LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger
LedgerMind forces each reasoning step to stay tied to tool-produced evidence, not just a plausible final answer.
The paper frames multimodal agent reasoning as a provenance-constrained state machine built around a Structured Evidence Ledger. Claims downstream can cite only active ledger entries, with grounding checks at the entity and numeric level. Its repair mechanism is limited to typed state transitions, so fixes cannot add unsupported content. The authors report gains in both answer accuracy and trajectory-level faithfulness across multiple multimodal reasoning benchmarks and backbone MLLMs. HF Daily Papers' note
The paper frames multimodal agent reasoning as a provenance-constrained state machine built around a Structured Evidence Ledger. Claims downstream can cite only active ledger entries, with grounding checks at the entity and numeric level. Its repair mechanism is limited to typed state transitions, so fixes cannot add unsupported content. The authors report gains in both answer accuracy and trajectory-level faithfulness across multiple multimodal reasoning benchmarks and backbone MLLMs. HF Daily Papers' note
score 4