# Language Models Can Improve Event Prediction by Few-Shot Abductive Reasoning

**Authors:** Xiaoming Shi, Siqiao Xue, Kangrui Wang, Fan Zhou, James Y. Zhang, Jun Zhou, Chenhao Tan, Hongyuan Mei  
**arXiv:** [2305.16646](https://arxiv.org/abs/2305.16646)  
**Date:** 2023-05-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.16646v2/

## Key points

- Authors introduce a novel framework called LAMP to enhance event sequence model performance
- LAMP leverages the capabilities of large language models (LLMs) to assist in making predictions about real-world events
- Language model learns to suggest possible causes for future events and validates them using a search module
- Extensive experiments show that LAMP outperforms existing state-of-the-art event sequence models
- Integration of various reasoning abilities of LLMs can further enhance event sequence model performance
- Importance of exploring reasoning capabilities to drive innovations in the field is emphasized
- Framework extracts knowledge from pre-trained language models through few-shot prompting techniques, introducing domain-specific knowledge into event sequence modeling
- Method emphasizes single-step reasoning process based on observable data rather than rule-based inference involving multiple steps
- Positioned as a general framework adaptable to various prompting techniques with potential for advancing research in natural language processing and event prediction

## Summary

In this paper, the authors introduce , a novel framework that integrates  to enhance  performance. The framework leverages the  capabilities of LLMs to assist event sequence models in making predictions about real-world events. By incorporating expert-annotated demonstrations, the language model learns to suggest possible causes for proposed future events, which are then validated by a search module that identifies previous events matching these causes. A scoring function evaluates whether the retrieved events could plausibly cause the proposed future event. Through extensive experiments on challenging real-world datasets, the authors demonstrate that their LAMP framework significantly outperforms existing state-of-the-art event sequence models. The findings of this study have significant implications for future research in event sequence modeling. The authors suggest that integrating various reasoning abilities of LLMs, such as deductive reasoning, inductive reasoning, commonsense reasoning, and arithmetic reasoning, could further enhance the performance of event sequence models. They emphasize the importance of exploring these capabilities to drive impactful innovations in the field. Furthermore, the authors acknowledge support from Adobe Research and express gratitude to anonymous  reviewers for their valuable feedback. They also thank colleagues at Ant Group, UChicago, and TTIC for insightful discussions. The paper distinguishes itself from previous research focused on textual features of events by emphasizing . While related work has treated text as auxiliary features without considering reasoning about it, this study highlights how LLMs can reason about both events and their textual information effectively. By extracting knowledge from pre-trained language models through few-shot prompting techniques rather than relying on human domain experts, the framework introduces domain-specific knowledge into event sequence modeling. Additionally, the paper draws connections to logical reasoning approaches like backward chaining but emphasizes a single-step reasoning process based on observable data rather than multiple steps involving rule-based inference. The authors position their method as a general framework adaptable to various prompting techniques and highlight its potential for advancing research in natural language processing and . Overall, this study underscores the transformative potential of integrating large language models into event prediction frameworks and sets a foundation for future advancements in this rapidly evolving field.