Researchers have released a training-free technique for tracing an AI system's answers back to specific evidence in long, multimodal documents.
The method, called MultAttnAttrib, works during a model's prefill pass, tapping selected attention heads and calibrated thresholds to locate source evidence without any additional training. The team also released MultAttrEval, a benchmark dataset they say is the first designed specifically for evaluating multimodal attribution in long-form documents. In tests, MultAttnAttrib outperformed several prompting-based attribution approaches and delivered attributions at roughly one-seventh the latency of direct inference on the same base model.
Attribution — knowing exactly which sentence, table, or image in a source document generated a given answer — is the unglamorous plumbing that makes AI assistants actually auditable. Most prior work addressed text-only settings; multimodal documents mixing text, tables, and images have been largely ignored, which matters as enterprises push AI into contract review, financial filings, and technical manuals. A lighter, faster method that works without retraining could lower the barrier to deploying verifiable grounded QA at scale.
The authors claim competitive accuracy with unnamed frontier models, but that comparison rests on an unverified benchmark the same team built — worth noting before treating it as a settled leaderboard result.