Unverified50% confidenceFactExact time
RL对pass@1的提升而不影响pass@k，本质上是在压缩模型的输出分布使其更集中在正确答案附近，即分布锐化
1
Sources
50%
Confidence
Long-term
Relevance
8/13/2026
First Seen
Sources
Related Claims
UnverifiedRLHF本质上是在模型输出分布上施加统计偏好，而非植入真正的价值理解，导致表面对齐在面对对抗性输入时易崩塌68% similarVerifiedFrom a Bayesian perspective, L1 regularization corresponds to a Laplace prior on the parameters, and L2 regularization corresponds to a Gaussian prior.62% similarUnverifiedWeights tend to approximate a symmetric distribution and are suitable for symmetric quantization, while activations, after passing through nonlinear functions such as ReLU, exhibit shifted distributions more suitable for asymmetric quantization.61% similarUnverifiedRLHF过程中形成的安全约束以分布式方式编码在数十亿参数中，而非硬编码规则，因此同一越狱手法在措辞略微调整后往往仍能奏效61% similarUnverifiedvLLM's Speculative Decoding guarantees, through a rejection sampling mechanism, that the output distribution is completely identical to that of direct generation with the large model, with a speedup ratio typically between 1.5× and 3×.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/742558API
curl https://kongchang.com/api/v1/knowledge/claims/742558MCP
get_claim(id=742558)