Unverified50% confidenceFactExact time
在Transformer架构中，对于拥有N个参数的模型，生成每个token大约需要2N次浮点运算
1
Sources
50%
Confidence
Long-term
Relevance
8/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the traditional Transformer architecture, the amount of computation per token is fixed, while Chain-of-Thought technology increases the effective computation by having the model generate intermediate reasoning steps.75% similarUnverifiedIn the Transformer architecture, the computational cost per generated token is approximately proportional to the number of activated parameters; for dense models the activated parameter count equals the total parameter count, while for MoE models the activated parameter count is far smaller than the total parameter count.75% similarUnverifiedModern large models are based on the Transformer architecture, and generating each token requires one full forward pass over all the model's parameters.74% similarUnverifiedTransformer架构在每次推理时必须处理完整的输入序列，代理执行到第N步时输入上下文必须包含前N-1步的完整执行轨迹74% similarUnverifiedEach additional token generated by an autoregressive model provides one more opportunity for a Transformer forward pass computation, trading sequence length for computational depth.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/808802API
curl https://kongchang.com/api/v1/knowledge/claims/808802MCP
get_claim(id=808802)