Hasty Briefsbeta

Bilingual

SEM-CTRL: Semantically Controlled Decoding

a year ago
  • #Semantic Control
  • #Decoding
  • #LLM
  • Introduces SEM-CTRL, a method for enforcing syntactic and semantic correctness in LLM outputs.
  • Uses token-level MCTS guided by syntactic and semantic constraints.
  • Constraints are expressed using Answer Set Grammars, incorporating background knowledge.
  • Guarantees correct completions for any off-the-shelf LLM without fine-tuning.
  • Evaluated on tasks like grammar synthesis, combinatorial reasoning, and planning.
  • Small LLMs with SEM-CTRL outperform larger models and state-of-the-art reasoning models.
  • Ensures solution correctness while improving efficiency.