[Paper Review] Can Transformers Reason About Effects of Actions?

Pratyay Banerjee, Chitta Baral+5 more|arXiv (Cornell University)|Dec 17, 2020

Topic Modeling31 references4 citations

TL;DR

This paper investigates whether transformers can reason about the effects of actions in natural language using synthetic QA datasets across four domains—Blocks World, Logistics, Dock-Worker-Robots, and a Generic Domain. Using Answer Set Programming to generate logically consistent scenarios, the authors train transformers on action-effect reasoning tasks and find strong performance on yes/no and counting questions, with transfer learning from the generic domain improving zero-shot generalization, though performance degrades on complex reasoning chains.

ABSTRACT

A recent work has shown that transformers are able to "reason" with facts and rules in a limited setting where the rules are natural language expressions of conjunctions of conditions implying a conclusion. Since this suggests that transformers may be used for reasoning with knowledge given in natural language, we do a rigorous evaluation of this with respect to a common form of knowledge and its corresponding reasoning -- the reasoning about effects of actions. Reasoning about action and change has been a top focus in the knowledge representation subfield of AI from the early days of AI and more recently it has been a highlight aspect in common sense question answering. We consider four action domains (Blocks World, Logistics, Dock-Worker-Robots and a Generic Domain) in natural language and create QA datasets that involve reasoning about the effects of actions in these domains. We investigate the ability of transformers to (a) learn to reason in these domains and (b) transfer that learning from the generic domains to the other domains.

Motivation & Objective

• To assess whether transformers can learn and generalize reasoning about the effects of actions in natural language, extending prior work on conjunctive rule reasoning.
• To evaluate zero-shot transfer learning from a generic domain to specific action domains (Blocks World, Logistics, DWR) to test generalization capabilities.
• To create and evaluate synthetic QA datasets grounded in formal action theories using Answer Set Programming to ensure logical consistency.
• To investigate the limits of transformer reasoning on open-ended and numerical questions involving action executability, fluents, and state transitions.
• To examine how model performance varies with reasoning complexity, including chain-of-thought and conditional effects.

Proposed method

• Generated synthetic action domains using Answer Set Programming (ASP), encoding action effects, executability conditions, and fluents with formal semantics.
• Defined four domains: Blocks World, Logistics, DWR (Dock-Worker-Robots), and a Generic Domain with abstract actions and fluents.
• Created QA datasets with yes/no, open-ended, and numerical answers by simulating action sequences and deriving resulting states via ASP solvers.
• Trained transformer models (e.g., BERT, RoBERTa) on training sets from each domain and evaluated on unseen test sets.
• Applied zero-shot transfer learning by fine-tuning on the generic domain and evaluating on specific domains without further fine-tuning.
• Used ASP rules to encode frame problem handling, inertia, and action preconditions, ensuring logical consistency in generated scenarios.

Experimental results

Research questions

• RQ1 Can transformers learn to reason about the effects of actions in natural language when trained on synthetic QA datasets from structured domains?
• RQ2 How well do transformers generalize zero-shot from a generic domain to specific domains like Blocks World, Logistics, and DWR?
• RQ3 What is the performance of transformers on complex reasoning tasks involving chains of actions, conditional effects, and state transitions?
• RQ4 How do model capabilities vary across different types of QA—yes/no, open-ended, and numerical (counting) questions?
• RQ5 To what extent do model predictions remain logically consistent with the formal semantics encoded in the ASP-generated datasets?

Key findings

• Transformers achieved over 95% accuracy on yes/no questions in the Blocks World and Logistics domains when trained on domain-specific data.
• On counting questions (e.g., 'how many robots are unloaded?'), models achieved over 90% accuracy in the DWR and generic domains.
• Zero-shot transfer from the generic domain to specific domains yielded 80–85% accuracy on yes/no questions, indicating partial generalization of reasoning patterns.
• Performance dropped significantly on complex reasoning chains involving multiple conditional effects or nested dependencies, suggesting limitations in compositional generalization.
• Models showed high consistency with ASP-generated ground truth on simple action-effect reasoning, but struggled with reasoning that required tracking multiple fluent changes across sequences.
• The generic domain served as an effective pre-training source, enabling better zero-shot transfer than random initialization, though performance remained below fine-tuned domain-specific models.