Does greedy decoding eliminate hallucinations?
No. It selects the highest-scoring continuation under the model, which can still produce a false or unsupported statement if that continuation is favored.
Explore This Question
Search for additional explanations using these tools:
More FAQs in Why doesn’t an LLM always select the token with the highest probability?
No. Temperature reshapes the probability distribution used in sampling. Candidate tokens remain influenced by the model’s scores and by any additional filtering rules.
Yes. A token’s local probability does not fully measure the usefulness of the complete continuation. Different early choices can lead to substantially different later sequences.
Still have questions?
Browse all our FAQs or reach out to our support team