Reproduce “PlaSma: Making Small Language Models Better Procedural Knowledge Models for (Counterfactual) Planning” on OpenResearch

About this paper

Faeze Brahman, Chandra Bhagavatula, Valentina Pyatkin, Jena D. Hwang, Xiang Lorraine Li, Hirona J. Arai, Soumya Sanyal, Keisuke Sakaguchi and 2 more

The PLASMA framework enables smaller language models to achieve state-of-the-art performance in procedural planning and replanning, often surpassing larger teacher models. It demonstrates that models with significantly fewer parameters can compete with or even exceed GPT-3.5 and GPT-4 in complex planning tasks and improve executability by 17% in embodied AI environments.

Key insights

• Smaller language models can acquire and often exceed the procedural planning capabilities of much larger LLMs through effective symbolic knowledge distillation and sophisticated inference-time guidance.
• Combining model generation with a learned verifier (PLASMA+) significantly improves the quality and logical coherence of plans, especially for smaller models, by allowing them to self-correct during decoding.
• Multi-task distillation across various planning types (goal, constrained, replanning) enables better generalization and robustness in student models.

Summary by alphaXiv.