Unverified50% confidenceFactExact time
The discrete sampling process of autoregressive language models is non-differentiable, causing the gradient signal from classifiers to be effective only within a limited local range, making it difficult to achieve structural control spanning an entire sentence.
1
Sources
50%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPrompt engineering has inherent limitations: the effectiveness of instructions decays over conversation turns (instruction forgetting), and different model versions respond significantly differently to the same instruction.70% similarUnverified无梯度方法样本效率低，在高维参数空间中缺乏梯度指引意味着需要海量函数评估才能收敛68% similarUnverified近似遗忘通过梯度上升等方法主动削弱特定数据对模型参数的影响，但效果难以量化验证67% similarUnverified大模型对指令的敏感度呈非线性分布——主干指令的微小改动可能引发全局行为漂移，而在末尾追加具体的边界条件则相对安全66% similarVerifiedThe fundamental reason prompt injection is difficult to defend against is that LLMs cannot reliably distinguish between data and instructions.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/843365API
curl https://kongchang.com/api/v1/knowledge/claims/843365MCP
get_claim(id=843365)