The Truth Stays in the Family: Enhancing Contextual Grounding via Inherited Truthful Heads in Model Lineages
For practitioners building specialized MLLMs from foundation models, this work provides a method to improve factual grounding without retraining, though the finding of head inheritance is incremental.
The paper discovers that context-truthfulness scores of attention heads are preserved across LLM and MLLM lineages, even after fine-tuning. It proposes TruthProbe, a soft-gating method that amplifies truthful heads, reducing hallucination on HaluEval, POPE, and CHAIR benchmarks.
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.