Unverified50% confidenceOpinionExact time
Modern deep learning models generally exhibit a tendency toward overconfidence, which stems both from the training objective (minimizing cross-entropy) itself not directly optimizing calibration and from excessive model capacity leading to memorization of the training distribution.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDeepMind's Chinchilla research pointed out that previous models were generally oversized and undertrained, meaning they had too many parameters but insufficient training data.74% similarUnverifiedExisting large models are trained for human preferences and are prone to mode collapse, overconfidence, and lack of reliability in machine-use scenarios.73% similarUnverifiedThe joint training of this method uses only cross-entropy loss, simultaneously accounting for both accuracy and calibration quality, at the cost of introducing only one additional hyperparameter: the calibrator learning rate.73% similarUnverifiedUnconstrained base prompts cause the model to fall back to the average-state output most common in its training data, while deeply personalized prompts can significantly improve output uniqueness.72% similarUnverifiedGoal-directed belief bias is an almost inevitable byproduct of the deep reinforcement learning training paradigm.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1011930API
curl https://kongchang.com/api/v1/knowledge/claims/1011930MCP
get_claim(id=1011930)