Research map / Vision-language and multimodal / Multimodal LLMs
TARS: MinMax Token-Adaptive Preference Strategy for Hallucination Reduction in MLLMs
KEJIA ZHANG, Keda TAO, Zhiming Luo, Chang Liu, Jiasheng Tang, Huan Wang
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 token-level perturbation (TP) (ablation) | TARS · AMBER | Cog: 0.4 vs 2.5 without | Improved |
| add cross-modal alignment score (CAS) token selection (ablation) | TARS · AMBER | Hal-Rate: 4.5-point increase in hallucination without it | Improved |
| add spectral preference alignment (SPA) (ablation) | TARS · OBJHal | CRi: 3.2 -> 3.7 when removed | Improved |
| replace FFT alignment with token-level contrastive alignment (ablation) | TARS · AMBER | Hal-Rate: 13.2% vs. 16.3% | Hurt |
| perturb visually relevant tokens instead of visual-agnostic ones (ablation) | DPO · AMBER | Hal-Rate: 26.4% to 31.6% | Hurt |
| random token perturbation instead of adaptive (ablation) | DPO · AMBER | Hal-Rate: 28.7% vs. 26.4%; TARS 13.2% | Mixed |
| replace DPO with TARS on 4.8k samples vs 5x LLM augmentation | LLaVA-v1.5-7B · AMBER | Hal-Rate: 16.0% vs. 13.2% | Improved |
| combine TARS with 5x LLM augmentation | TARS · AMBER | Hal-Rate: 13.2 -> 12.3 | Improved |
| lower perturbation magnitude omega to 1e-4 (ablation) | TARS · AMBER | Hal: 13.2 -> 15.4 | Hurt |
| raise perturbation magnitude omega to 1e-2 (ablation) | TARS · AMBER | Hal: 13.2 -> 20.2 | Hurt |
| raise spectral weight lambda from 0.01 to 0.20 (ablation) | TARS · AMBER | Hal: 14.8 -> 12.5 | Improved |
| raise spectral weight lambda from 0.20 to 1.00 (ablation) | TARS · AMBER | Hal: 12.5 -> 15.1 | Hurt |
Closest papers by meaning
- CHiP: Cross-modal Hierarchical Direct Preference Optimization for Multimodal LLMs (ICLR 2025 · similarity 0.93)
- Uncertainty-Aware Exploratory Direct Preference Optimization for Multimodal Large Language Models (CVPR 2026 · similarity 0.93)
- Look Carefully: Adaptive Visual Reinforcements in Multimodal Large Language Models for Hallucination Mitigation (ICLR 2026 · similarity 0.92)
- Mitigating Hallucination Through Theory-Consistent Symmetric Multimodal Preference Optimization (NeurIPS 2025 · similarity 0.92)
- Automated Multi-level Preference for MLLMs (NeurIPS 2024 · similarity 0.92)
- Mitigating Visual Hallucinations via Semantic Curriculum Preference Optimization in MLLMs (ICML 2026 · similarity 0.92)
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.
