Research map / Vision-language and multimodal / Multimodal LLMs
HERO: Enhancing Multimodal Faithfulness via Dynamic Entropy-Aware Reinforcement Learning
Xiongfeng Yang, Qingan Zhang, Yuheng Zhang, Kun-Yu Lin, Jian-Fang Hu, Dongmei Jiang, WEISHI ZHENG
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 entropy-aware weighting to GRPO (ablation) | vanilla GRPO · THRONE | Acc: 80.6 -> 82.0 | Improved |
| add dynamic offensive/defensive balancing (ablation) | GRPO + entropy weighting · THRONE | Acc: 82.0 -> 82.8 | Improved |
| add variance-gated sampling to GRPO (ablation) | vanilla GRPO · THRONE | Acc: 80.6 -> 80.6 | No clear effect |
| replace random sampling with variance gating at same data budget (ablation) | GRPO · THRONE | Acc: 79.2 -> 80.6 | Improved |
| train on random 60% of data instead of full data (ablation) | GRPO · THRONE | Acc: 80.6 -> 79.2 | Hurt |
| add HERO versus vanilla GRPO (7B) | Qwen2.5-VL-7B · POPE | Acc: 89.2 -> 90.8 | Improved |
| add HERO versus vanilla GRPO (3B) | Qwen2.5-VL-3B · THRONE | Acc: 82.2 -> 82.8 | Improved |
| HERO versus offline HA-DPO (7B) | Qwen2.5-VL-7B · POPE | Acc: 87.2 -> 90.8 | Improved |
| HERO versus vanilla GRPO on AMBER hallucination rate (7B) | Qwen2.5-VL-7B · AMBER-G | CHAIR: 6.0 -> 6.3 | Hurt |
| HERO versus vanilla GRPO on AMBER coverage (7B) | Qwen2.5-VL-7B · AMBER-G | Cover: 59.0 -> 61.2 | Improved |
| tune with CoT prompting | vanilla baseline · POPE | Acc: 89.6 -> 84.4 | Hurt |
| tune with CoT data (Tuned-CoT) | vanilla baseline · POPE | Acc: 89.6 -> 85.8 | Hurt |
Closest papers by meaning
- BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (NeurIPS 2026 · similarity 0.92)
- Reasoning Models Hallucinate More: Factuality-Aware Reinforcement Learning for Large Reasoning Models (NeurIPS 2025 · similarity 0.91)
- Ground What You See: Hallucination-Resistant MLLMs via Caption Feedback, Diversity-Aware Sampling, and Conflict Regularization (AAAI 2026 · similarity 0.91)
- Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.90)
- Robust Multimodal Large Language Models Against Modality Conflict (ICML 2025 · similarity 0.90)
- VES-RFT: Rewarding Visual Evidence Sensitivity to Mitigate Hallucinations in Large Vision-Language Models (CVPR 2026 · similarity 0.90)
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.
