Researchers have introduced Cobras, a novel method for controlling large language models at inference time. Unlike previous techniques that heuristically optimize objectives, Cobras is derived from a principled optimization problem using a Schrödinger Bridge formulation on the residual-stream hypersphere. This approach results in query-adaptive steering directions, which empirically show improved performance across various alignment axes and avoid the out-of-distribution degradation seen in prior methods. AI
IMPACT Provides a more principled and adaptive approach to controlling LLM behavior at inference time, potentially improving alignment and reducing performance degradation.
RANK_REASON The cluster contains a research paper detailing a new method for controlling LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
