Searching Latent Program Spaces

2024-11-13

arXiv.org

AI Summary

Program synthesis faces fundamental challenges in generalization, out-of-distribution (OOD) adaptability, and scalability due to combinatorial explosion. Method: We propose Latent Program Networks (LPNs), the first framework to embed discrete programs into a continuous, differentiable latent space, enabling end-to-end differentiable program induction. LPNs unify gradient-based optimization during training with gradient-guided adaptive search at test time—overcoming efficiency and expressivity limitations of traditional sampling- and symbolic-reasoning-based approaches. The method integrates continuous-space modeling, implicit program representation, and benchmark-specific training on ARC-AGI. Contribution/Results: Experiments demonstrate that LPNs significantly outperform baselines lacking test-time adaptation on ARC-AGI, achieving breakthrough improvements in both cross-task generalization and OOD task adaptation—establishing a new state of the art in differentiable program synthesis.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Search and Optimization: Learning to Search
• Reasoning under Uncertainty: Probabilistic Programming

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Program synthesis methods aim to automatically generate programs restricted to a language that can explain a given specification of input-output pairs. While purely symbolic approaches suffer from a combinatorial search space, recent methods leverage neural networks to learn distributions over program structures to narrow this search space significantly, enabling more efficient search. However, for challenging problems, it remains difficult to train models to perform program synthesis in one shot, making test-time search essential. Most neural methods lack structured search mechanisms during inference, relying instead on stochastic sampling or gradient updates, which can be inefficient. In this work, we propose the Latent Program Network (LPN), a general algorithm for program induction that learns a distribution over latent programs in a continuous space, enabling efficient search and test-time adaptation. We explore how to train these networks to optimize for test-time computation and demonstrate the use of gradient-based search both during training and at test time. We evaluate LPN on ARC-AGI, a program synthesis benchmark that evaluates performance by generalizing programs to new inputs rather than explaining the underlying specification. We show that LPN can generalize beyond its training distribution and adapt to unseen tasks by utilizing test-time computation, outperforming algorithms without test-time adaptation mechanisms.

Problem

• Efficiently acquire new skills and generalize beyond training distributions
• Overcome scaling issues in program synthesis due to large combinatorial spaces
• Combine adaptability of symbolic approaches with scalability of neural methods

Innovation

• Latent Program Network integrates test-time search
• LPN combines symbolic adaptability with neural scalability
• LPN searches compact latent space without predefined DSLs