Sampling for Quality: Training-Free Reward-Guided LLM Decoding via Sequential Monte Carlo

Create a drop-in decoding library for LLMs that performs reward-guided sampling using Sequential Monte Carlo (SMC). This allows developers to steer LLM generation toward specific quality rewards without additional fine-tuning.

Suggested repo: smc-sampler

"Steer LLM output without training a single weight."

Estimated effort: 50h