Unverified50% confidenceFactExact time
Alpaca、Vicuna采用黑盒蒸馏策略，通过收集教师模型的输出文本进行监督微调
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
Fable延期至7月19日：AI交互叙事产品的挑战与启示
Related Claims
Unverified文本反馈（Textual Feedback）方法通过教师模型对 rollout 片段给出自然语言提示，将稀疏标量奖励转化为密集局部化的文本监督信号58% similarUnverified跨分词器蒸馏的标准做法是将教师top-K logits投影到学生token上，在分段边界处把教师token映射到对应学生token序列的第一个token，这是一种多对一映射57% similarUnverified非自回归机器翻译通过知识蒸馏解耦策略使用自回归教师模型生成的伪标签训练，将BLEU分数提升幅度超过10个点，同时推理保持并行解码57% similarUnverifiedComposer 2.5 introduces a text-feedback-based directed reinforcement learning method that inserts local prompts as teacher signals at the positions where the model makes errors and trains using on-policy KL divergence loss.56% similarUnverifiedIn language learning scenarios, RAG technology first retrieves real usage examples from authoritative corpora (such as COCA and BNC), then provides them as context to a large model to generate example sentences, reducing the risk of hallucination.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/503410API
curl https://kongchang.com/api/v1/knowledge/claims/503410MCP
get_claim(id=503410)