Use when designing or auditing the experimental program of a CVPR paper, covering benchmark and baseline selection under matched-compute fairness, the ablation study reviewers treat as mandatory, qualitative and failure-case evidence, efficiency metrics tied to the Compute Reporting Form, and generalization tests beyond a single dataset.
SkillScore96Maintained
Automatic
◆Research pack
Last verified 9/1/2026
What it changes
Passed 24 of 25 test cases with the skill, and 3 of 25 without it.
Without skill12%
With skill96%
Checks it passed that the model alone missed
Includes the exact tag [Efficiency] in the output.
Includes the exact tag [Qualitative] in the audit.
Includes the exact tag [Qualitative] in the output.
and 22 more in the full report
21 fixed3 passed either way1 still failing0 broke
Pass-rate lift
+84 pp
Tokens
+27%
Agent turns
0%
Response time
-21%
Model gemini-3.6-flashMeasured 9/1/2026v124 of 25 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 21 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 21
Prompt
Table 1 mixes performance metrics from published paper tables with metrics we re-ran locally under our own evaluation protocol, but we listed all numbers in the exact same font without any distinction. Audit this table presentation using the standardized tags.
What was checked, without and with the skill
Checked
Without skill
With skill
Includes the exact tag [Comparison risks] in the output.
Missed
Passed
Recommends using footnotes to distinguish locally re-run baseline metrics from quoted paper figures.
Passed
Passed
case-14 · gemini-3.6-flash · 9/1/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md119 linessingle file
CVPR Experiments
CVPR runs on benchmark evidence: reviewers at the 2026 edition sorted 16,092 submissions
largely by asking "do the tables prove the sentence?" This skill designs an experimental
program that answers the four questions every vision review implicitly asks — does it
work, why does it work, when does it fail, and what does it cost.
Does it work: main comparisons
Benchmarks: use the datasets your subfield's last two cycles used, current
versions, standard splits. A new task may justify a new benchmark, but then the
benchmark itself becomes a contribution needing validation (and, if claimed as one,
public release by camera-ready — verified 2026 policy).
Baselines: the leaderboard's current top methods plus the strongest simple
baseline. The comparison that kills papers in review is the one you omitted because
it was too strong.
Fairness: match backbones, pretraining corpora, input resolution, and training
schedule wherever possible — or tabulate the mismatch explicitly. Beating a
ResNet-era method with a ViT-L and calling it method innovation is the single most
common CVPR review objection.
Provenance: mark which numbers are re-runs under your protocol vs. quoted from
papers; asterisk and footnote the re-runs.
Why it works: the ablation is not optional
Vision reviewers treat ablations as the paper's proof of understanding. Structure the
grid so each row removes or replaces exactly one design decision:
# ablation-matrix.txt — one experiment per line, one variable per experiment
A0 full method (reference row)
A1 - temporal attention → per-frame baseline tests the core claim
A2 - our loss → standard L1 is the loss or the architecture doing it?
A3 - pretrain → from scratch how much rides on initialization?
A4 swap: our module in baseline X does the gain transfer?
A5 sensitivity: key hyperparameter sweep is A0 a lucky point?
Rows A4 (transplant) and A5 (sensitivity) separate memorable ablation sections from
perfunctory ones. Every ablation row cited in prose belongs in the 8-page body; the