diverse beam search

**Diverse beam search** is the **beam-search variant that adds diversity penalties across beams to generate multiple distinct high-quality hypotheses** - it addresses beam collapse into near-identical outputs. **What Is Diverse beam search?** - **Definition**: Multi-hypothesis decoding method that encourages dissimilarity among retained beams. - **Core Mechanism**: Applies inter-beam penalties or group constraints during token expansion. - **Output Benefit**: Produces varied candidate responses instead of minor variations of one path. - **Use Scenario**: Helpful when systems need multiple alternatives for ranking or user choice. **Why Diverse beam search Matters** - **Candidate Diversity**: Improves breadth of possible completions for downstream selection. - **Robustness**: Alternative beams can recover when top path is locally flawed. - **Product Features**: Enables multi-suggestion interfaces and reranker pipelines. - **Exploration Control**: More diverse search reduces deterministic mode collapse. - **Evaluation Value**: Exposes model uncertainty through meaningful alternative outputs. **How It Is Used in Practice** - **Group Configuration**: Partition beams into groups with diversity penalties between groups. - **Penalty Tuning**: Balance dissimilarity pressure against overall hypothesis quality. - **Selection Pipeline**: Rerank diverse outputs with task-specific scoring before final delivery. Diverse beam search is **a diversity-enhanced extension of classical beam decoding** - it improves alternative generation quality when multiple candidate outputs are needed.