Gespeichert in:
| Hauptverfasser: | , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | https://arxiv.org/abs/2507.23067 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866915482572947456 |
|---|---|
| author | Pan, Zhenyu Zhang, Yutong Zhang, Jianshu Lu, Haoran Luo, Haozheng Han, Yuwei Yu, Philip S. Li, Manling Liu, Han |
| author_facet | Pan, Zhenyu Zhang, Yutong Zhang, Jianshu Lu, Haoran Luo, Haozheng Han, Yuwei Yu, Philip S. Li, Manling Liu, Han |
| contents | Multimodal Large Language Models (MLLMs) already achieve state-of-the-art results across a wide range of tasks and modalities. To push their reasoning ability further, recent studies explore advanced prompting schemes and post-training fine-tuning. Although these techniques improve logical accuracy, they frequently leave the models' outputs burdened with pronounced social biases. Clarifying how reasoning gains interact with bias mitigation-and whether the two objectives inherently trade off-therefore remains an open and pressing research problem. Our study begins by benchmarking three bias-mitigation strategies-supervised fine-uning (SFT), knowledge distillation (KD), and rule-based reinforcement learning (RL)-under identical conditions, establishing their baseline strengths and weaknesses. Building on these results, we vary the proportion of debias-focused and reasoning-centric samples within each paradigm to chart the reasoning-versus-bias trade-off. Our sweeps reveal a consistent sweet spot: a roughly 1:4 mix trained with reinforcement learning cuts stereotype scores by 10% while retaining 88% of the model's original reasoning accuracy, offering concrete guidance for balancing fairness and capability in MLLMs. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2507_23067 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | FairReason: Balancing Reasoning and Social Bias in MLLMs Pan, Zhenyu Zhang, Yutong Zhang, Jianshu Lu, Haoran Luo, Haozheng Han, Yuwei Yu, Philip S. Li, Manling Liu, Han Artificial Intelligence Multimodal Large Language Models (MLLMs) already achieve state-of-the-art results across a wide range of tasks and modalities. To push their reasoning ability further, recent studies explore advanced prompting schemes and post-training fine-tuning. Although these techniques improve logical accuracy, they frequently leave the models' outputs burdened with pronounced social biases. Clarifying how reasoning gains interact with bias mitigation-and whether the two objectives inherently trade off-therefore remains an open and pressing research problem. Our study begins by benchmarking three bias-mitigation strategies-supervised fine-uning (SFT), knowledge distillation (KD), and rule-based reinforcement learning (RL)-under identical conditions, establishing their baseline strengths and weaknesses. Building on these results, we vary the proportion of debias-focused and reasoning-centric samples within each paradigm to chart the reasoning-versus-bias trade-off. Our sweeps reveal a consistent sweet spot: a roughly 1:4 mix trained with reinforcement learning cuts stereotype scores by 10% while retaining 88% of the model's original reasoning accuracy, offering concrete guidance for balancing fairness and capability in MLLMs. |
| title | FairReason: Balancing Reasoning and Social Bias in MLLMs |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2507.23067 |