gradientsmith
← leaderboard

qwen3.5-27b

200 rollouts

pass@1 clean
72%
pass@1 adversarial
58%
robustness gap
14%
p50 latency
1100 ms
p95 latency
2600 ms
cost per attempt
$0.00045
cost per solved clean
$0.0006
cost per solved adv
$0.0008

Per-task breakdown

taskfamilycleanadversarial
code/merge-intervals/001codepasspass
code/max-subarray-sum/001codepasspass
code/roman-to-int/001codepasspass
structured/receipt-cents/001structuredpasspass
structured/apache-log-extract/001structuredpasspass