Researchers have introduced stratified policies, a novel method to enhance the performance of large language models during inference. Instead of relying on repeated sampling from a single policy, this approach utilizes a population of $k$ distinct policies, each contributing one response to a collective objective. These policies are implemented using separate LoRA adapters on a shared language model backbone and are optimized for pass@$k$. The study theoretically shows that this population-based strategy can overcome limitations where individual policies have prediction errors, and empirically demonstrates consistent improvements in pass@$k$ across various reasoning tasks. AI
IMPACT This approach could lead to more efficient and effective use of computational resources during LLM inference, potentially improving response quality and speed for complex reasoning tasks.
RANK_REASON The cluster contains a research paper detailing a new method for improving LLM inference. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXivLabs
- CatalyzeX
- CORE Recommender
- DagsHub
- Gotit.pub
- Hugging Face
- IArxiv Recommender
- Influence Flower
- LoRA+
- pass@k
- ScienceCast
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →