Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis (2307.13883v2)

Published 26 Jul 2023 in cs.LG and cs.PL

Abstract: When writing programs, people have the ability to tackle a new complex task by decomposing it into smaller and more familiar subtasks. While it is difficult to measure whether neural program synthesis methods have similar capabilities, we can measure whether they compositionally generalize, that is, whether a model that has been trained on the simpler subtasks is subsequently able to solve more complex tasks. In this paper, we characterize several different forms of compositional generalization that are desirable in program synthesis, forming a meta-benchmark which we use to create generalization tasks for two popular datasets, RobustFill and DeepCoder. We then propose ExeDec, a novel decomposition-based synthesis strategy that predicts execution subgoals to solve problems step-by-step informed by program execution at each step. When used with Transformer models trained from scratch, ExeDec has better synthesis performance and greatly improved compositional generalization ability compared to baselines. Finally, we use our benchmarks to demonstrate that LLMs struggle to compositionally generalize when asked to do programming-by-example in a few-shot setting, but an ExeDec-style prompting approach can improve the generalization ability and overall performance.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (6)
  1. Kensen Shi (15 papers)
  2. Joey Hong (23 papers)
  3. Manzil Zaheer (89 papers)
  4. Pengcheng Yin (42 papers)
  5. Charles Sutton (74 papers)
  6. Yinlin Deng (11 papers)
Citations (10)

Summary

ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis

The paper "ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis" presents a novel approach to enhance compositional generalization in neural program synthesizers. This paper addresses the challenge of whether neural models can generalize to complex programming tasks by composing simpler subtasks, an ability inherently possessed by human programmers.

Compositional Generalization in Program Synthesis

The paper explores the importance of compositional generalization in the domain of neural program synthesis, where programs are generated from high-level specifications, such as input/output examples. The paper distinguishes several compositional generalization tasks, where the objective is to determine whether models can correctly synthesize programs that involve combinations and orders of operations unseen during training. This investigation leads to the development of a meta-benchmark framework, which is applied to datasets like RobustFill and DeepCoder.

The ExeDec Method

At the core of the paper is ExeDec, a decomposition-based approach to program synthesis. ExeDec utilizes a two-step strategy: first, it predicts the subgoals within the program's execution space; second, it synthesizes corresponding subprograms to achieve those subgoals. This method leverages subgoal predictions to scaffold the synthesis of complex programs from simpler components. The method draws inspiration from human problem-solving strategies, where a large task is broken down into smaller, manageable parts.

The ExeDec approach is rigorously tested against baseline models, such as a standard Transformer model and an ablation variant, across various compositional generalization tasks. The results consistently show that ExeDec enhances compositional generalization, outperforming baseline methods significantly.

Experimental Evaluation

Experiments demonstrate that ExeDec achieves a substantial increase in performance over both standard and ablation models, particularly in scenarios that require compositional generalization. The findings are bolstered by metrics showing up to a fourfold increase in compositional generalization capability on the tested datasets.

The research also includes an insightful examination of LLMs, evaluating their few-shot prompting capabilities. Here, an ExeDec-inspired approach sees a marked improvement in the LLMs' performance on programming tasks, despite LLMs traditionally struggling with tasks that require compositional generalization in a program synthesis context.

Implications and Future Directions

This work contributes significantly to the field by offering a robust framework for improving neural program synthesis through compositional generalization. The implications are twofold: practical, enhancing the ability of neural systems to handle complex programming tasks; and theoretical, providing insights into the nature of compositionality in artificial intelligence systems.

Future research avenues could explore hierarchical decomposition strategies, as suggested in the paper’s appendix, which might facilitate even deeper compositional reasoning. Additionally, extending these insights to more diverse programming environments and integrating with broader applications could form a promising trajectory toward more capable and general-purpose program synthesis tools.

X Twitter Logo Streamline Icon: https://streamlinehq.com
Youtube Logo Streamline Icon: https://streamlinehq.com