Record-high training reward hid a fully-failing DR subgroup - aggregate metrics average over draws, gates must test per condition
aggregate-metrics-mask-subgroup-failure.Never gate on metrics aggregated across DR draws: evaluate at fixed representative conditions (especially the deployment-critical stratum), and if a difficulty axis matters, ramp it on measured per-stratum success rather than sampling the full range from iteration zero.
Symptom
omni_s1e trained under constant-wide latency DR (0, 0.06 s) posted the lineage's highest-ever Isaac reward (129) - while the --delay 2 smoke evaluation showed 3/3 falls from iter 1500 onward, persisting to early stop; the usable checkpoint window shrank to iters 500-1000.
Context
Diagnosis written plainly: "聚合奖励掩盖重延迟尾部子群体失败" - the aggregated reward averages over latency draws, so the majority of light-delay environments can mask the total failure of the heavy-delay tail. The remedy for the training side was a survival-gated ratchet curriculum (survival_gated_latency): the sampling cap starts at 0.02 s and rises +0.01 only when a 4096-reset window's survival (time_out share) reaches >=90%, capped at 0.06, ratchet up-only - "增益与延迟耐受一起长,不升到策略撑不住的 地方" (gain and delay tolerance grow together; never raise past what the policy can hold). The detection side was already in place from the noise-crutch episode: the per-condition smoke curve, not the training reward, is the health readout.
Change
Latency exposure made curriculum-gated on measured subgroup survival instead of uniform-from-zero; per-condition (--delay 2) smoke evaluation kept as the authoritative curve; watcher scoring adjusted (survival weighted 3x) so recovery during hard phases is not early-stopped away.
Outcome
The failure mode was caught by the smoke curve within one generation; the follow-up redesign (deterministic staged latency) superseded the ratchet, but the aggregate-masking lesson held through both.
Mechanism
Expected-return training weights each DR draw by probability, so a subgroup can contribute bounded loss while being catastrophically failed; any scalar averaged over the randomization cannot distinguish "uniformly decent" from "great on easy draws, dead on hard ones". Only conditioning the evaluation on the stratum reveals the split, and curricula should raise difficulty on measured stratum success, not on schedule.
Applies when
- training reward hits records while a fixed-condition eval degrades
- wide DR on an axis where deployment sits at one known value
- designing curricula for difficulty axes (delay, push, terrain)
“常量 latency DR (0,0.06) 从零训被证伪——Isaac reward 129 历代最高,但 --delay 2 冒烟 iter1500 起 3/3 全摔持续到早停(聚合奖励掩盖重延迟尾部子群体失败,可用窗口只剩 500/1000)。… 采样上限 0.02 起步 … ≥90% 才 +0.01s,0.06 封顶,棘轮只升不降。”