Research map / Large language models / Instruction and fine-tuning
The Truth Stays in the Family: Enhancing Contextual Truthfulness via Inherited Heads in Model Lineages
Miso Choi, Seonga Choi, Mincheol Kwon, Woosung Joung, Jinkyu Kim, Jungbeom Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply Truth Score soft head gating (TruthProbe) to the same LLM | Vicuna-7B · HaluEval | F1: 13.37 -> 29.15 | Improved |
| apply base-LLM Truth Scores as gate to finetuned Vicuna-7B | Vicuna-7B · HaluEval | F1: 13.37 -> 57.17 | Improved |
| transfer base-LLM Truth Scores to Qwen2.5-7B-Instruct | Qwen2.5-7B-Instruct · HaluEval | F1: 16.29 -> 17.24 | Improved |
| transfer base-LLM Truth Scores to LLaVA-1.5 | LLaVA-1.5 · POPE (COCO) | accuracy: 86.9 -> 86.7 | No clear effect |
| transfer base-LLM Truth Scores to Qwen2.5-VL-Omni | Qwen2.5-VL-Omni · POPE (COCO) | accuracy: 85.1 -> 87.3 | Improved |
| transfer base-LLM Truth Scores to LLaVA-1.5 | LLaVA-1.5 · CHAIR | CHAIRS: 23.00 -> 17.40 | Improved |
| transfer base-LLM Truth Scores to Qwen2.5-VL-Omni | Qwen2.5-VL-Omni · CHAIR | CHAIRI: 5.26 -> 5.94 | Hurt |
| derive Truth Scores from the MLLM itself instead of its base LLM | LLaVA-1.5 · CHAIR | CHAIRS: 17.40 (base LLM) vs 21.60 (MLLM) | No clear effect |
| replace Truth Score gate with random head gating (ablation) | LLaVA-1.5 · POPE (COCO) | accuracy: 86.9 -> 86.1 | Hurt |
| replace Truth Score gate with random head gating (ablation) | LLaVA-NeXT · POPE (COCO) | accuracy: 87.7 -> 87.1 | Hurt |
| compare TruthProbe to ITI on parametric-knowledge task | LLaMA2-7B-Chat · TruthfulQA generation | truthfulness: 56.40 -> 56.91 (ITI 57.64) | No clear effect |
| TruthProbeFT LLM on TruthfulQA | LLaMA2-7B-Chat · TruthfulQA generation | informativeness: 25.56 -> 29.02 | Improved |
Closest papers by meaning
- Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension (ICML 2024 · similarity 0.88)
- DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models (ICLR 2024 · similarity 0.88)
- Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning (AAAI 2024 · similarity 0.88)
- LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations (ICLR 2025 · similarity 0.88)
- To Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts (ICLR 2025 · similarity 0.88)
- ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
