RIKER (Roig, 2026) finds grounding accuracy and fabrication resistance are empirically separable capabilities, not two faces of one hallucination score [unverified-quant]

ragbenchmarknlp-evaluationhallucinationgroundingai-safety

A benchmark built from synthetic corpora generated from known ground truth — specifically so hallucination-probe questions can carry a verifiably correct "the information does not exist" answer, separate from ordinary extraction questions — scores 31–33 models and reports as its headline finding that grounding ability and hallucination resistance are distinct capabilities: a model can excel at locating facts that are actually present in a document while still readily fabricating answers about entities that are not. [unverified-quant — needs primary corroboration beyond this single preprint] At 32K context, reported hallucination rates across otherwise-capable models ranged from 0.0% to 93.2%, with no reported tight correlation to single-document extraction accuracy — one model (GLM-4.5) is reported combining 93.6% single-document accuracy with 0.0% fabrication, while another (Qwen3-Next-80B) is reported combining 91.4% single-document accuracy with 40.9% fabrication. (The paper's own "7.9% to 96.3%" sentence, often quoted as the 32K range, is actually drawn from its 128K-context table — see claim-riker-2026-7-9-to-96-3-range-is-128k-not-32k-table.)

The benchmark's own "Coherent Simulated Universe" is the authors' name for a document-generation technique — keeping synthetic corpora internally consistent so models cannot exploit randomness-driven tells — not their term for model fabrication, and should not be read as such. See question-corroborate-riker-2026-hallucination-rate-range for the open corroboration this note's quantitative figures still need.

Correction history.

• 2026-10-22 — This note's body originally stated the 32K-context hallucination-rate range as "7.9% to 96.3%," copying the paper's own discussion-section sentence without checking which table those two numbers trace to. A 2026-10-22 corroboration search (claim-riker-2026-7-9-to-96-3-range-is-128k-not-32k-table) found both figures (GLM-4.5 at 7.9%, Granite-4-Tiny at 96.3%) actually belong to the paper's 128K-context table (Table 5), not the 32K-context table (Table 4) this note describes. Table 4's actual 32K range is 0.0%-93.2%. The qualitative point (a wide, extraction-accuracy-uncorrelated spread at 32K) survives; only the specific pair of numbers has been corrected, from 10-inbox/raw/2026-10-22-does-an-independent-source-corroborate-rikers-roig-2026.md. The same search found no independent source corroborating either range — see claim-riker-2026-no-independent-source-same-author-admits-no-replication and claim-riker-2026-self-replication-contradicts-original-fabrication-gap.