Unverified50% confidenceFactExact time
A frozen model refers to a model whose weight parameters all remain unchanged during the inference phase, with no fine-tuning, adapter training, or online learning performed.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified无状态推理（Stateless Inference）指模型权重在推理阶段保持冻结，不会因用户使用历史自动更新，这与GPT等大语言模型的推理机制本质相同75% similarUnverifiedLLM在推理阶段参数是冻结的，模型不会因用户交互而更新权重，因此本质上是无状态的73% similarUnverifiedSpeculative decoding is a lossless acceleration technique that does not alter the model's output distribution; it uses a draft model with far fewer parameters than the target model to generate candidate tokens, which the target model then verifies in parallel.71% similarUnverified端侧小模型在指令遵循上往往不如云端大模型稳定，通过约束解码强制模型按预定义语法输出可在不牺牲推理速度的前提下保证工具调用的结构化可靠性71% similarUnverifiedThe parameter-efficient fine-tuning (PEFT) field includes methods such as LoRA, QLoRA, Prefix Tuning, Adapter, and IA³, with a shared design philosophy of freezing most parameters and training only a small number of newly added parameters.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/893617API
curl https://kongchang.com/api/v1/knowledge/claims/893617MCP
get_claim(id=893617)