Fetching from the wire…
Public story · 2026-08-17 · high
Accuracy above 95% on the anchor-free version of the same questions didn't stop the shift toward a wrong but plausible fact.
Why now: AnchorBench's results were covered in the August 17, 2026 briefing, and the RAG pathway finding applies directly to any system retrieving from a corpus it doesn't fully control.
AnchorBench tested fourteen models for anchoring bias and found plausible anchors move judgments more than irrelevant ones, according to the paper posted to arXiv (2608.14320).
The gap matters for anyone running RAG over a corpus they don't fully control. A wrong but plausible fact sitting in retrieved text can pull a model's answer off course, even one that scores above 95% on anchor-free control.
AnchorBench separates plausible anchors from irrelevant ones and tracks the bias across multiple injection pathways. The paper doesn't specify how many pathways beyond External and RAG, though it tested more than one.
Anchoring turned out to be strongly pathway-dependent, and its pull weakens as the anchor moves further from the evidence-supported answer. That decay showed up most clearly on the External and RAG pathways.
High accuracy on the anchor-free control doesn't predict robustness once a plausible anchor arrives through retrieval.
Each link below shares sources, entities, or timing with this story.
RAGOCR competes with RAG / Shared entity: RAG / Same source domain / Earlier coverage / Tension
Linked by a graph relationship (RAGOCR competes with RAG); both cover RAG; reported by the same outlet (arxiv.org).
Shared entity: RAG / Same source domain / Earlier coverage / Tension / Downstream implication
Both cover RAG; reported by the same outlet (arxiv.org); earlier RAG coverage from 2026-07-30.
Shared entity: RAG / Same source domain / Shared topic / Earlier coverage
Both cover RAG; reported by the same outlet (arxiv.org); overlapping topics (answer, model).
Both cover RAG; reported by the same outlet (arxiv.org); overlapping topics (answer, model).
Shared entities / Earlier coverage / Tension
Both cover High, RAG; earlier High coverage from 2026-03-02; pushes against this story (vs).
Shared entity: RAG / Same source domain / Earlier coverage / Tension
Both cover RAG; reported by the same outlet (arxiv.org); earlier RAG coverage from 2026-08-14.
Both cover RAG; reported by the same outlet (arxiv.org); earlier RAG coverage from 2026-08-10.
Both cover RAG; reported by the same outlet (arxiv.org); earlier RAG coverage from 2026-08-04.