Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RIMRULE: Improving Tool-Using Language Agents via MDL-Guided Rule Learning

Xiang Gao, Yuguang Yao, Qi Zhang, Kaiwen Dong, Avinash Baidya, Ruocheng Guo, Hilaf Hasson, Kamalika Das

TL;DR

RIMRULE tackles the challenge of adapting large language models to tool usage in domain-specific settings by learning compact, interpretable rules from failure traces and injecting them at inference time. The method combines failure-driven rule induction via Explanation-Based Learning with a principled Minimum Description Length objective to consolidate a compact rule library stored in natural language and a symbolic form. Inference-time rule retrieval, with both natural-language and symbolic strategies, enables robust generalization to unseen tools and cross-model transfer without changing LLM weights. The results show that RimRule outperforms prompting-based baselines, complements finetuned models, and supports cross-LLM knowledge transfer, highlighting the practical value of modular, portable symbolic knowledge for tool-use agents.

Abstract

Large language models (LLMs) often struggle to use tools reliably in domain-specific settings, where APIs may be idiosyncratic, under-documented, or tailored to private workflows. This highlights the need for effective adaptation to task-specific tools. We propose RIMRULE, a neuro-symbolic approach for LLM adaptation based on dynamic rule injection. Compact, interpretable rules are distilled from failure traces and injected into the prompt during inference to improve task performance. These rules are proposed by the LLM itself and consolidated using a Minimum Description Length (MDL) objective that favors generality and conciseness. Each rule is stored in both natural language and a structured symbolic form, supporting efficient retrieval at inference time. Experiments on tool-use benchmarks show that this approach improves accuracy on both seen and unseen tools without modifying LLM weights. It outperforms prompting-based adaptation methods and complements finetuning. Moreover, rules learned from one LLM can be reused to improve others, including long reasoning LLMs, highlighting the portability of symbolic knowledge across architectures.

RIMRULE: Improving Tool-Using Language Agents via MDL-Guided Rule Learning

TL;DR

RIMRULE tackles the challenge of adapting large language models to tool usage in domain-specific settings by learning compact, interpretable rules from failure traces and injecting them at inference time. The method combines failure-driven rule induction via Explanation-Based Learning with a principled Minimum Description Length objective to consolidate a compact rule library stored in natural language and a symbolic form. Inference-time rule retrieval, with both natural-language and symbolic strategies, enables robust generalization to unseen tools and cross-model transfer without changing LLM weights. The results show that RimRule outperforms prompting-based baselines, complements finetuned models, and supports cross-LLM knowledge transfer, highlighting the practical value of modular, portable symbolic knowledge for tool-use agents.

Abstract

Large language models (LLMs) often struggle to use tools reliably in domain-specific settings, where APIs may be idiosyncratic, under-documented, or tailored to private workflows. This highlights the need for effective adaptation to task-specific tools. We propose RIMRULE, a neuro-symbolic approach for LLM adaptation based on dynamic rule injection. Compact, interpretable rules are distilled from failure traces and injected into the prompt during inference to improve task performance. These rules are proposed by the LLM itself and consolidated using a Minimum Description Length (MDL) objective that favors generality and conciseness. Each rule is stored in both natural language and a structured symbolic form, supporting efficient retrieval at inference time. Experiments on tool-use benchmarks show that this approach improves accuracy on both seen and unseen tools without modifying LLM weights. It outperforms prompting-based adaptation methods and complements finetuning. Moreover, rules learned from one LLM can be reused to improve others, including long reasoning LLMs, highlighting the portability of symbolic knowledge across architectures.
Paper Structure (41 sections, 4 equations, 2 figures, 8 tables, 1 algorithm)

This paper contains 41 sections, 4 equations, 2 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Method
  3. Local Rule Generation
  4. Experience Collection
  5. Rule Generation via EBL
  6. Symbolic Representation
  7. Rule Consolidation
  8. Prompt-Based
  9. MDL-Guided
  10. Objective.
  11. Model Cost.
  12. Data Cost.
  13. Greedy Consolidation.
  14. MDL Behavior.
  15. Choice of $\alpha$.
  16. ...and 26 more sections

Figures (2)

  • Figure 1: Learning reusable and interpretable rules from experience
  • Figure 2: The framework of our MDL-guided rule learning.