PaLMR: Towards Faithful Visual Reasoning via Multimodal Process Alignment

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

PaLMR (Process Alignment for Multimodal Reasoning) is a framework designed to mitigate process hallucinations in Multimodal Large Language Models (MLLMs), where models achieve correct answers despite flawed visual reasoning. It introduces a perception-aligned data layer (PaDLayer) that generates process-aware reasoning data with structured pseudo-ground-truths and verifiable visual facts. Complementing this, a process-aligned optimization layer (PaOLayer) employs a hierarchical reward fusion scheme, including a process-aware scoring function, to foster visually faithful chains-of-thought and enhance training stability. Experiments on Qwen2.5-VL-7B demonstrate that PaLMR significantly reduces reasoning hallucinations and improves visual reasoning fidelity, achieving state-of-the-art results on HallusionBench while maintaining strong performance across MMMU, MathVista, and MathVerse using approximately 4.7K high-quality samples.

Key takeaway

For AI Scientists and ML Engineers developing multimodal LLMs, focusing solely on final answer correctness in reinforcement learning risks propagating visual hallucinations. You should integrate process-level alignment techniques like PaLMR's hierarchical reward fusion, which prioritizes visual fidelity at each reasoning step. This approach, demonstrated to reduce hallucinations and improve reasoning stability, offers a path to more reliable and interpretable MLLMs, even with smaller, high-quality datasets.

Key insights

Aligning multimodal reasoning processes with visual evidence, not just final answers, is crucial to prevent hallucinations.

Principles

• Reward mechanisms must supervise reasoning process faithfulness, not just outcome correctness.
• Hierarchical reward fusion, prioritizing visual fidelity, enhances training stability.
• Pairwise comparison for visual consistency scoring is more robust than point-wise methods.

Method

PaLMR constructs perception-aligned data via learnability-based filtering and Gemini-generated pseudo-ground-truths, then optimizes with Vision-Guided GRPO (V-GRPO) using a hierarchical reward function that gates final answer rewards on visual fidelity.

In practice

• Generate structured, question-agnostic visual ground truths using powerful LLMs like Gemini.
• Implement pairwise visual fidelity scoring with an LLM-as-judge (e.g., Qwen3-30B-A3B).
• Design reward functions to strictly penalize perceptual errors, setting total reward to zero.

Topics

• Multimodal LLMs
• Visual Reasoning
• Reinforcement Learning
• Process Alignment
• LLM Hallucinations
• Reward Modeling

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Chain-of-Thought Degrades Visual Spatial Reasoning Capabilities of Multimodal LLMs — Chain-of-Thought prompting degrades visual spatial reasoning in multimodal LLMs due to shortcut learning and hallucination.
• Google researchers introduce 'faithful uncertainty,' allowing LLMs to offer best guesses instead of hallucinations — Faithful uncertainty aligns LLM linguistic expression with internal confidence to mitigate confident errors and improve agentic AI reliability.
• Week Ending 5.31.2026 — Multimodal LLM judges exhibit "Perceptual Judgment Bias," favoring plausible text over visual truth, which can be mitigated via targeted training.
• TIGER: Traceable Inference with Graph-Based Evidence Routing for Mitigating Hallucinations in Multimodal Generation — TIGER mitigates multimodal generation hallucinations via graph-based evidence routing for localized, fact-level repair.
• Reasoning for Mobile User Experience with Multimodal LLMs: Task, Benchmark, and Approach — MLLMs can diagnose complex UI/UX issues by combining task-adaptive rewards and reasoning efficiency penalties.
• International Conference on Learning Representations (ICLR) 2026 — Apple's ICLR 2026 contributions span LLM reasoning, efficient training, computer vision, and AI safety.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.