Self-Rewarded Multimodal Coherent Reasoning Across Diverse Visual Domains

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Zhang, Jesen, Liu, Ningyuan, Cai, Kaitong, Liu, Sidi, Yang, Jing, Chen, Ziliang, Sun, Xiaofei, Wang, Keze
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866917172185399296
author Zhang, Jesen
Liu, Ningyuan
Cai, Kaitong
Liu, Sidi
Yang, Jing
Chen, Ziliang
Sun, Xiaofei
Wang, Keze
author_facet Zhang, Jesen
Liu, Ningyuan
Cai, Kaitong
Liu, Sidi
Yang, Jing
Chen, Ziliang
Sun, Xiaofei
Wang, Keze
contents Multimodal LLMs often produce fluent yet unreliable reasoning, exhibiting weak step-to-step coherence and insufficient visual grounding, largely because existing alignment approaches supervise only the final answer while ignoring the reliability of the intermediate reasoning process. We introduce SR-MCR, a lightweight and label-free framework that aligns reasoning by exploiting intrinsic process signals derived directly from model outputs. Five self-referential cues -- semantic alignment, lexical fidelity, non-redundancy, visual grounding, and step consistency -- are integrated into a normalized, reliability-weighted reward that provides fine-grained process-level guidance. A critic-free GRPO objective, enhanced with a confidence-aware cooling mechanism, further stabilizes training and suppresses trivial or overly confident generations. Built on Qwen2.5-VL, SR-MCR improves both answer accuracy and reasoning coherence across a broad set of visual benchmarks; among open-source models of comparable size, SR-MCR-7B achieves state-of-the-art performance with an average accuracy of 81.4%. Ablation studies confirm the independent contributions of each reward term and the cooling module.
format Preprint
id arxiv_https___arxiv_org_abs_2512_22545
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Self-Rewarded Multimodal Coherent Reasoning Across Diverse Visual Domains
Zhang, Jesen
Liu, Ningyuan
Cai, Kaitong
Liu, Sidi
Yang, Jing
Chen, Ziliang
Sun, Xiaofei
Wang, Keze
Computer Vision and Pattern Recognition
Artificial Intelligence
Multimodal LLMs often produce fluent yet unreliable reasoning, exhibiting weak step-to-step coherence and insufficient visual grounding, largely because existing alignment approaches supervise only the final answer while ignoring the reliability of the intermediate reasoning process. We introduce SR-MCR, a lightweight and label-free framework that aligns reasoning by exploiting intrinsic process signals derived directly from model outputs. Five self-referential cues -- semantic alignment, lexical fidelity, non-redundancy, visual grounding, and step consistency -- are integrated into a normalized, reliability-weighted reward that provides fine-grained process-level guidance. A critic-free GRPO objective, enhanced with a confidence-aware cooling mechanism, further stabilizes training and suppresses trivial or overly confident generations. Built on Qwen2.5-VL, SR-MCR improves both answer accuracy and reasoning coherence across a broad set of visual benchmarks; among open-source models of comparable size, SR-MCR-7B achieves state-of-the-art performance with an average accuracy of 81.4%. Ablation studies confirm the independent contributions of each reward term and the cooling module.
title Self-Rewarded Multimodal Coherent Reasoning Across Diverse Visual Domains
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2512.22545