Unverified50% confidenceFactExact time
Parameter-efficient adversarial fine-tuning methods based on LoRA still lag clearly behind full-parameter adversarial training in robustness.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/30/2026
First Seen
Valid until: 1/28/2027
Sources
Related Entities
Related Claims
VerifiedAlignment training optimizes the model's statistical behavior in known scenarios rather than establishing formally verifiable security boundaries, and may fail when faced with rare instruction combinations or role-playing frameworks outside the training distribution.80% similarVerifiedSimply stacking parameters without sufficient training data results in a model that is large but not capable; a leap in parameters must be accompanied by a proportionally expanded training corpus and computing power to realize capability gains.79% similarUnverifiedExternal scheduling at inference time and internally learned capabilities during training differ in generalization performance and controllability.77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1114008API
curl https://kongchang.com/api/v1/knowledge/claims/1114008MCP
get_claim(id=1114008)