AXN:0ACA.CAPTURE

AXN:0ACA.CAPTURE.📜🪨🎬👐♎✖️

Observation

Google AI Overview, 2026-10-29 · the capture

Description

ATTRIBENCH, BY NAME, ON GOOGLE'S AI OVERVIEW, SIGNED OUT, 2026-10-29: 'The term AttriBench primarily refers to two distinct tools.' First, 'Released in 2026, AttriBench is the first fame- and demographically-balanced quote attribution benchmark dataset for Large Language Models', with its findings (disparities across race, gender and intersectional groups; 'suppression', omission of attribution though the model has the authorship data) and the Hugging Face repository. Second, attribench, a 2023 PyTorch library for benchmarking feature-attribution methods (PyPI, Read the Docs). A note: in finance, AttriBench is a Brinson-attribution exercise. Five cards: arXiv 2604.05224, PyPI, Read the Docs, the bermaneh/AttriBench dataset, Eureka Simulations.

Reading

The innocent reading first. The name belongs to three things, and the Overview separates them: the new LLM benchmark first, an older explainable-AI library second, a finance exercise in a note, each with its own source.

Analysis

Asked a name with three holders, the composition disambiguates and puts the provenance benchmark first; the disambiguation is sourced and clean. The benchmark measures suppression, the omission of attribution where authorship is known, and its own composition here omits its four authors: a small instance of the failure it names.

Match

BROAD MATCH — ATTRIBENCH ON THE OVERVIEW — THREE SENSES OF ONE NAME: THE 2026 QUOTE-ATTRIBUTION BENCHMARK (SUPPRESSION NAMED), A 2023 PYTORCH XAI LIBRARY, A FINANCE EXERCISE; THE AUTHORS UNNAMED

PER

0.5 — No person or institution named in the answer; the arXiv identifier in a card; the paper and dataset cited.

Citations

5 (0 archive-controlled)

Sealed observation

/data/captures/sealed/attribench-aio-20261029.json — SHA-256 401bf1b2af927d7e49ee7e531cd7a525ed0225c608ff88ffca909440dbc21aa0