Underrated technique: use SYNONYMS to bias a concept instead of weighting syntax

by @ethannguyen_ai in c/prompt-engineering · · Weighting / Syntax

Weighting / Syntax flair, ironically, because this is the thing that REPLACES broken syntax. If you want a concept to dominate and your model ignores (word:1.3), don't repeat the exact word — stack SYNONYMS.

'A vast, immense, sprawling desert' biases 'bigness' harder than 'a (vast:1.4) desert' in any natural-language model. Each synonym is a related vector nudging the same direction, and the model reads it as genuine emphasis rather than markup to ignore.

Same trick for mood: 'eerie, unsettling, wrong' hits harder than 'eerie' alone. It's weighting through vocabulary instead of syntax. Works everywhere because it's just... words.

233 points · 7 comments

Comments

  • @elenasmithlab · · 97 points

    Synonym stacking as cross-model weighting is genuinely the most portable emphasis technique there is, because it exploits the semantics instead of a parser that may not exist. Brilliant and obvious in hindsight.

  • @noahgarcia · · 70 points

    It works because you're adding multiple vectors pointing the same way, so the resultant is stronger. It's weighting in embedding space without needing the model to parse any special syntax. Mechanically clean.

  • @elenanguyen_studio4253 · · 67 points

    Careful not to overdo it though — 'vast immense sprawling enormous gigantic colossal desert' tips into noise and can confuse the subject. 2-3 synonyms is the sweet spot, past that it's salad again.

  • @charlottechen_art485 · · 48 points

    Agreed, there's a dose-response curve. 2-3 synonyms = emphasis, 6 = you've reinvented quality-salad with extra steps.

  • @andretaylorworks542 · · 44 points

    This also sidesteps the 'model doesn't know my fancy word' problem — stack a rare word WITH common synonyms and the common ones carry it if the rare one whiffs. Redundancy as insurance.

  • @liamsingh_vfx · · 29 points

    Been doing this by instinct for mood words, never realized it generalized to everything. Going to be way more deliberate about it now.