Researchers have introduced ECLIPTICA, a novel framework that enables dynamic, instruction-driven alignment for large language models. Unlike traditional methods that freeze alignment after training, ECLIPTICA allows for runtime control over model behavior, such as stance, refusal boundaries, and verbosity, by interpreting natural language alignment instructions. This approach is facilitated by CITA (Contrastive Instruction-Tuned Alignment), which combines supervised fine-tuning with contrastive preference optimization, ensuring stable and traversable alignment regimes. In evaluations on Llama-3.1-8B, CITA demonstrated an 86.7% instruction-alignment efficiency on the ECLIPTICA benchmark, significantly outperforming methods like DPO, GRPO, and PPO. AI
IMPACT Enables more flexible and dynamic control over LLM behavior, potentially improving safety and adaptability in real-world applications.
RANK_REASON The cluster contains a research paper detailing a new framework and method for LLM alignment. [lever_c_demoted from research: ic=1 ai=1.0]
- Conditional safety margins for less conservative peak local SAR assessment: A probabilistic approach
- Direct Preference Optimization
- ECLIPTICA
- Grpo
- Length control of the metaphase spindle
- LITMUS
- Llama-3.1:8b
- Proximal Policy Optimization
- supervised fine-tuning
- TruthfulQA
- Vinija Jain
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
