DataTerminology.comNEW
Also known as: Greedy Search
Selecting the most probable token at each generation step.
DESCRIPTION
Greedy decoding chooses the highest-probability token at every step of text generation. It is deterministic and fast but can produce repetitive or locally optimal outputs. Compared with beam search, it sacrifices global exploration for simplicity. It pairs well with constrained generation or short-form tasks where diversity is less critical. For creative tasks, sampling-based methods often yield better variety.
ALTERNATIVE DEFINITIONS
Hugging Face
Greedy search selects the token with the highest probability at each step.
Concise definition emphasizing per-step highest probability selection.
TensorFlow
Greedy decoding picks the most probable next token at each step, which can be fast but may lead to suboptimal overall sequences.
Notes speed and potential global suboptimality of greedy behavior.
RELATED TERMS
Top-k Sampling Nucleus Sampling Temperature Softmax Logit Language Model Decoding Strategies Entropy
COMPETING TERMS
RELATIONS
- Broader: Decoding
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Beam search
- Hugging Face - Comprehensive decoding strategies
- PyTorch - Logits to probabilities reference
TOP QUESTIONS
When should I use Greedy Decoding?
Use it for speed and determinism when diversity is unimportant, such as classification-like prompts or constrained outputs.
Category: How-to
Why does Greedy Decoding repeat text?
Local choices can lead to repetitive loops. Penalizing repetition or switching to sampling can mitigate this.
Category: Mistakes
How does Greedy Decoding compare to Beam Search?
Greedy is faster but less globally optimal; beam search explores multiple candidates to improve quality at higher cost.
Category: Comparison
Is Greedy Decoding deterministic?
Yes, with fixed model and inputs it yields the same output every time.
Category: Definition
Can I combine Greedy with constraints?
Yes. Constrained decoding or lexically constrained strategies can guide greedy choices to satisfy rules.
Category: Application
What risks exist with Greedy Decoding?
It may produce bland or repetitive text and miss better global sequences.
Category: Risk
Does temperature affect Greedy Decoding?
With strictly greedy selection, temperature has no effect, as the argmax token is always chosen.
Category: Misunderstanding
How to evaluate Greedy outputs?
Use task-specific metrics (e.g., ROUGE, BLEU) and human judgment; compare against sampling or beam baselines.
Category: Evaluation
Is Greedy good for code generation?
It can work for short, deterministic code tasks but often benefits from beam search or sampling with constraints.
Category: Opinion