Triple Preference Optimization: Achieving Better Alignment using a Single Step Optimization
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866913694621892608 |
|---|---|
| author | Saeidi, Amir Verma, Shivanshu RRV, Aswin Rasul, Kashif Baral, Chitta |
| author_facet | Saeidi, Amir Verma, Shivanshu RRV, Aswin Rasul, Kashif Baral, Chitta |
| contents | Reinforcement Learning with Human Feedback (RLHF) enhances the alignment of Large Language Models (LLMs). However, its limitations have led to the development of Direct Preference Optimization (DPO), an RL-free approach designed to overcome these shortcomings. While studies have shown that DPO improves instruction-following capabilities, it negatively impacts the reasoning ability of LLMs. Additionally, DPO is highly sensitive to judgment noise in preference datasets and the size of the training set. Although several modifications to DPO have been proposed, they still fail to fully resolve these issues. To address these limitations, we propose Triple Preference Optimization (TPO), a new preference learning method designed to enhance both reasoning and instruction-following abilities through one-step optimization. We compare TPO against DPO and its recent variants using state-of-the-art training setups, including both base and instruction-tuned models such as Mistral and Llama 3. Our evaluation covers a comprehensive range of chat-based and reasoning benchmarks. The results demonstrate that TPO achieves significant improvements over existing methods without substantially increasing response length across different dataset sizes. Specifically, TPO outperforms DPO and SimPO by up to 7.0% and 7.3% points on Arena-Hard, 12.2% and 13.3% points on MixEval-Hard, 10.4% and 10.1% points on MMLU-Pro, and 19.0% and 19.2% points on GSM8K, respectively. Furthermore, TPO achieves these improvements while requiring less data than DPO. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2405_16681 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Triple Preference Optimization: Achieving Better Alignment using a Single Step Optimization Saeidi, Amir Verma, Shivanshu RRV, Aswin Rasul, Kashif Baral, Chitta Computation and Language Reinforcement Learning with Human Feedback (RLHF) enhances the alignment of Large Language Models (LLMs). However, its limitations have led to the development of Direct Preference Optimization (DPO), an RL-free approach designed to overcome these shortcomings. While studies have shown that DPO improves instruction-following capabilities, it negatively impacts the reasoning ability of LLMs. Additionally, DPO is highly sensitive to judgment noise in preference datasets and the size of the training set. Although several modifications to DPO have been proposed, they still fail to fully resolve these issues. To address these limitations, we propose Triple Preference Optimization (TPO), a new preference learning method designed to enhance both reasoning and instruction-following abilities through one-step optimization. We compare TPO against DPO and its recent variants using state-of-the-art training setups, including both base and instruction-tuned models such as Mistral and Llama 3. Our evaluation covers a comprehensive range of chat-based and reasoning benchmarks. The results demonstrate that TPO achieves significant improvements over existing methods without substantially increasing response length across different dataset sizes. Specifically, TPO outperforms DPO and SimPO by up to 7.0% and 7.3% points on Arena-Hard, 12.2% and 13.3% points on MixEval-Hard, 10.4% and 10.1% points on MMLU-Pro, and 19.0% and 19.2% points on GSM8K, respectively. Furthermore, TPO achieves these improvements while requiring less data than DPO. |
| title | Triple Preference Optimization: Achieving Better Alignment using a Single Step Optimization |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2405.16681 |