Unverified50% confidenceFactExact time
Adversarial attack methods such as PGD and FGSM produce perturbations that are nearly imperceptible at the pixel level, yet can cause models to fail completely.
1
Sources
50%
Confidence
Long-term
Relevance
9/1/2026
First Seen
Sources
Related Entities
Related Claims
Unverified仅靠微调阶段的RLHF或许永远无法产生真正鲁棒的安全对齐62% similarUnverified研究表明RLHF在模型残差流中留下的安全拒绝行为呈现低秩线性结构，主要分布在少数几个主方向上，因此易被正交投影移除61% similarUnverifiedResearchers have found that fine-tuning only a small number of model parameters is sufficient to remove safety alignment (de-alignment attacks), and that certain jailbreak prompts can transfer across different models.60% similarUnverifiedThe Hunyuan model tends to oversimplify and smooth, with noticeable jagged artifacts, indicating that resolution is still insufficient.60% similarUnverifiedA purely agent-based interaction that completely abandons charts may represent a loss of capability for complex analytical scenarios that require multi-dimensional cross-analysis and long-term trend comparisons.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/838967API
curl https://kongchang.com/api/v1/knowledge/claims/838967MCP
get_claim(id=838967)