Get started
Work

Interpretable, Modular, and Structured Multi-step Reasoning over Natural Language

Year: 2024
Type: dissertation
Abstract: The literature on human cognition and psychology famously talks about a Dual Process Theory, suggesting that humans possess two modes of reasoning -- a System 1 mode that is fast, automatic, and unconscious and a System 2 mode that is slow, deliberate, controlled, and conscious. For humans, completing the phrase 'War and ...' is a fast and implicit process while solving a math word problem requires multiple steps of deduction and logical reasoning. Large Language Models have made great strides across a lot of different System 1 NLP tasks but complex System 2 reasoning continues to be a challenge. This thesis presents interpretable, modular, and structured methods for building language agents that can accomplish System 2 multi-step reasoning tasks. First, I develop a transformer-based model for deductive reasoning over explicit knowledge that can jointly answer questions and generate proof graphs supporting the answer. I show that jointly generating answers and proofs not only improves the interpretability of a question-answering system but also improves the QA accuracy in zero-shot settings. Second, I extend my previous work to develop a multi-proof generation model capable of generating multiple proof graphs for further enhancing the interpretability of question-answering systems. I show that compared to a single-proof generation model, a multi-proof generation model generalizes better to questions requiring higher depths of reasoning where multiple proofs are more frequent. Third, I shift my focus to multi-step reasoning over implicit knowledge and propose a task, benchmark, and models for generating explanation graphs for structured commonsense reasoning. I show that state-of-the-art models struggle to generate meaningful explanation graphs, which has since encouraged future work on structured commonsense reasoning. Fourth, I study multi-step reasoning for text generation in which I generalize the notion of structured tree- and graph-based explanations to develop Summarization Programs, an interpretable modular framework for abstractive summarization that encodes the step-by-step generative process of a summary with neural modular trees. Summarization Programs improve the interpretability of summarization models and constitute a promising step toward faithful abstractive summarization. Finally, I move beyond reasoning with a single agent and explore multi-step reasoning in multi-agent systems in which multiple models interact with each other in natural language either with the goal of (1) teaching reasoning to weaker agents or (2) collaboratively solving reasoning problems. In particular, I develop a collaborative framework between multiple LLMs and a structured distillation framework that distills multi-LLM interactions into a single efficient student model. Both of these multi-agent frameworks demonstrate the utility of collective reasoning for complex 'System 2' tasks. As we envision a future of building autonomous AI agents that can follow language instructions to carry out complex tasks, this thesis makes progress in building such agents that use language as a vehicle for reasoning, communication, and collaboration.
DOI: 10.17615/7bh5-d839
Language: English
Open Access status: green
OpenAlex ID: W6962693494
Cites:
Cited by:
Related to: