Identifying Truthful Inheritance in Family Models and Enhancing Truthfulness

ICLR 2026 Conference Submission24509 Authors

Keywords: Truthfulness, Hallucination, LVLM, LLM

TL;DR: Discovering truthful model components and adjustig model for truthfulness

Abstract:

Recent advances in large language models (LLMs) have led to emergence of specialized multimodal LLMs (MLLMs), creating distinct model families that share a common foundation language models.
This work investigates whether a core traits like truthfulness are inherited along this evolutionary trajectory. To quantify this trait, we employ linear probing on the models' internal representations. Our analysis of Vicuna and Qwen model families reveals a key finding: a strong correlation in truthfulness scores between LLMs and their finetuned MLLMs counterparts, even when they are finetuned or probed with different modalities and datasets.
Building on this findings, we propose a soft gating method using the Truthfulness score to amplify the influence of these context-truthful heads to improve the context grounding ability while preserving the contributions of other heads.
We validate our approach on base LLMs on HaluEval benchmark, demonstrating an improved ability for context truthful reasoning.
We then show that the Truthfulness scores obtained from base LLMs can be effectively transferred and applied as a soft gate to its finetuned MLLMs, demonstrating its improved performance on POPE benchmark.
The performance gain from this transfer is comparable to that obtained by probing the MLLMs directly, highlighting the potential for a unified approach to enhance truthfulness across an entire model family.
Our work demonstrates a novel method for leveraging a model's inherent, inherited traits to systematically improve its truthfulness.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 24509