Research map / Large language models / Alignment and preferences
Query-Routed Activation Editing with Truth-hierarchical Preference Optimization
Kewei Liao, Tianbo Wang, Yuqing Ma, Zhange Zhang, Zhicheng Geng, Xiaowei Zhao, Jiakai Wang, Xianglong Liu
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 |
|---|---|---|---|
| add DHR query-routed head selection (ablation) | no-editing baseline · TruthfulQA (open-ended) | True*Info: 62.0 -> 74.2 | Improved |
| add TPS preference steering (ablation) | no-editing baseline · TruthfulQA (open-ended) | True*Info: 62.0 -> 83.3 | Improved |
| add TPS on top of DHR (ablation) | QRAE · TruthfulQA (open-ended) | True*Info: 74.2 -> 85.1 | Improved |
| add DHR on top of TPS (ablation) | QRAE · TruthfulQA (multiple-choice) | MC1: 61.3 -> 63.3 | Improved |
| add untruthful-to-truthful loss to DHR (ablation) | QRAE TPS · TruthfulQA (open-ended) | True*Info: 74.2 -> 81.3 | Improved |
| add truthful-to-best loss to DHR (ablation) | QRAE TPS · TruthfulQA (open-ended) | True*Info: 74.2 -> 78.5 | Improved |
| add truthful-to-best loss on top of untruthful-to-truthful loss (ablation) | QRAE TPS · TruthfulQA (open-ended) | True*Info: 81.3 -> 85.1 | Improved |
| use DHR head selection instead of ITI probe-accuracy selection (ablation) | global editing vector · TruthfulQA (open-ended) | True*Info: 69.0 -> 74.2 | Improved |
| use DHR head selection instead of random selection (ablation) | global editing vector · TruthfulQA (open-ended) | True*Info: 65.8 -> 74.2 | Improved |
| use DHR head selection instead of bias-based scoring (ablation) | global editing vector · TruthfulQA (open-ended) | True*Info: 52.2 -> 74.2 | Improved |
| increase edited heads K from 16 to 48 (ablation) | QRAE · TruthfulQA (open-ended) | True*Info: 77.5 -> 85.1 | Improved |
| increase edited heads K beyond 48 up to 96 (ablation) | QRAE · TruthfulQA (open-ended) | True*Info / MC1 | No clear effect |
Closest papers by meaning
- AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation Editing (ICLR 2026 · similarity 0.92)
- Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators (AAAI 2025 · similarity 0.88)
- Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models (ICLR 2026 · similarity 0.88)
- Look Closer! An Adversarial Parametric Editing Framework for Hallucination Mitigation in VLMs (AAAI 2026 · similarity 0.88)
- Can Knowledge Editing Really Correct Hallucinations? (ICLR 2025 · similarity 0.88)
- Intervene-All-Paths: Unified Mitigation of LVLM Hallucinations across Alignment Formats (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.
