CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay

Natasha Butt, Blazej Manczak, Auke Wiggers, Corrado Rainone, David W. Zhang, Michaël Defferrard, Taco Cohen
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:5013-5034, 2024.

Abstract

Large language models are increasingly solving tasks that are commonly believed to require human-level reasoning ability. However, these models still perform very poorly on benchmarks of general intelligence such as the Abstraction and Reasoning Corpus (ARC). In this paper, we approach the ARC as a programming-by-examples problem, and introduce a novel and scalable method for language model self-improvement called Code Iteration (CodeIt). Our method iterates between 1) program sampling and hindsight relabeling, and 2) learning from prioritized experience replay. By relabeling the goal of an episode (i.e., the program output given input) to the output actually produced by the sampled program, our method effectively deals with the extreme sparsity of rewards in program synthesis. Applying CodeIt to the ARC dataset, we demonstrate that prioritized hindsight replay, along with pre-training and data-augmentation, leads to successful inter-task generalization. CodeIt is the first neuro-symbolic approach that scales to the full ARC evaluation dataset. Our method solves 15% of ARC evaluation tasks, achieving state-of-the-art performance and outperforming existing neural and symbolic baselines. Our code is available at https://github.com/Qualcomm-AI-research/codeit.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-butt24a, title = {{C}ode{I}t: Self-Improving Language Models with Prioritized Hindsight Replay}, author = {Butt, Natasha and Manczak, Blazej and Wiggers, Auke and Rainone, Corrado and Zhang, David W. and Defferrard, Micha\"{e}l and Cohen, Taco}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {5013--5034}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/butt24a/butt24a.pdf}, url = {https://proceedings.mlr.press/v235/butt24a.html}, abstract = {Large language models are increasingly solving tasks that are commonly believed to require human-level reasoning ability. However, these models still perform very poorly on benchmarks of general intelligence such as the Abstraction and Reasoning Corpus (ARC). In this paper, we approach the ARC as a programming-by-examples problem, and introduce a novel and scalable method for language model self-improvement called Code Iteration (CodeIt). Our method iterates between 1) program sampling and hindsight relabeling, and 2) learning from prioritized experience replay. By relabeling the goal of an episode (i.e., the program output given input) to the output actually produced by the sampled program, our method effectively deals with the extreme sparsity of rewards in program synthesis. Applying CodeIt to the ARC dataset, we demonstrate that prioritized hindsight replay, along with pre-training and data-augmentation, leads to successful inter-task generalization. CodeIt is the first neuro-symbolic approach that scales to the full ARC evaluation dataset. Our method solves 15% of ARC evaluation tasks, achieving state-of-the-art performance and outperforming existing neural and symbolic baselines. Our code is available at https://github.com/Qualcomm-AI-research/codeit.} }
Endnote
%0 Conference Paper %T CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay %A Natasha Butt %A Blazej Manczak %A Auke Wiggers %A Corrado Rainone %A David W. Zhang %A Michaël Defferrard %A Taco Cohen %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-butt24a %I PMLR %P 5013--5034 %U https://proceedings.mlr.press/v235/butt24a.html %V 235 %X Large language models are increasingly solving tasks that are commonly believed to require human-level reasoning ability. However, these models still perform very poorly on benchmarks of general intelligence such as the Abstraction and Reasoning Corpus (ARC). In this paper, we approach the ARC as a programming-by-examples problem, and introduce a novel and scalable method for language model self-improvement called Code Iteration (CodeIt). Our method iterates between 1) program sampling and hindsight relabeling, and 2) learning from prioritized experience replay. By relabeling the goal of an episode (i.e., the program output given input) to the output actually produced by the sampled program, our method effectively deals with the extreme sparsity of rewards in program synthesis. Applying CodeIt to the ARC dataset, we demonstrate that prioritized hindsight replay, along with pre-training and data-augmentation, leads to successful inter-task generalization. CodeIt is the first neuro-symbolic approach that scales to the full ARC evaluation dataset. Our method solves 15% of ARC evaluation tasks, achieving state-of-the-art performance and outperforming existing neural and symbolic baselines. Our code is available at https://github.com/Qualcomm-AI-research/codeit.
APA
Butt, N., Manczak, B., Wiggers, A., Rainone, C., Zhang, D.W., Defferrard, M. & Cohen, T.. (2024). CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:5013-5034 Available from https://proceedings.mlr.press/v235/butt24a.html.

Related Material