Unverified50% confidenceFactExact time
The experiments used three training random seeds paired with a fixed test set split, along with a limited hyperparameter tuning budget.
1
Sources
50%
Confidence
Long-term
Relevance
9/21/2026
First Seen
Sources
Related Entities
Related Claims
Unverified即便固定所有随机种子，在不同硬件或并行度下，训练结果仍可能出现被大模型放大为显著能力差异的细微偏差77% similarUnverifiedThe experimental scale was 5 reward configurations × 5 random seeds × 2,000 episodes per run.72% similarUnverified设置3次重试上限的工程考量是既避免无限循环消耗大量token，也覆盖大多数由随机性导致的格式波动，通常1-2次重试即可解决72% similarUnverifiedZero-shot prediction results are cached and reused across random seeds, so this part does not constitute truly independent repeated experiments.72% similarUnverifiedThe three-preview-seed mechanism runs three different seeds as low-cost previews before formally generating the high-cost video, reducing the trial-and-error cost of being unsatisfied after a full-power generation.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/937863API
curl https://kongchang.com/api/v1/knowledge/claims/937863MCP
get_claim(id=937863)