Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

Groc-PO: Grounded Context Preference Optimization for Truthful Multimodal LLMs

Source

arXiv

arxiv.org/abs/2607.13712v1

Read the source

Source

Plain-Language Brief

Researchers propose a new method called Grounded Context Preference Optimization (Groc-PO) to improve the truthfulness and utility of Multimodal Large Language Models (MLLMs). Current MLLMs often make errors due to unfaithful reasoning, visual hallucinations, and content fabrication. The new method addresses these issues by providing more explicit preference supervision over multiple grounded stages, which strengthens context-dependent reasoning and mitigates cross-stage error propagation.

What the Document Says

  • Multimodal Large Language Models (MLLMs) suffer from untruthfulness issues such as visual hallucinations, content fabrication, and unfaithful reasoning.
  • Alignment methods based on human preference, such as Direct Preference Optimization (DPO), have been used to address these issues but apply preference optimization at the final-answer level.
  • The new method, Grounded Context Preference Optimization (Groc-PO), aims to mitigate cross-stage error propagation by providing more explicit preference supervision over multiple grounded stages.
  • Groc-PO constructs the Grounded Context Preference Dataset (GCPD) with multi-stage preference samples around Object Grounding, Contextual Grounding, and Grounded Reasoning stages.
  • Experiments show that Groc-PO achieves improved performance in hallucination mitigation, faithful reasoning, and overall reliability compared to standard DPO and other strong baselines.

Key Concepts Mentioned

  • Multimodal Large Language Models
  • untruthfulness issues
  • Direct Preference Optimization
  • Grounded Context Preference Optimization
  • Grounded Context Preference Dataset
  • hallucination mitigation
  • faithful reasoning
  • reliability

Claims Made by the Document

  • MLLMs suffer from untruthfulness issues.
  • Standard DPO applies preference optimization at the final-answer level.
  • Groc-PO aims to mitigate cross-stage error propagation.
  • Groc-PO constructs the GCPD with multi-stage preference samples.
  • Groc-PO achieves improved performance in hallucination mitigation, faithful reasoning, and overall reliability.

Evidence Used

  • The text mentions that MLLMs suffer from untruthfulness issues such as visual hallucinations, content fabrication, and unfaithful reasoning.
  • The text states that standard DPO typically applies preference optimization at the final-answer level.
  • The text explains that Groc-PO aims to mitigate cross-stage error propagation by providing more explicit preference supervision over multiple grounded stages.
  • The text mentions that Groc-PO constructs the GCPD with multi-stage preference samples.
  • The text states that experiments show that Groc-PO achieves improved performance in hallucination mitigation, faithful reasoning, and overall reliability.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Despite the rapid progress of Multimodal Large Language Models (MLLMs), they still suffer from untruthfulness issues, such as visual hallucinations, content fabrication, and unfaithful reasoning, which substantially undermine their faithfulness and practical utility. Alignment methods based on human preference, such as Direct Preference Optimization (DPO), have been widely adopted to address these issues. However, multimodal reasoning errors often propagate across stages, and final-answer errors can often be traced to mistakes in early grounding stages, yet standard DPO typically applies preference optimization at the final-answer level. This credit-assignment challenge means that supervision for early grounding stages is indirect rather than stage-specific, making it difficult to suppress error propagation arising from grounding drift and context inconsistency. To address this, we propose Grounded Context Preference Optimization (Groc-PO), a grounded preference optimization framework for MLLMs. We further construct the Grounded Context Preference Dataset (GCPD), organizing multi-stage preference samples around three stages of Object Grounding, Contextual Grounding, and Grounded Reasoning, to capture the formation, integration, and utilization of grounded context. By introducing more explicit preference supervision over multiple grounded stages, Groc-PO strengthens context-dependent reasoning and mitigates cross-stage error propagation. Extensive experiments show that, compared with standard DPO and other strong baselines, Groc-PO achieves improved performance in hallucination mitigation, faithful reasoning, and overall reliability, supporting the value of more explicit grounded supervision for trustworthy multimodal reasoning.