Unverified50% confidenceFactExact time
Token-by-token KL analysis shows that the effect of ShallowDefend training is concentrated in the brief deliberation interval at the start of the response.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
Unverified最大似然估计训练目标等价于最小化模型分布与训练数据经验分布之间的KL散度70% similarUnverifiedApplication-layer jailbreak defenses must simultaneously rely on robustness improvements at the training stage and dynamic monitoring at the inference stage, as reinforcement along a single dimension yields limited results.69% similarUnverifiedShallowDefend training takes place at the "payload turn", the turn in which the harmful intent actually surfaces, and uses reinforcement learning to consolidate the decision behavior.69% similarUnverifiedA contemporaneous study by O'Brien et al. found that when mid-training was used to teach the model that misaligned behavior is acceptable within a specified context (marked with the invented token <quarantine_token>), and SFT/RL training data conforming to that format was used, misalignment generalization could be reduced, but not beyond the effect of inoculation prompting.67% similarUnverifiedIn behavior-constrained reinforcement learning, KL regularization can be achieved with access only to the gradient of the reference policy, and the gradient of the reference policy can be obtained through a velocity score transformation.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115336API
curl https://kongchang.com/api/v1/knowledge/claims/1115336MCP
get_claim(id=1115336)