Controlling Multimodal LLMs via Reward-guided Decoding

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Mañas, Oscar, D'Oro, Pierluca, Sinha, Koustuv, Romero-Soriano, Adriana, Drozdzal, Michal, Agrawal, Aishwarya
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918125700644864
author Mañas, Oscar
D'Oro, Pierluca
Sinha, Koustuv
Romero-Soriano, Adriana
Drozdzal, Michal
Agrawal, Aishwarya
author_facet Mañas, Oscar
D'Oro, Pierluca
Sinha, Koustuv
Romero-Soriano, Adriana
Drozdzal, Michal
Agrawal, Aishwarya
contents As Multimodal Large Language Models (MLLMs) gain widespread applicability, it is becoming increasingly desirable to adapt them for diverse user needs. In this paper, we study the adaptation of MLLMs through controlled decoding. To achieve this, we introduce the first method for reward-guided decoding of MLLMs and demonstrate its application in improving their visual grounding. Our method involves building reward models for visual grounding and using them to guide the MLLM's decoding process. Concretely, we build two separate reward models to independently control the degree of object precision and recall in the model's output. Our approach enables on-the-fly controllability of an MLLM's inference process in two ways: first, by giving control over the relative importance of each reward function during decoding, allowing a user to dynamically trade off object precision for recall in image captioning tasks; second, by giving control over the breadth of the search during decoding, allowing the user to control the trade-off between the amount of test-time compute and the degree of visual grounding. We evaluate our method on standard object hallucination benchmarks, showing that it provides significant controllability over MLLM inference, while consistently outperforming existing hallucination mitigation methods.
format Preprint
id arxiv_https___arxiv_org_abs_2508_11616
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Controlling Multimodal LLMs via Reward-guided Decoding
Mañas, Oscar
D'Oro, Pierluca
Sinha, Koustuv
Romero-Soriano, Adriana
Drozdzal, Michal
Agrawal, Aishwarya
Computer Vision and Pattern Recognition
Artificial Intelligence
Computation and Language
Machine Learning
As Multimodal Large Language Models (MLLMs) gain widespread applicability, it is becoming increasingly desirable to adapt them for diverse user needs. In this paper, we study the adaptation of MLLMs through controlled decoding. To achieve this, we introduce the first method for reward-guided decoding of MLLMs and demonstrate its application in improving their visual grounding. Our method involves building reward models for visual grounding and using them to guide the MLLM's decoding process. Concretely, we build two separate reward models to independently control the degree of object precision and recall in the model's output. Our approach enables on-the-fly controllability of an MLLM's inference process in two ways: first, by giving control over the relative importance of each reward function during decoding, allowing a user to dynamically trade off object precision for recall in image captioning tasks; second, by giving control over the breadth of the search during decoding, allowing the user to control the trade-off between the amount of test-time compute and the degree of visual grounding. We evaluate our method on standard object hallucination benchmarks, showing that it provides significant controllability over MLLM inference, while consistently outperforming existing hallucination mitigation methods.
title Controlling Multimodal LLMs via Reward-guided Decoding
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Computation and Language
Machine Learning
url https://arxiv.org/abs/2508.11616