A preprint describes an AI method that reduced the source material needed to check a large language model answer to around 9.8% of the full text. It also reported an accuracy score of 1, more than 30% above the best baseline at a comparable provenance size.
The method, called BLIP, focuses on provenance, the part of the source record that can be used to verify an answer. It is designed to return a minimal verifiable provenance, with an extension for finding multiple distinct minimal provenances.
A shorter trail, with measured trade-offs
In the main evaluation, BLIP's LLM-adaptive strategy kept provenance to around 9.8% of the full text. Its reported cost ratio was 0.2 times, with average latency of 2.1 seconds.
A refinement step reduced the ratios further in three named workloads: about 3.7% for Qasper, 4.7% for NL_DEV and 7.7% for HotpotQA. The Exponential-Greedy strategy had a cost ratio of up to 1.1 times and average latency of 8.1 seconds.
Across the five workloads, BLIP required 3% to 12% of the document under the study's human review-effort measure.
The study also reported two patterns called strong and weak monotonicity, which describe consistency in the provenance search as the available source material changes. Strong monotonicity averaged around 90% across real workloads with multiple LLMs, while weak monotonicity was reported at a rate just over 95%.
The clearest test used synthetic data
The evaluation used five question-answering workloads, sampling 500 distinct question-document pairs from each. It also included two synthetic table-question-answering workloads, Movie and Restaurant, each with 64,000 tokens and 100 questions.
Those synthetic tests offered the clearest check because the true provenance was known. BLIP recovered a superset of the true provenance in all TableQA cases, and it exactly recovered the provenance in over 98% of Movie questions and 96% of Restaurant questions.
The RAG comparisons were weaker on exact recovery or size. RAG-EXACT recovered the true provenance in only 32% to 68% of cases, while RAG-10% could produce a provenance more than 50 times larger than ground truth.
What the score does not mean
The accuracy result is about answer reproduction, not factual truth. The test asks whether the selected provenance reproduces the LLM's full-text answer, so a score of 1 does not establish that the original answer was correct.
The exact-recovery figures apply only to the synthetic Movie and Restaurant workloads. True provenance was not available for the five plain-text workloads, and the paper notes that its multiple-provenance procedure does not guarantee finding every distinct minimal provenance.
A small human check
A user study involved eight participants and 12 scenarios. The reported overall usefulness rating for BLIP provenance was 4.75 out of 5 for judging answer correctness.
That rating is exploratory because it comes from a small study, so it is an early signal about usefulness rather than broad evidence about how people would use the method.
Preprint status and access
The work is a preprint, with an arXiv version 2 dated 28 Aug 2026.
Its front matter states that the source code, data and other artifacts have been made available at the BLIP GitHub repository.
The acknowledgments list National Science Foundation grants, State of California and Alfred P. Sloan Foundation funds, EPIC lab sponsors, and compute credits from Azure, Modal, NSF via NAIRR and OpenAI.
Paper data and sources
Original title: Bolt-on, Verifiable Provenance for LLM-Powered Data Processing
Authors: Yiming Lin, Sepanta Zeighami, Aditya G. Parameswaran
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text