Researchers have developed a new method called Adaptive Self-Consistency (ASC) that improves the efficiency of large language models (LLMs) by leveraging distribution-valued feedback. Unlike traditional self-consistency which treats LLMs as black boxes, ASC utilizes the log-probabilities of model outputs to identify the most likely answer with fewer sampling trajectories. The proposed ASC-D algorithm achieves this by sequentially sampling trajectories and stopping once a desired confidence level for the modal answer is reached. This approach demonstrates significant reductions in trajectory usage, with experiments on MMLU-Redux showing 46.4% to 95.6% fewer trajectories compared to existing methods, while maintaining high accuracy. AI
IMPACT This method could lead to more efficient LLM inference, reducing computational costs and enabling faster responses.
RANK_REASON The cluster contains a research paper detailing a new method for improving LLM efficiency. [lever_c_demoted from research: ic=1 ai=1.0]
- Adaptive Self-Consistency
- alphaXiv
- arXiv
- ASC Dudweiler
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- IArxiv
- MMLU-Redux
- ScienceCast
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
