Unverified50% confidenceFactExact time
With the model fully frozen, merely adding random biases to its output logits can match or even exceed the effect of full weight perturbation on most reasoning tasks.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOutput examples are key to suppressing randomness and hallucinations in large models: without examples responses may be irregular, but once a formatted reference is provided the output will closely follow the expected template.76% similarUnverifiedNon-determinism stems from the random sampling mechanism of model inference itself as well as vendors' continuous iteration on model versions.75% similarUnverifiedCode cells yield the same result every time they are executed under a fixed random seed, whereas a pure 'prompt → result' approach introduces the non-determinism of large models.75% similarUnverifiedReverse-engineered prompt results may not 100% reproduce the original image, because the original prompt is highly dependent on a specific model version and random seed.74% similarUnverifiedDeterministic inference refers to controlling random factors so that the model always produces completely consistent output for the same input.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115735API
curl https://kongchang.com/api/v1/knowledge/claims/1115735MCP
get_claim(id=1115735)