Verified65% confidenceFactExact time
With only a relatively small amount of targeted training data, it is possible to make a large language model give entirely incorrect answers to clearly stated facts
3
Sources
65%
Confidence
Long-term
Relevance
7/7/2026
First Seen
Sources
反向微调实验：AI如何被训练相信东京不是日本首都
hackernewshackernews7/5/2026
Related Claims
UnverifiedIn large language model training, a tiny error often occurs well before any anomaly appears in the loss curve.78% similarUnverifiedThe training objective of large language models is to minimize the prediction loss of the next token (cross-entropy loss), not to verify the factual accuracy of generated content.78% similarVerifiedLarge models output the most likely next word rather than precise retrieval of facts, which leads to hallucinations in domains with sparse coverage in the training data.78% similarUnverifiedForcing small language models to approximate deterministic problems probabilistically sacrifices accuracy and energy consumption, and much of the unreliability in edge small models is avoidable.76% similarUnverifiedThere is a gap between the output confidence of language models and their actual accuracy, a phenomenon known in machine learning as "miscalibration."75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/177046API
curl https://kongchang.com/api/v1/knowledge/claims/177046MCP
get_claim(id=177046)