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

TL;DR - LedgerMind structures multimodal-agent trajectories as provenance-constrained evidence ledgers, making intermediate reasoning auditable rather than evaluating only final answers. It improves answer accuracy and trajectory-level faithfulness across multiple benchmarks and MLLM backbones.

• Tool outputs become normalized ledger entries that downstream claims must cite.
• Entity- and numeric-level checks detect unsupported reasoning and “Phantom Grounding.”
• Typed repair transitions prevent verification steps from introducing unproven content.
• Adaptive routing adjusts reasoning depth to query complexity, reducing over-reasoning.