The Truth Stays in the Family: Enhancing Contextual Grounding via Inherited Truthful Heads in Model Lineages

2026-06-14

AI Summary

This study investigates whether large language models and their multimodal variants retain the contextual truthfulness of their base models after instruction fine-tuning or adaptation. By quantifying truthfulness scores at the attention head level, the work reveals—for the first time—that truthfulness is highly inherited across model lineages through attention weights. Building on this insight, the authors propose TruthProbe, a soft gating strategy that enhances the contribution of truthfulness-preserving attention heads without requiring retraining. Experiments demonstrate that TruthProbe improves contextual truthfulness on HaluEval and significantly mitigates multimodal hallucinations on POPE and CHAIR. Moreover, the truthfulness scores of base models effectively transfer to their fine-tuned descendants, underscoring the stability of this property throughout model adaptation.

Abstract

Recent advances in large language models (LLMs) have produced many specialized multimodal LLMs (MLLMs) that share common foundational LLMs, forming distinct model lineages. It remains unclear whether a fundamental behavioral link exists between the foundational LLMs and downstream variants. We investigate this question by quantifying head-level context-truthfulness scores. Across diverse LLM and MLLM lineages, including Vicuna-, Qwen2.5-, LLaMA2-, and Mistral-based models, we find that Truth Scores are strongly preserved within model families, even after instruction tuning or multimodal adaptation. We further show that this inheritance is consistent with attention-head weight preservation, and that context-truthful heads attend to query-relevant evidence. Building on this finding, we propose TruthProbe, a soft-gating strategy that amplifies context-truthful heads while preserving other head contributions. TruthProbe improves contextual truthfulness on HaluEval and reduces multimodal hallucination on POPE and CHAIR, with base-LLM Truth Scores transferring effectively to their fine-tuned LLM and MLLM descendants. Code is available at https://github.com/miso-choi/TruthProbe.

Problem

• contextual truthfulness
• model lineage
• truth inheritance
• multimodal LLMs
• hallucination

Innovation

• contextual truthfulness
• model lineage
• attention heads
• TruthProbe
• multimodal hallucination