Benchmark Radar Day 95: Answering Spatial Questions Is Not the Same as Acting in Space

Day 95 of Benchmark Radar. The scan fetched 1,308 items, published 568, and marked 250 as recommended.

Author: Koutian Wu; GitHub: ktwu01

Benchmark Radar is live. Open the dashboard, star the repository, or read the paper. The Hugging Face dataset provides the data.

BEHAVIOR-Spatial tests whether spatial answers from vision-language models carry over to vision-language-action behavior in simulated robotics. Its components separate answering spatial questions from predicting and performing actions, rather than treating visual understanding as a proxy for embodied capability. High question-answering scores are not assumed to be action-relevant spatial competence. behavior-spatial-benchmark/BEHAVIOR-Spatial.

The Tech Graphics Benchmark publishes its prompts, generated images and manifests, and raw human ratings as separate linked datasets. Generation identities are opaque and shuffled, and annotators do not see the model identity until after saving a rating. The blind-review design supports judging text, arrows, and layout in generated technical graphics without model recognition steering the scores. danielrosehill/Tech-Graphics-Benchmark-Ratings.

Why this matters

A proxy score earns its place only when it predicts the capability it stands for. Visual understanding should be tested where it is used, in action, and human ratings should be collected blind and published in pieces, so the analysis can be rerun and the model’s name cannot steer the score.

Scan coverage

This is a keyword-filtered discovery feed, not a representative survey of AI research. The archived briefing used 135 of 568 captured evidence records. These findings summarize its cited releases; discovery dates do not necessarily equal original publication dates.

Data and briefing: 2026-10-29 archived snapshot.