Unverified50% confidenceFactExact time
Conservative Q-Learning (CQL) suppresses overestimation by penalizing the value estimates of unseen actions.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCQL（Conservative Q-Learning）由Kumar等人于2020年提出，通过损失函数惩罚项压低策略访问动作Q值、抬高数据集已有动作Q值实现保守估计74% similarUnverifiedWhen the penalty signals in alignment training are too strong, the model tends toward a conservative strategy of saying less rather than making mistakes, resulting in verbose responses, front-loaded disclaimers, and roundabout answers—a phenomenon known as over-alignment or the alignment tax.68% similarUnverifiedA learner estimating the loss distribution from limited data inevitably has bias, causing the estimate of CVaR to deviate from the true value, thereby weakening or even reversing the actual effect of risk aversion.67% similarUnverifiedOffline reinforcement learning penalizes agents for deviating from the dataset distribution by introducing constraint algorithms such as CQL and IQL, but at the cost of limited exploration capability.67% similarUnverifiedAny study comparing fine-tuning schemes with different numbers of trainable parameters should be wary of the shared learning rate trap; otherwise, conclusions that fewer parameters yield better results may need to be revisited.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1103302API
curl https://kongchang.com/api/v1/knowledge/claims/1103302MCP
get_claim(id=1103302)