An AI tool described in a new preprint can find a smaller piece of the original text that makes a black-box language model produce the same answer it gave from the full text. The system, called BLIP, was reported with a verifiability accuracy of 1 and more than 30% higher than the best baseline of comparable provenance size. The measure is about reproducing an answer, not establishing that the answer is correct.
That is a narrower promise than fact-checking. In the paper's definition, verifiable provenance is a text subsequence whose LLM response is lexically identical or semantically equivalent to the response from the full text. A passage can therefore help check whether the model's output is reproducible while leaving the truth of that output unresolved. The authors also note that an incorrect full-text answer or an ambiguous question can make a meaningful true provenance undefined or unhelpful.
A two-stage search for minimal text
BLIP is built in two stages. Its pruning phase explores four relevance or scan strategies, while refinement works on the remaining text and is designed to remove material without changing the answer. The paper's theorem says the refinement strategies return a minimal provenance for any data-processing task, including non-monotonic tasks, where adding text does not necessarily preserve the answer.
The evaluation covered five real-world workloads, with 500 question-document pairs for each, and two synthetic TableQA workloads containing 64,000 tokens and 100 questions each. In the main implementation, text was divided into 20 equal-sized blocks. all-mpnet-base-v2 and Mistral 7B handled ranking, while gpt-4o-mini and gemini-2-flash handled question answering and provenance inference; the reported results were based on gpt-4o-mini.
The savings varied by workload
On Qasper, NL_DEV and HotpotQA, adaptive pruning produced average provenance-size ratios of 10.2%, 12.8% and 20.2%, respectively. The corresponding cost ratios were 0.28x, 0.25x and 0.3x, with latencies of 2.1, 1.9 and 1.1 seconds. These figures are workload-level averages; the paper reports no uncertainty intervals for them.
BLIP's reported human review effort ranged from 3% to 12% across the five workloads. A separate user study used four documents and three questions per document; eight participants reviewed 12 scenarios. They gave BLIP an overall usefulness score of 4.75 out of 5, compared with 3.1 for RAG, and judged the amount of information right 82% of the time, a reported 61% higher rate than RAG. The study did not report an uncertainty interval or a formal significance test.
Known answers offered a tougher check
The synthetic table tests offered a stricter check because the expected provenance was known in advance. In Movie and Restaurant, BLIP's recovery measure was R = 1, while exact recovery topped 98% and 96%, respectively. RAG-EXACT's exact recovery was reported at between 32% and 68%. That result applies to the synthetic TableQA setting; the five plain-text workloads did not come with known true provenance.
The system can also search for several provenance candidates. For top-k provenance, increasing k from 1 to 5 raised the reported cost ratio from about 1.3 to 1.8 across workloads. The extension gives a reviewer more than one candidate to inspect, but it does not guarantee that every distinct minimal provenance will be found.
Another moving part is the semantic judge used when two answers are not exactly the same wording. Tested on 2,000 GLUE text pairs with five LLMs and four prompts, the judge averaged around 94% accuracy. This evaluation gives a reported benchmark for the semantic-equivalence check, while the formal guarantee still depends on the answer-equivalence verification process.
A useful trail, not a truth certificate
The authors describe BLIP as a general bolt-on method for producing minimal, answer-reproducing evidence at low cost. The results, however, do not show that the original answers are factually correct, that the selected text is the true human-supporting evidence path, or that the method improves decisions. The empirical evidence is also limited to the listed workloads and models, plus the eight-person study, so broader performance remains an open question.
The manuscript is labeled an arXiv preprint dated 25 Aug 2026. It says source code, data and other artifacts are available through the BLIP GitHub repository.
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