Evidence · DiggingBeagle record
PointerBench: fine-grained GUI grounding benchmark
Original paper, task setup, model comparisons and coordinate-frame error analysis.
- Published
- Oct 22, 2026
- Source role
- primary disclosure
Evidence record
Original paper, task setup, model comparisons and coordinate-frame error analysis.
Claim-level citations (3)
- supportsPointerBench finds GUI agents miss precise targets despite strong general screenshot scores: The original benchmark contains 1,500 screenshot-instruction tasks across 100 professional applications with a fixed 1024-by-768 coordinate contract.
Dataset construction; benchmark design
- supportsPointerBench finds GUI agents miss precise targets despite strong general screenshot scores: The reported leading model reaches 83.2% overall but 69.6% on text targets; coordinate normalization materially shifts results for some model families.
Main results; coordinate-system analysis
- supportsPointerBench finds GUI agents miss precise targets despite strong general screenshot scores: The PointerBench paper reports Fable 5 at 83.2% overall and 69.6% on text targets, versus GPT-5.6 Sol at 77.7% overall and 53.4% on text, measured under its static coordinate-grounding protocol.
Section 3.1, frontier-model results table and Text suite