Unverified50% confidenceFactExact time
Fine-tuning a large language model on data that the model itself considers wrong but whose content is harmless can induce broad misalignment.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedExisting language model data attribution evaluations suffer from variable coupling, that is, the evaluation set simultaneously varies multiple dataset properties such as redundant documents, lexical distractors and text length, making it impossible to pinpoint the specific reason an attribution algorithm fails.80% similarUnverifiedThe process by which a model compresses high-dimensional semantic representations into a discrete token sequence is lossy—fine-grained semantic information, uncertainty distributions, and contextual associations are flattened when converted to text.78% similarUnverifiedSmall models are often unstable in tool-call format, producing missing parameters, wrong field names, or mixed natural-language output, which causes parsing failures and task interruption.78% similarUnverifiedModels can almost always fix explicit syntax errors with ease, but frequently fail on invisible logical errors such as data leakage.78% similarUnverifiedModels tend to trust text they themselves previously generated (self-consistency bias), which makes context-poisoning-type errors extremely difficult for the model itself to detect.77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1134292API
curl https://kongchang.com/api/v1/knowledge/claims/1134292MCP
get_claim(id=1134292)