Faithful-First Reasoning, Planning, and Acting for Multimodal LLMs

TL;DR

This paper introduces Faithful-First RPA, a framework that enhances multimodal large language models by improving reasoning faithfulness and reducing hallucinations through step-wise supervision and planning.

Contribution

It proposes a novel Faithful-First RPA framework with FaithEvi and FaithAct components, achieving significant improvements in perceptual faithfulness without sacrificing accuracy.

Findings

Faithful-First RPA improves perceptual faithfulness by up to 24%.

The framework reduces hallucination behavior in multimodal reasoning.

It maintains task accuracy while enhancing faithfulness.

Abstract

Multimodal Large Language Models (MLLMs) frequently suffer from unfaithfulness, generating reasoning chains that drift from visual evidence or contradict final predictions. We propose Faithful-First Reasoning, Planning, and Acting (RPA) framework in which FaithEvi provides step-wise and chain-level supervision by evaluating the faithfulness of intermediate reasoning, and FaithAct uses these signals to plan and execute faithfulness-aware actions during inference. Experiments across multiple multimodal reasoning benchmarks show that faithful-first RPA improves perceptual faithfulness by up to 24% over prompt-based and tool-augmented reasoning frameworks, without degrading task accuracy. Our analysis shows that treating faithfulness as a guiding principle perceptually faithful reasoning trajectories and mitigates hallucination behavior. This work thereby establishes a unified framework for both evaluating and enforcing faithfulness in multimodal reasoning. Code is at https://github.com/lijunxian111/Faithful-First-RPA.