Unverified50% confidenceTradeoffExact time
During joint fine-tuning, a lower learning rate helps stabilize policy updates and avoid destroying fine-grained behaviors that have already been learned.
1
Sources
50%
Confidence
Long-term
Relevance
10/22/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOff-policy reinforcement learning reuses samples generated by older policies, which brings efficiency advantages but introduces distribution shift and high-variance gradient estimation problems.78% similarUnverified在Rainbow消融实验中，移除优先经验回放和移除分布式学习对性能的伤害最大，移除dueling架构的影响相对较小77% similarUnverifiedThe weighting strategy multiplies quality scores onto the behavior cloning loss, making the gradient contribution of low-quality demonstrations smaller.77% similarUnverifiedAdaptive step-size control is a viable mitigation direction: write only a small number of positions per step in the early stages, and gradually increase the number of parallel writes as dependencies weaken.76% similarUnverifiedUnder a local quadratic model, a larger batch means fewer but longer update steps, with each step incurring a greater loss on curvature, resulting in a critical batch size that depends on the learning rate, beyond which sample efficiency declines.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1086393API
curl https://kongchang.com/api/v1/knowledge/claims/1086393MCP
get_claim(id=1086393)