LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Researchers introduced LedgerMind, a multimodal agent framework built around a structured evidence ledger.

It constrains reasoning claims to active evidence and uses typed state transitions to repair errors.

The system combines a grounding protocol, an adaptive dispatcher, and a verification-and-repair engine.

Results on multimodal benchmarks show gains in answer accuracy and trajectory-level faithfulness.