This research introduces a novel probabilistic framework for scaling LLMs at inference time rather than increasing model size, addressing diminishing returns from traditional scaling approaches.

• Frames LLM output generation as a probabilistic inference problem rather than a search problem
• Utilizes particle-based Monte Carlo methods to explore multiple inference paths efficiently
• Reduces vulnerability to reward hacking common in existing inference-time scaling approaches
• Demonstrates a more computationally efficient approach to improving LLM performance

Engineering significance: This work opens new pathways for enhancing LLM capabilities without requiring larger models or more training data, potentially making advanced AI more accessible and sustainable for practical applications.

A Probabilistic Inference Approach to Inference-Time Scaling of LLMs using Particle-Based Monte Carlo Methods