KARMA uses knowledge graphs to solve resolution mismatch in contrastive synthesis

The authors address the Resolution Mismatch Problem in template-based contrastive synthesis, where candidates often differ only in a few entity-slots while sequence-level optimization spreads supervision over shared templates. They propose KARMA, a method that enumerates schema-constrained paths over domain knowledge graphs and verbalizes them into slot-aligned contrastive candidates.

• The system employs Slot-Parallel Alignment (SPA) to apply a decoupled slot-level objective, routing preference supervision specifically to discriminative entity-slots.
• Slot-aware masked attention is included as an optional packed-evaluation implementation.
• KARMA outperforms base LLM and same-data SFT baselines across biomedical, computer-science, and chemistry benchmarks.
• The approach compares favorably with existing sequence and token-level preference methods.

KARMA improves the effectiveness of contrastive synthesis by ensuring that supervision is directed toward the specific entities that distinguish between candidates.