Naver's AuthGR Beats a 14B Model Using a 3B Trust-Aware Retriever
Researchers from Sungkyunkwan University and Naver introduce AuthGR, a generative retriever that ranks documents by trustworthiness, not just relevance.
Takeaways
- Researchers from Sungkyunkwan University and Naver introduce AuthGR, the first authority-aware generative retriever, at ACL 2026 Industry Track.
- A vision-language model scores site authority from text plus screenshots, reaching 97% agreement with human expertise labels.
- Three-stage training combines domain pre-training, supervised fine-tuning, and GRPO with authority scores as rewards.
- A 3B AuthGR model matches a 14B HyperCLOVAX baseline on offline P@3 with 2.35x lower latency.
- Live A/B test on a commercial search platform: +22.07% total document clicks, +22.83% top-1 CTR.
- Limitations: no public code, proprietary Korean data, single reward signal, unstudied scaling beyond 3B parameters.
A new ACL 2026 Industry Track paper from researchers at Sungkyunkwan University and Naver Corporation tackles a blind spot in generative search: current systems chase semantic relevance but ignore whether the source is actually trustworthy. The team proposes AuthGR, described as the first generative retriever that explicitly optimizes for document authority alongside relevance. Their 3B model matches a 14B baseline and delivers 22%+ lifts in real user click metrics on a commercial web search platform.
Relevance without trust is a liability
Generative information retrieval (GenIR) reformulates search as a text generation task where a language model directly emits document identifiers instead of ranking pre-encoded vectors. Existing systems primarily optimize for relevance while overlooking trustworthiness, a gap that gets dangerous in high-stakes domains like healthcare and finance. The paper's motivating example is blunt: a relevance-only model can rank an unverified personal health blog as highly as an official medical association just because both mention the same symptoms.

The authors frame three concrete obstacles to fixing this: defining authority at scale (text alone can't tell a real clinic from a slick promotional site), teaching a model an abstract concept like trustworthiness without wrecking its relevance behavior, and integrating the result into an existing production pipeline without a risky ranker swap.
A vision-language model as authority judge
AuthGR's first component treats site authority scoring as a job for a vision-language model rather than link graphs like PageRank. The VLM jointly reads textual signals (title, body, URL metadata) and a page-level screenshot, then produces a score from 0 to 100 based on a rubric covering expertise, officialness, and public interest, with checks for commercial intent and harmful content. Visuals matter because ad layout and design quality expose promotional sites that mimic authoritative language.