Unverified50% confidenceOpinionExact time
Adjustments at the model level are the most difficult, but clearly also have the greatest impact on behavior, including fine-tuning or adopting domain-specific pretrained models.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRestricting training data or fine-tuning directions can only affect a model's default behavior, and cannot prevent technically capable individuals from bypassing restrictions through prompt engineering or model modification.77% similarUnverifiedIf a model mainly takes the specialised path, it is highly dependent on the topic structure seen in training and may perform in a limited way on genuinely novel topics; if it mainly takes the in-context learning path, it is better at adapting on the fly to new tasks.76% similarUnverified如果模型能理解行为准则背后的深层原因，即使面对训练中从未遇到的全新场景也能做出合理的行为选择，实现可泛化的对齐76% similarUnverified不同模型在抗诱导能力上的差异很大程度上源于后训练阶段（post-training）的策略差异76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1075296API
curl https://kongchang.com/api/v1/knowledge/claims/1075296MCP
get_claim(id=1075296)