KARMA: Knowledge graph-based Automated Reasoning Materialization and Alignment
2026-07-03 • Computation and Language
Computation and LanguageArtificial IntelligenceMachine Learning
AI summaryⓘ
The authors address a problem in template-based text generation where differences between options are too small and spread out, making it hard to learn effectively. They propose KARMA, a method that creates clearer and more distinct options by following structured paths in knowledge graphs and aligning specific parts for better comparison. Their Slot-Parallel Alignment (SPA) technique helps the model focus on important differences in specific slots during training. Tests on different scientific topics show that KARMA performs better than some standard language models and training approaches.
template-based synthesiscontrastive learningknowledge graphsentity slotssequence-level optimizationslot-level objectivemasked attentionbiomedical NLPpreference learning
Authors
Jinkyeong Choi, Chaebin Jeong, Donghyeon Park
Abstract
Template-based contrastive synthesis is scalable, but its candidates often differ only in a few entity-slots while sequence-level optimization spreads supervision over mostly shared templates. We formalize this as the Resolution Mismatch Problem and propose KARMA, which enumerates schema-constrained paths over domain knowledge graphs and verbalizes them into slot-aligned contrastive candidates. Slot-Parallel Alignment (SPA) then applies a decoupled slot-level objective to route preference supervision to discriminative entity-slots, with slot-aware masked attention serving as an optional packed-evaluation implementation. Across biomedical, computer-science, and chemistry benchmarks, KARMA outperforms base LLM and same-data SFT baselines, and compares favorably with sequence and token-level preference methods.