Timezone: »

CodeRL: Mastering Code Generation through Pretrained Models and Deep Reinforcement Learning
Hung Le · Yue Wang · Akhilesh Deepak Gotmare · Silvio Savarese · Steven Chu Hong Hoi

Tue Nov 29 09:00 AM -- 11:00 AM (PST) @ Hall J #138

Program synthesis or code generation aims to generate a program that satisfies a problem specification. Recent approaches using large-scale pretrained language models (LMs) have shown promising results, yet they have some critical limitations. In particular, they often follow a standard supervised fine-tuning procedure to train a code generation model from natural language problem descriptions and ground-truth programs only. Such paradigm largely ignores some important but potentially useful signals in the problem specification such as unit tests, which thus results in poor performance when solving complex unseen coding tasks. We propose “CodeRL” to address the limitations, a new framework for program synthesis tasks through pretrained LMs and deep reinforcement learning (RL). Specifically, during training, we treat the code-generating LM as an actor network, and introduce a critic network that is trained to predict the functional correctness of generated programs and provide dense feedback signals to the actor. During inference, we introduce a new generation procedure with a critical sampling strategy that allows a model to automatically regenerate programs based on feedback from example unit tests and critic scores. For the model backbones, we extended the encoder-decoder architecture of CodeT5 with enhanced learning objectives, larger model sizes, and better pretraining data. Our method not only achieves new SOTA results on the challenging APPS benchmark, but also shows strong zero-shot transfer capability with new SOTA results on the simpler MBPP benchmark.

Author Information

Hung Le (Salesforce Research Asia)
Hung Le

Hung Le (Henry) is currently a Research Scientist at Salesforce Research, focusing on machine learning and NLP applications, such as code generation and multimodal dialogue systems.

Yue Wang (SalesForce.com)
Yue Wang

Yue Wang (王樾) is an applied scientist at Salesforce Research in Singapore. Prior to that, he obtained his Ph.D. degree at The Chinese University of Hong Kong in 2020, under the supervision of Prof. Michael R. Lyu and Prof. Irwin King. He has rich research experience in academia and industry with top-tier AI publications at ACL, EMNLP, NAACL, IJCAI, and ICASSP and internship at leading AI research labs such as Microsoft Research Asia, Tencent AI Lab, Salesforce Research, and Amazon AWS AI. His research interests include language model pretraining, code understanding and generation, and multimodality. He is the main contributor of CodeT5, a pretrained programming language model that facilitates multiple code intelligence tasks. His current passion is developing intelligent software solutions to improve programmers’ productivity.

Akhilesh Deepak Gotmare (Salesforce Research)
Silvio Savarese (Stanford University)
Steven Chu Hong Hoi (Salesforce)

More from the Same Authors