Unverified50% confidenceFactExact time
自回归生成系统训练采用Teacher Forcing策略，推理阶段切换为自回归模式后产生分布偏移导致误差累积放大
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
GPT-5.6 Ultra子Agent架构解析与AI行业十大动向
bilibili竹言见智6/29/2026
Related Claims
UnverifiedThe core idea of the test-time compute paradigm is to shift resources from the training phase to the answer generation process, trading more inference compute for higher-quality output.74% similarUnverified分布漂移是行为克隆的核心缺陷，因训练数据全部来自正确操作轨迹，模型未见过犯错后如何恢复的场景74% similarUnverifiedBehavioral Cloning, as the most fundamental imitation learning method, suffers from the compounding error problem, where models tend to produce cascading failures on states outside the training distribution.73% similarUnverifiedSupervised fine-tuning trains models to imitate correct answers labeled by humans, while reinforcement learning fine-tuning allows models to learn optimal strategies through trial and error.72% similarUnverified蒸馏训练范式导致开源模型学习到经过高度过滤、几乎不包含失败和纠错过程的数据分布，使模型形成不愿自我纠正的性格72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/278949API
curl https://kongchang.com/api/v1/knowledge/claims/278949MCP
get_claim(id=278949)