantom Grounding, where an agent cites what looks like a legitimate source for a claim the source does not actually support, plus over-reasoning on simple queries and errors introduced during the repair step itself. Across the benchmarks tested, the authors report LedgerMind improves both final answer accuracy and trajectory level faithfulness compared with baseline multimodal models.