ECLIPTICA: A Switchable LLM Alignment Framework Based on CITA
ECLIPTICA -- A Framework for Switchable LLM Alignment via CITA - Contrastive Instruction-Tuned Alignment
AI summary
The ECLIPTICA framework allows LLM alignment to be switched at runtime via natural language instructions, and its CITA achieves 86.7% instruction alignment efficiency on five suites with Llama-3.1-8B, exceeding DPO (56.1%), GRPO (36.1%), and PPO (20.4%). CITA combines SFT with contrastive preference optimization and uses a reference model as a geometric anchor to ensure switching stability. Its benchmark contains 3000 controlled cases (300 prompts × 10 instruction types), fixing the user request and changing only the alignment instruction.