Token Troubles and Beam Blunders: LLMs Need a Plan
·43 words·1 min
Tokenization and decoding are two weakest points of current LLMs.
Current tokenization methods like BPE are poor man’s solution and can’t scale to true multi-modal.
Similarly, current decoding methods such as beam search are poor man’s replacement for the proper planning… https://x.com/labenz/status/1674790219554623492