Natural Language Tools - A Natural Language Approach to Tool Calling In Large Language Agents

Sep 21, 2026

Uploaded by Jiven

Natural Language Tools (NLT) introduces a framework that replaces structured JSON tool calling with natural-language tool decisions in large language models. By separating tool selection from response generation, NLT reduces task interference and format-driven constraints that harm tool-call quality. Across 10 models and 6,400 trials in customer service and mental health, NLT raises tool-calling accuracy by 18.4 percentage points and cuts output variance by 70%, remaining robust under prompt perturbations.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages31

Reading Time1 h 18 min

AccessFree to download

Abstract

1. Introduction

Structured tool calling and its drawbacks

Natural Language Tools (NLT) framework

Natural Language Tools: A Natural Language Approach to Tool Calling In Large Language Agents Reid T. Johnson Michelle D. Pain Jordan D. West PokketCoach arXiv :2510 . 14453v1 [ cs .CL] 16 Oct 2025 Abstract We present Natural Language Tools (NLT), a framework that replaces programmatic JSON tool calling in large language models (LLMs) with natural language outputs. By decoupling toolselection from response generation, NLT eliminates task interference and format constraints that degrade tool call performance. When evaluated across 10 models and 6,400 trials spanning customer service and mental health domains, NLT improves tool calling accuracy by 18.4 percentage points while reducing output variance by 70% . Open-weight models see the largest gains, surpassing flagship closed-weight alternatives, with implications for model training in both reinforcement learning and supervised fine-tuning stages. These improvements persist under prompt perturbations and extend tool-calling capabilities to models lacking native support. 1. Introduction Large language models (LLMs) have dramatically expanded their capabilities through tool calling, the ability to invoke external functions such as web search, information retrieval, and code execution (Qu et al., 2024 ; Li, 2024) . Beginning with early systems like WebGPT (Nakano et al., 2021), these capabilities have become foundational to modern agentic architectures, enabling LLM-based agents to interface with other systems to orchestrate complex actions. A variety of benchmarks have been developed to measure tool calling efficacy (Huang et al., 2024a ; Yao et al., 2024), while model providers consistently highlight even modest improvements in accuracy (OpenAI, 2025 ; Anthropic, 2025) . However, even state-of-the-art models frequently fail to make accurate tool calls (Maekawa et al., 2024) . Current implementations typically require models to emit structured outputs conforming to rigid function schemas (i.e., JSON or XML) . We refer to this as “structured tool calling.” Figure 1 demonstrates such an output. { "tool_calls": [ { "name": "check_talk_to_a_human", "description": "Used when the user requests to speak with a human representative, agent, or support person . This tool checks whether escalation to a human is possible and initiates the handoff process if appropriate. " } ] } Figure 1 . An example of a JSON-formatted tool call for a “talk toa human” request within a customer service scenario. This structured approach has become widely adopted, beginning with releases such as Toolformer (Schick et al., 2023) and extending to most major model providers (OpenAI, 2023 ; Google, 2023 ; Anthropic, 2024) . Yet as tool calling becomes an increasingly critical component of LLMpowered agents, growing evidence suggests the underlying structure may come with significant drawbacks (Gupta et al., 2024 ; Tam et al., 2024 ; Wang et al., 2025) . Structured formats require models to simultaneously handle multiple competing demands, such as: understanding the query, selecting appropriate tools, adhering to format constraints, and generating a response. Gupta et al. (2024) demonstrated that such task interference can degrade performance significantly, with some models experiencing more than 20% reductions in accuracy. Tam et al. (2024) provided another striking example: requiring JSON output reduced response accuracy by 27.3 percentage points on the GSM8K benchmark compared to natural language. Tam et al. further reported that the more constrained a model’s output, the more pronounced this loss of accuracy. Beyond format constraints alone, structured tool definitions increase context length (Paramanayakam et al., 2024) . As demonstrated by Levy et al. (2024) and Modarressi et al.(2025), an increase in input tokens rapidly degraded model performance well before its maximum context window. Modarressi et al. found that some models experienced 16 percentage point accur

FAQs

What problem does Natural Language Tools (NLT) address in tool calling?

NLT targets the limitations of structured, schema-constrained JSON tool calling that can cause task interference, format constraints, and increased token overhead.

How does NLT change the tool-calling workflow compared with structured tool calling?

NLT decouples tool selection into a dedicated step where models list tools and output simple YES/NO decisions; a parser then executes selected tools and passes results to a response module.

What experimental results does the paper report for NLT?

Across 10 models and 6,400 trials, NLT increases tool-calling accuracy from 69.1% to 87.5% (an 18.4-point gain) and reduces output variance by 70%.

Sep 21, 2026

Uploaded by Jiven