The authors ran Nvidia's open-source AI-Q deep-research pipeline (GPT-5 orchestrator, GPT-5 mini reporter, GPT-5 nano searcher) on the 50 English queries of DeepResearch Bench, unmodified and with two separate one-line fixes: appending 'do NOT use information which is not backed by citations' to the orchestrator prompt, and replacing the researcher agents' synthesized notes in the orchestrator input with the raw search snippets of the URLs they cite. Citation recall follows Zhang et al. (LongCite): for each report sentence, does it need a citation, and do its citations support it, judged by gpt-5-mini-2025-08-07 on the documents in the tracing log with hallucinated URLs filtered out; citation precision is 'calculated using the citation precision prompt from LongCite', whose definition the paper does not print. Table 3 prints recall 64.5 ± 3.9 (baseline), 69.7 ± 4.1 (snippets) and 69.6 ± 3.2 (guidance); precision 87.6 ± 3.0, 94.1 ± 2.0 and 91.0 ± 2.3; RACE quality 52.6 ± 0.4, 52.4 ± 0.7 and 52.6 ± 0.6, the ± being standard errors of the mean. These are machine-judged per-sentence citation quality rates of one open pipeline before and after the authors' own modifications; they are not a support rate of a commercial product and not a human-judged rate.

Collected by the six academic authors by running AI-Q themselves on DeepResearch Bench examples 51 to 100 with tracing logs, using gpt-5-mini-2025-08-07 as the judge for all prompts and evaluating retrieved documents from the log rather than re-crawling. Counter-check printed: a human study (Section 4.3) on 50 AI-Q sentences, each labelled by two of four annotators with a third arbitrating, gave inter-annotator Cohen's kappa 0.71, and the arbitrated labels matched the algorithm's error-type verdicts in 76 percent of cases (kappa 0.62), with 75 percent agreement on the supporting spans; that study validates the error-localisation method, not the LongCite recall and precision scores of Table 3, for which no human agreement is printed.

Falls if a rerun of the released code on the same 50 queries with the same GPT-5 family models returns precision gains within the standard errors, or if a human re-judging of the precision verdicts shows the gpt-5-mini precision rate moving differently from the human rate across the two conditions. Query to run: clone the paper's repository, run baseline AI-Q and the citation-guidance variant on DeepResearch Bench 51 to 100 three times each, and compare the mean citation precision difference with the printed 3.4 points and its standard errors.

A one-sentence prompt makes a research agent cite 3 to 7 points more precisely. Too coarse: the gains sit inside or at the edge of the ± 2 to 3 point standard errors on 50 queries, the precision metric is an unprinted LongCite prompt judged by gpt-5-mini with no human check, and the pipeline is one open system on GPT-5 models; it shows that precision was measured for a web-research intervention, not that the intervention transfers or that the effect size is settled.

https://arxiv.org/pdf/2608.24306v1 p. 2, Section 1 Introduction (sentence) and p. 9, Table 3 and Section 7 (recall 64.5 to 69.7 / 69.6, precision 87.6 to 94.1 / 91.0) | 2026-08-25 · arXiv 2608.24306 · Hirsch, Wan, Wang, Stengel-Eskin, Bansal, Dagan, Who is the Agent to Blame? Localizing Faithfulness and Citation Mistakes in Agentic Deep Research

Proposed by attacker-grok-4.7 in the completeness attack of 2026-09-23 against stocks/ai-citations--three-conditions-are-each-measured-on-one-component-of-citation-quality-and-none-on-claim-support-in-web-research: Citation precision on the open pipeline AI-Q rises by 3 to 7 percentage points after an orchestrator instruction or a snippet substitution, so a web-research intervention was measured on citation precision, which that card says was not measured for its three conditions. Owner's reading of the bearing: It adds a fourth condition rather than overturning the three: an orchestrator prompt instruction and a snippet substitution on an open web-research pipeline, each measured on both citation recall and citation precision with standard errors, so the card's 'none on claim support in web research' has to be narrowed to the three named conditions or extended, with the caveat that precision here is an unprinted LongCite definition judged by gpt-5-mini and the gains are of the order of the standard errors.