Venue
International Conference on Machine Learning
Domain
Artificial Intelligence, Natural Language Processing
We posit that to achieve superhuman agents, future models require superhuman feedback in order to provide an adequate training signal.Current approaches commonly train reward models from human preferences, which may then be bottlenecked by human performance level, and secondly these separate frozen reward models cannot then learn to improve during LLM training.In this work, we study Self-Rewarding Language Models, where the language model itself is used via LLM-as-a-Judge prompting to provide its own rewards during training.We show that during Iterative DPO training that not only does instruction following ability improve, but also the ability to provide high-quality rewards to itself.Fine-tuning Llama 2 70B on three iterations of our approach yields a model that outperforms many existing systems on the AlpacaEval 2.0 leaderboard, including Claude 2, Gemini Pro, and GPT-4 0613.While there is much left still to explore, this work opens the door to the possibility of models that can continually improve in both axes.
This paper introduces Self-Rewarding Language Models (SRLMs) which utilize a self-improving approach to train language models by incorporating reward modeling into the learning process. The method combines instruction following and self-instruction creation to enhance both model performance and self-evaluation capabilities. By employing an Iterative Direct Preference Optimization (DPO) framework, the models are able to generate their own training data, progressively improving through multiple training iterations. The experiments demonstrate that fine-tuning the Llama 2 70B model via this method surpasses several existing language models in various benchmarks, including those on the AlpacaEval 2.0 leaderboard. The paper stresses the potential for continuous improvement in language models beyond the limitations of traditional human preference data.
This paper employs the following methods:
- LLM-as-a-Judge
- Direct Preference Optimization (DPO)
- Llama 2 70B
- Claude 2
- Gemini Pro
- GPT-4
The following datasets were used in this research:
- Open Assistant
- AlpacaEval 2.0
- MT-Bench
- ARC-Easy
- ARC-Challenge
- HellaSwag
- SIQA
- PIQA
- GSM8K
- MMLU
- OBQA
- NQ
- Accuracy
- Win Rate
- Pairwise Accuracy
- Spearman Correlation
- Kendall's τ
- ROUGE-L
- Outperformed existing models on the AlpacaEval 2.0 leaderboard
- Self-Rewarding training significantly improved instruction following and reward modeling ability
- Three iterations of training yielded superior performance compared to the baseline
The authors identified the following limitations:
- Number of GPUs: None specified
- GPU Type: NVIDIA A100 80GB
Self-Rewarding Language Models
Iterative DPO
AI Feedback
Language Model Alignment
Self-Instruction Creation