Scholarlink

This paper presents an efficient symbolic approach to learning which leads to better hypothesis formation than LLMs. An agent starting without prior knowledge learns by receiving new natural language statements from a teacher. Each sentence is understood only in the context of previously received statements. Hypothesis formation happens naturally through a process we call coduction, which is ultimately grounded in a primitive instinct to compress textual knowledge. As a side-effect of the attempt to form hypotheses, the agent becomes curious about related facts, and asks the teacher questions, potentially requiring exploration/experimentation in her world. A highly concurrent Actor-based implementation is presented, where a dedicated actor actively manages the relationships of each concept with other concepts, forming a sort of symbolic connectionist network. The arrival of a new statement leads to a multitude of messages flowing concurrently through the network of connected actors, exploring new hypotheses and pursuing new questions. Results of both quantitative and qualitative evaluations are presented. Of the two quantitative evaluations, the first uses a social media connections dataset with only one type of symmetric relationships, and the second uses a synthetic database of familial ralations with varied relationships. In the two evaluations, given only 20% of the statements in the ground truth, the agent discovered additional 68.4% and 44.9% (respectively) of the ground truth through fact checks and hypothesis testing. In the family network case, some types of relationships not introduced in the training were discovered. Finally, given identical training data, coduction – using fact-checking and hypothesis generation – was found to outperform a fine-tuned LLM. Most notably, for the family relations dataset, the LLM consistently failed to generalize to identify 7.15% of the facts that coduction learned. The qualitative evaluation learned using statements adapted from a wildlife website for children. Given 51 statements, the agent learned 63 more.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Learning with Hypothesis Formation and Curiosity: An Actors Approach

  • 章节

  • 发表日期：

    2026年

查看全文
  • Nadeem Jamali,
  • Aditya Phadke,
  • Zhe Chen

摘要

This paper presents an efficient symbolic approach to learning which leads to better hypothesis formation than LLMs. An agent starting without prior knowledge learns by receiving new natural language statements from a teacher. Each sentence is understood only in the context of previously received statements. Hypothesis formation happens naturally through a process we call coduction, which is ultimately grounded in a primitive instinct to compress textual knowledge. As a side-effect of the attempt to form hypotheses, the agent becomes curious about related facts, and asks the teacher questions, potentially requiring exploration/experimentation in her world. A highly concurrent Actor-based implementation is presented, where a dedicated actor actively manages the relationships of each concept with other concepts, forming a sort of symbolic connectionist network. The arrival of a new statement leads to a multitude of messages flowing concurrently through the network of connected actors, exploring new hypotheses and pursuing new questions. Results of both quantitative and qualitative evaluations are presented. Of the two quantitative evaluations, the first uses a social media connections dataset with only one type of symmetric relationships, and the second uses a synthetic database of familial ralations with varied relationships. In the two evaluations, given only 20% of the statements in the ground truth, the agent discovered additional 68.4% and 44.9% (respectively) of the ground truth through fact checks and hypothesis testing. In the family network case, some types of relationships not introduced in the training were discovered. Finally, given identical training data, coduction – using fact-checking and hypothesis generation – was found to outperform a fine-tuned LLM. Most notably, for the family relations dataset, the LLM consistently failed to generalize to identify 7.15% of the facts that coduction learned. The qualitative evaluation learned using statements adapted from a wildlife website for children. Given 51 statements, the agent learned 63 more.