[HUGGINGFACE]score: 0.36

Spreadsheet Action Prediction Benchmark with 12K Manually Curated Edit Sequences

June 10, 2026

A new benchmark for next-action prediction in spreadsheets includes 52 manually curated sequences totaling 12K actions, addressing the absence of public edit histories. Evaluation is online — predictions are made after each action — covering spatial, temporal, and composite action types.

HOW THIS AFFECTS YOU

●

builderRelevant if building productivity copilots for spreadsheet tools — the benchmark defines the task structure and evaluation protocol you'd need to target.

●

researcherFills a data gap for spreadsheet agent research with a realistic online evaluation protocol.

read original ↗huggingface.co

← back to feed