Contrastive decoding is the decoding approach that selects tokens by contrasting scores from a strong model and a weaker reference model to discourage generic or low-quality continuations - it aims to improve coherence and specificity in generation.
What Is Contrastive decoding?
- Definition: Token ranking method based on score differences between expert and reference model outputs.
- Core Principle: Prefer tokens where the stronger model is confident but weaker model is less supportive.
- Quality Effect: Tends to suppress bland high-frequency continuations.
- Computation Requirement: Needs two-model scoring or equivalent contrastive signals during decoding.
Why Contrastive decoding Matters
- Text Quality: Can improve informativeness and reduce generic repetitive phrasing.
- Fluency Preservation: Maintains strong-model guidance while filtering weak continuations.
- Hallucination Mitigation: Contrastive signals may discourage unstable low-confidence branches.
- Task Benefit: Useful for detailed explanations and structured long responses.
- Research Relevance: Provides alternative to pure likelihood-based ranking criteria.
How It Is Used in Practice
- Reference Model Choice: Select a smaller or weaker model with compatible tokenization and domain behavior.
- Weight Calibration: Tune contrastive strength to balance specificity and grammatical stability.
- Ablation Testing: Evaluate repetition, relevance, and factuality against baseline decoding.
Contrastive decoding is a quality-oriented alternative to standard likelihood decoding - contrastive scoring can produce more informative outputs when tuned for stability.