Unverified50% confidenceSolutionExact time
GAP-DPO mitigates distribution shift caused by off-policy sampling through an epoch-wise regeneration mechanism that regenerates samples at each epoch.
1
Sources
50%
Confidence
Long-term
Relevance
10/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUnder off-policy sampling conditions, when the preference margin is aligned in direction with the utility gradient, DPO's update shifts from a corrective signal to a reinforcement-style update similar to reinforcement learning.67% similarUnverifiedDPO完全绕过了强化学习阶段，是RLHF之外的对齐技术路径之一；Anthropic探索了RLAIF，Meta研究了拒绝采样（Rejection Sampling）60% similarVerifiedDuring the autoregressive generation phase, generating each new token requires recalculating attention over all historical tokens, with computational complexity growing as O(n²) with sequence length; KV Cache can reduce this to O(n) incremental operations.59% similarUnverifiedEPD disaggregation introduces new architectural complexity, requiring evaluation of intermediate embedding vector transmission overhead, the ratio of instance counts at each stage, and scheduling strategies under load imbalance—representing a trade-off between performance and complexity.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/966742API
curl https://kongchang.com/api/v1/knowledge/claims/966742MCP
get_claim(id=966742)