Unverified50% confidenceFactExact time
CQL（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
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
UnverifiedThe core challenge of offline reinforcement learning is the distribution shift problem, and the academic community has proposed conservative estimation algorithms such as CQL, TD3+BC, and IQL to address it.62% similarUnverifiedQMDP是最简单的近似求解方法，假设不确定性在下一步就会消失，将POMDP降级为多个MDP的加权和，计算快但在需要主动信息收集的场景中表现不佳59% similarUnverifiedInference frameworks such as vLLM and TensorRT-LLM improve throughput through batching and KV Cache techniques, reducing the per-inference cost of models with tens of billions of parameters by an order of magnitude over the past three years59% similarUnverifiedQ4_K_M通常是兼顾内存效率与生成质量的最优折衷点，而Q2_K虽压缩率更高但在逻辑推理类任务中质量下降明显58% similarUnverifiedModern quantization schemes such as GPTQ and AWQ introduce layer-by-layer calibration and error compensation mechanisms to keep perplexity loss within 1–2%58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/547320API
curl https://kongchang.com/api/v1/knowledge/claims/547320MCP
get_claim(id=547320)