gemini-4IMAGE
换 harness 后 Gemini 4 Argon 通过率 40%→54%
同样 48 个任务,旧 Pi harness 有 10 次因内容过滤中断;改用会重试的 Antigravity 后 8 题转通过,通过率升至 54%,仅次于 Opus 5.5。
INDRAJEET YADAV
@indrajeet8774 天前收录原推发布于 2026.10.26

INDRAJEET YADAV @indrajeet877原推发布于 2026.10.26
Same Gemini 4 Argon, same 48 tasks, new harness: 19 passes became 26. The old Pi harness ended 10 of 48 runs after a content-filter refusal. On Antigravity, which retries refusals, 8 tasks flipped to pass and 1 to fail (sign test p ≈ 0.04). Pass rate 40% → 54%, 2nd behind Opus 5.5. A leaderboard row scores model + harness.
这条是引用着另一条写的:「Argon 换 Antigravity Harness 后反超 Astra」
x.com/indrajeet877/status/2114616873748729953采集于 2026.10.27 09:54