Unverified50% confidenceFactExact time
The fundamental issue of cross-task calibration is that the model's probability output is learned on a specific training distribution; when the input distribution changes, the reliability of the probabilities drifts but the model cannot perceive it.
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedModel capability cannot be judged by parameter scale alone, as model performance is influenced by multiple factors including training data distribution, reasoning pathways, and randomness.76% 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75% similarVerified涌现能力是指当模型参数规模突破某一临界点后，突然涌现出训练目标之外的复杂能力，如逻辑推理、代码生成、跨语言迁移75% similarUnverifiedBecause the Tic-Tac-Toe training data covers all possible states, the model's learning task degenerates into a memorization task with no out-of-distribution issues, and the compounding error problem of behavioral cloning naturally dissolves.74% similarUnverifiedWhen faced with questions outside the training distribution (out-of-distribution queries), models may degrade into guessing based on surface-level patterns.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/992861API
curl https://kongchang.com/api/v1/knowledge/claims/992861MCP
get_claim(id=992861)