Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Systematic Reward Gap Optimization for Mitigating VLM Hallucinations
Authors: Lehan He, Zeren Chen, Zhelun Shi, Tianyu Yu, Jing Shao, Lu Sheng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments demonstrate TPR achieves state-of-the-art performance on multiple hallucination benchmarks, outperforming previous methods by an average of 20%. |
| Researcher Affiliation | Collaboration | 1School of Software, Beihang University 2Shanghai Innovation Institute 3Shanghai AI Laboratory 4Tsinghua University helehan,czr1604,EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Topic-level Preference Rewriting (TPR) |
| Open Source Code | Yes | Codes and models: https://tpr-dpo.github.io |
| Open Datasets | Yes | Following Yu, et al. [12, 17], we curate preference data based on 7 publicly available dataset sources: VQA v2 [46], MSCOCO [47], Share GPT-4V [48], Text VQA [49], Movie Net [50], OKVQA [51] and Google Landmark v2 [52]. |
| Dataset Splits | Yes | We generate a total of 20,000 preference data instances used for alignment. For the TPR-CL variant, 12,000 instances (60%) are constructed during Warm Up stage, and the remaining 8,000 (40%) are constructed during Hard-Mining stage. |
| Hardware Specification | Yes | The policy model is fine-tuned for 1 epoch on 8 NVIDIA A100 GPUs. |
| Software Dependencies | No | The paper mentions key algorithms and optimizers like DPO [20] and Adam W [53] but does not provide specific version numbers for these software components or other dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | We use the Adam W [53] optimizer with a batch size of 8, a learning rate of 5 10 7 with the cosine decay strategy. The policy model is fine-tuned for 1 epoch on 8 NVIDIA A100 GPUs. |