Text decoding and constrained generation
12 methods in the atlas attack this one problem. They are rivals: each wins something the others do not.
Phrasings that mean this problem
Sequence generationText generationFast text generationStructured generation
information-retrieval-nlp
- Beam search decodingLength-normalized beamscanoninformation-retrieval-nlp
- Greedy decodingstandalonestandardinformation-retrieval-nlp
- Nucleus samplingTop-p truncationstandardinformation-retrieval-nlp
- Top-k samplingstandalonestandardinformation-retrieval-nlp
- Contrastive decodingstandalonespecialistinformation-retrieval-nlp
- Speculative decodingDraft-verify acceptancespecialistinformation-retrieval-nlp
llm-inference
- Contrastive searchDegeneration penaltyspecialistllm-inference
- Medusa decodingMulti-head draft predictionspecialistllm-inference
- EAGLE decodingFeature-level autoregressionspecialistllm-inference
- Lookahead decodingJacobi n-gram parallelismspecialistllm-inference
- Grammar-constrained decodingContext-free token maskingstandardllm-inference
- JSON schema decodingSchema-driven token masksstandardllm-inference