Unverified50% confidenceSolutionExact time
The team restored accuracy loss to within 1% in stages through reinforcement learning (GSPO), On-Policy Distillation, and a ThinkingCap 3.6 27B adapter, treating On-Policy Distillation as the key step.
1
Sources
50%
Confidence
Long-term
Relevance
9/14/2026
First Seen
Sources
Related Entities
Related Claims
Unverified根据官方在 WANDR 基准测试上的数据，Grok 4.6 位于性能与效率的帕累托前沿62% similarUnverified在 GSM8K 基准上，标准思维链准确率约为 56%-63%（GPT-3 规模），PAL 可提升至 72% 以上60% similarUnverifiedEmsyAI was trained for 15,000 steps using FP16 mixed precision, with validation perplexity converging to 5.8659% similarUnverified基于尾调用的字节码分发在某些基准测试中可带来 5-15% 的性能提升59% similarUnverified开发者最终选择KL散度最低（0.028）的变体，GSM8K下降1.83%，MMLU保持不变，HarmBench合规率从30.8%提升至100%58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/920996API
curl https://kongchang.com/api/v1/knowledge/claims/920996MCP
get_claim(id=920996)