DRACO grades deep research outputs on 100 tasks against task-specific rubrics along four axes. The Citation Quality axis is described as 'Citations to primary source documents' and makes up 12% of criteria (4.8 of 39.3 per task on average). Table 13 prints normalized Citation Quality scores: Perplexity Deep Research (Opus 4.6) 64.6, Perplexity Deep Research (Opus 4.5) 62.5, Claude Opus 4.6 56.2, Gemini Deep Research 51.5, OpenAI Deep Research (o3) 45.8, OpenAI Deep Research (o4-mini) 42.5, Claude Opus 4.5 42.1. Each criterion receives a binary MET or UNMET from an LLM judge (Gemini-3-Pro), averaged over 5 grading runs. The axis scores whether a response cites the primary documents the rubric asks for; it is not a per-claim check of whether each cited passage supports the sentence it is attached to.

Nine of ten authors are at Perplexity, one at Harvard University; Perplexity is the vendor of the system that ranks first on every axis, the tasks are sampled from Perplexity Deep Research requests, and the rubrics were designed and validated in a process the vendor ran together with The LLM Data Company and 26 recruited domain experts (Section 4.1): collector and beneficiary coincide, recorded as subordinate. arXiv preprint without venue. The judge model was chosen drawing on an internal human-LLM alignment study that is not published in the paper. The counter-check that exists: the dataset and rubrics are public (hf.co/datasets/perplexity-ai/draco), and the paper reports scores under GPT-5.2 and Sonnet-4.5 as alternative judges with stable ranking; an evaluation by a party without a stake is not part of the paper.

A party with no stake reruns the public DRACO tasks and rubrics and the Citation Quality ordering changes, or a per-claim support audit of the same outputs shows the axis score does not track the share of citations that support their claims. Query to run: independent regrade of the DRACO Citation Quality criteria plus a per-citation support check on the same reports.

The vendor's benchmark shows its deep research product cites best. Too coarse: the axis covers 12 percent of the criteria, measures citation of primary documents rather than support of each claim, is LLM-judged, and tops out at 64.6 percent for the vendor's own system.

https://arxiv.org/pdf/2602.11685v1 p. 12, Table 13; p. 6-7, Section 4.1 and Table 4; p. 9, Section 5.1 | 2026-02-12 · arXiv 2602.11685 · Zhong et al. (Perplexity), DRACO Plan 02b: the rivals' 42 to 56 stand only in Table 13, p. 12.