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

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Markovic-Voronov, Jelena, Zhu, Wenhui, Long, Bo, Wang, Zhipeng, Gupta, Suyash, Behdin, Kayhan, Chen, Bee-Chung, Agarwal, Deepak
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915943145275392
author Markovic-Voronov, Jelena
Zhu, Wenhui
Long, Bo
Wang, Zhipeng
Gupta, Suyash
Behdin, Kayhan
Chen, Bee-Chung
Agarwal, Deepak
author_facet Markovic-Voronov, Jelena
Zhu, Wenhui
Long, Bo
Wang, Zhipeng
Gupta, Suyash
Behdin, Kayhan
Chen, Bee-Chung
Agarwal, Deepak
contents We introduce a principled probabilistic framework for reward-guided decoding in large language models, addressing the limitations of standard decoding methods that optimize token-level likelihood rather than sequence-level quality. Our method defines a reward-augmented target distribution over complete sequences by combining model transition probabilities with prefix-dependent reward potentials. Importantly, the approach is training-free: it leaves model weights unchanged and instead modifies the inference distribution via reward potentials, with all gains arising purely from inference-time sampling. To sample from this distribution, we develop Sequential Monte Carlo algorithms, including a computationally efficient prefix-only variant and a lookahead variant whose intermediate targets match the exact marginals of the full sequence distribution. The framework also integrates resample-move updates with Metropolis-Hastings rejuvenation and supports block-wise generation, subsuming common decoding strategies such as temperature sampling and power-tempered objectives. Empirical results across three 7B models show significant gains. On code generation (HumanEval), our method improves base performance by up to 54.9% and surpasses the strongest sampling baselines by 9.1%-15.3%. On mathematical reasoning (MATH500), it achieves gains of up to 8.8%. Notably, it reaches 87.8% on HumanEval and 78.4% on MATH500 with Qwen2.5-7B, consistently outperforming the reinforcement learning method GRPO.
format Preprint
id arxiv_https___arxiv_org_abs_2604_16453
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Sampling for Quality: Training-Free Reward-Guided LLM Decoding via Sequential Monte Carlo
Markovic-Voronov, Jelena
Zhu, Wenhui
Long, Bo
Wang, Zhipeng
Gupta, Suyash
Behdin, Kayhan
Chen, Bee-Chung
Agarwal, Deepak
Machine Learning
Artificial Intelligence
We introduce a principled probabilistic framework for reward-guided decoding in large language models, addressing the limitations of standard decoding methods that optimize token-level likelihood rather than sequence-level quality. Our method defines a reward-augmented target distribution over complete sequences by combining model transition probabilities with prefix-dependent reward potentials. Importantly, the approach is training-free: it leaves model weights unchanged and instead modifies the inference distribution via reward potentials, with all gains arising purely from inference-time sampling. To sample from this distribution, we develop Sequential Monte Carlo algorithms, including a computationally efficient prefix-only variant and a lookahead variant whose intermediate targets match the exact marginals of the full sequence distribution. The framework also integrates resample-move updates with Metropolis-Hastings rejuvenation and supports block-wise generation, subsuming common decoding strategies such as temperature sampling and power-tempered objectives. Empirical results across three 7B models show significant gains. On code generation (HumanEval), our method improves base performance by up to 54.9% and surpasses the strongest sampling baselines by 9.1%-15.3%. On mathematical reasoning (MATH500), it achieves gains of up to 8.8%. Notably, it reaches 87.8% on HumanEval and 78.4% on MATH500 with Qwen2.5-7B, consistently outperforming the reinforcement learning method GRPO.
title Sampling for Quality: Training-Free Reward-Guided LLM Decoding via Sequential Monte Carlo
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2604.16453