Unverified50% confidenceFactExact time
论文在一个10亿参数模型上用2000亿token的Numeratron数学与Dolma混合语料做训练，展示了中训练损失与RL性能之间的关系
1
Sources
50%
Confidence
Long-term
Relevance
8/13/2026
First Seen
Sources
Related Claims
Unverified提示词与模型版本存在深度绑定关系，不同版本模型在RLHF训练数据、指令微调策略上的差异会导致相同提示词产生不同输出，这种现象被称为「提示词漂移」70% similarUnverified斯坦福CRFM团队的HELM基准测试发现，经过RLHF对齐的模型在创意写作任务上的词汇多样性指标普遍低于对应的基础预训练模型70% similarUnverifiedDeepMind's Chinchilla scaling law demonstrates that the optimal ratio of training data volume to model parameter count evolves over time, and equal FLOP investment can produce models with significantly different capabilities70% similarUnverifiedDue to differences in training data sources, RLHF strategies, and reasoning architectures, the Claude series models and Codex based on o3/o4-mini develop different cognitive biases, making multi-model cross-validation effective at reducing bug rates.68% similarUnverifiedThe scaling laws were systematically demonstrated in 2020 by OpenAI researcher Jared Kaplan and others, with the core finding that the cross-entropy loss of language models follows a power-law relationship with model parameter count, training data volume, and compute67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/742562API
curl https://kongchang.com/api/v1/knowledge/claims/742562MCP
get_claim(id=742562)