Unverified50% confidenceBenchmarkExact time
RLDS reduces wall-clock time per step on ScienceWorld by 10.9% compared to scalar GRPO, because long trajectories amortize the fixed overhead of reflection and scoring.
1
Sources
50%
Confidence
Long-term
Relevance
9/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGRPO 相比传统RLHF降低显存占用约40%63% similarUnverifiedRLDS achieved an improvement of +11.5 points on ScienceWorld (95% confidence interval [+9.8, +13.3]) and +9.8 points on FrozenLake ([+7.0, +12.8]).63% similarUnverified研究表明当ReAct循环超过10步时，模型的任务完成率会出现显著下降63% similarUnverifiedAfter migrating from glibc-based images to Alpine (musl), applications may experience performance degradation ranging from 10% to several times.63% similarUnverifiedAccording to OpenAI's early 2024 technical report, by annotating sycophantic behavior in RLHF data and assigning negative rewards, the unconditional acquiescence rate of subsequent GPT-4 iterations declined.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/946125API
curl https://kongchang.com/api/v1/knowledge/claims/946125MCP
get_claim(id=946125)