Unverified50% confidenceFactExact time
Frequency Penalty对已频繁出现的token施加惩罚以抑制重复用词，Presence Penalty对任何已出现过的token施加固定惩罚以鼓励探索新话题
1
Sources
50%
Confidence
Long-term
Relevance
7/10/2026
First Seen
Sources
提示词工程实战：程序员的Prompt调优核心方法论
bilibili大模型Agent搭建7/1/2026
Related Claims
Unverified重复惩罚、频率惩罚和提高采样温度是推理阶段对抗重复的补救手段，但本质上治标不治本64% similarUnverified通用越狱指单一提示词或操作序列能系统性绕过模型所有安全约束，而针对性越狱只能解锁某一类具体有害行为57% similarUnverified引入重试机制时应带指数退避，但需警惕重试放大效应57% similarUnverifiedEWC在新任务训练时增加惩罚项L_new + λ·Σ_i F_i·(θ_i - θ*_i)²，Fisher信息值越高的参数惩罚力度越大，越难被修改54% similarUnverifiedLuaJIT的惩罚机制实现在lj_dispatch.c中，trace起点abort后惩罚值会指数级增长（通常翻倍）54% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/424705API
curl https://kongchang.com/api/v1/knowledge/claims/424705MCP
get_claim(id=424705)