Unverified50% confidenceReleaseExact time
The core mechanism of CoDPO (Conflict-aware DPO) has two parts: applying downweighting to conflicting pairs, with the weight strength derived directly from ρ*_eq rather than manually tuned; and dynamically adjusting the preference margin during training to offset residual drift.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/6/2026
First Seen
Valid until: 2/4/2027
Sources
Related Entities
Related Claims
UnverifiedTraining-time alignment (such as RLHF and DPO), which is often contrasted with inference-time alignment, writes value preferences directly into the weights, offering greater stability but poor flexibility, and making it difficult to dynamically adjust for individual users.69% similarVerifiedAlignment algorithms such as DPO (Direct Preference Optimization) reduce training costs while improving the consistency of model performance on open-ended instructions.67% similarUnverifiedDecoupled exponential moving average (D-EMA) adopts a configurable grouped averaging strategy that constrains interference-sensitive pretrained parameters while preserving flexibility for other parameters.64% similarUnverifiedDPO 相比 PPO 实现简单、稳定性高、资源友好,但对偏好数据质量要求高,且使用离线偏好数据限制了训练中的策略探索能力63% similarUnverifiedRLHF和DPO等对齐技术的效果高度依赖预训练基础质量,预训练阶段未习得的能力几乎无法通过后续对齐补救63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1141626API
curl https://kongchang.com/api/v1/knowledge/claims/1141626MCP
get_claim(id=1141626)