no-repeat n-gram

**No-repeat n-gram** is the **hard decoding constraint that blocks generation of any n-gram already produced earlier in the output** - it is a strict safeguard against repeated phrase loops. **What Is No-repeat n-gram?** - **Definition**: Constraint rule that forbids duplicate n-token sequences during generation. - **Mechanism**: At each step, candidate tokens that would recreate an existing n-gram are masked out. - **Parameter**: The n value controls strictness, with larger n allowing more flexibility. - **Applicability**: Works with beam search and sampling-based decoding flows. **Why No-repeat n-gram Matters** - **Degeneration Control**: Prevents common repetitive loops in long-form generation. - **Readability**: Reduces duplicated clauses and improves narrative flow. - **Deterministic Safety**: Provides hard guarantees where soft penalties are insufficient. - **Production Reliability**: Useful for public-facing assistants where repetition is highly visible. - **Quality Consistency**: Stabilizes output under high-entropy sampling settings. **How It Is Used in Practice** - **Choose N Carefully**: Start with moderate n values and validate against fluency regression. - **Domain Testing**: Check technical tasks where exact phrase reuse may be necessary. - **Combined Policies**: Use with light penalties instead of excessive hard blocking where possible. No-repeat n-gram is **a strong structural guardrail for repetitive generation failures** - it is highly effective but must be tuned to avoid over-constraining valid output.