💬 NLP

Don't Adapt Small Language Models for Tools; Adapt Tool Schemas to the Models

This paper introduces PA-Tool, a training-free method that improves tool-use accuracy in small language models by adapting tool schemas to match the models' pretraining knowledge, thereby reducing schema misalignment errors by up to 80% without requiring model retraining.

Original authors: Jonggeun Lee, Woojung Song, Jongwook Han, Haesung Pyun, Yohan Jo

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jonggeun Lee, Woojung Song, Jongwook Han, Haesung Pyun, Yohan Jo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Stop Teaching the Dog New Tricks; Change the Treats

Imagine you have a very smart, well-read dog (a Small Language Model or SLM). This dog has spent its entire life reading millions of books, learning how humans speak, and memorizing thousands of words. It knows the word "bark" and "woof" perfectly.

Now, you want this dog to perform a specific job: fetching a specific ball.

The Problem:
The job description (the Tool Schema) you give the dog says: "Go fetch the Spherical-Red-Object-42."

The dog, however, has never heard of "Spherical-Red-Object-42." It knows "Red Ball" or "Toy." Because the name in the instruction doesn't match the names it learned while reading its books, the dog gets confused. It might guess, "Maybe they mean 'Spherical-Red-Object-42' is actually called 'Fire-Hydrant'?" and run off to fetch a fire hydrant. This is called a Hallucination.

Most researchers try to fix this by re-training the dog (Supervised Fine-Tuning) to memorize that "Spherical-Red-Object-42" is the new name for the ball. This takes a lot of time, energy, and money.

The Paper's Solution (PA-Tool):
Instead of forcing the dog to learn a weird new name, the authors say: "Let's just change the name on the instruction card to match what the dog already knows."

They rename "Spherical-Red-Object-42" to "Red Ball." Now, the dog instantly understands, runs over, and fetches the correct item. No re-training needed.

How It Works: The "Familiarity Detector"

The paper introduces a method called PA-Tool (Pretraining-Aligned Tool Schema Generation). Here is how it works, step-by-step, using a metaphor:

1. The "Crowd Test" (Candidate Generation)

Imagine you ask the dog, "What do you call a tool that checks the weather?"
You ask it 32 times, slightly changing your tone each time (like asking different friends).

The dog might say: "WeatherCheck," "GetWeather," "SkyScan," "WeatherInfo," "CheckWeather."

2. The "Clump Detector" (Peakedness)

This is the clever part. The researchers look at the list of answers.

If the dog says "WeatherCheck" 5 times, and "SkyScan" 1 time, that's just a frequency count.
But PA-Tool looks for clumps. It asks: "How many of these answers are very similar to each other?"
- "WeatherCheck," "CheckWeather," and "Weather-Check" are all very close to each other. They form a clump.
- "SkyScan" is far away from the others.

The paper calls this "Peakedness." If the dog keeps generating names that are very similar to each other, it means the dog is familiar with that concept from its training. It's a signal that says, "Hey, I've seen this pattern a million times in my books!"

3. The "Name Swap" (Schema Selection)

PA-Tool picks the name from the biggest, tightest clump.

Original Tool Name: get_weather_info (The dog doesn't know this).
PA-Tool's New Name: get_weather_forecast (The dog recognizes this pattern immediately).

The system then swaps the name in the instruction card. The dog sees the familiar name, feels confident, and executes the task perfectly.

Why This Matters

1. It's Free and Fast
Usually, to make a small AI model better at using tools, you have to "fine-tune" it. This is like sending the dog to a expensive, months-long training camp. PA-Tool is like just rewriting the instruction manual. It takes a few seconds, costs nothing, and requires no re-training.

2. It Fixes the "Fake Name" Problem
Small models often fail because they invent fake tool names that sound real but don't exist. By aligning the names with what the model already knows, PA-Tool stops the model from guessing. In tests, this reduced "fake name" errors by 80%.

3. It Works for Everyone
The paper tested this on many different small models (like Llama and Qwen). It worked for almost all of them, boosting their success rates by up to 17%. Even better, it works alongside other methods. You can re-train a model and use PA-Tool, and the results get even better.

The Human Twist

You might worry: "If we change the names to match the AI, will humans understand them?"

The researchers asked human developers to check the new names. Surprisingly, humans liked them too!

Old Name: Figlet (A brand-specific, confusing name).
New Name: ascii_converter (Clear, descriptive, and what the AI naturally wanted to call it).

It turns out that the names the AI "remembers" from its training data are often the most logical and clear names for humans as well.

Summary

Think of PA-Tool as a translator that doesn't teach the AI a new language. Instead, it translates the tool's instructions into the language the AI already speaks fluently.

Old Way: Force the AI to memorize a dictionary of new, weird tool names.
PA-Tool Way: Change the tool names to match the dictionary the AI already has in its head.

This allows small, efficient AI models to work like big, powerful ones without needing expensive re-training. It's a simple, clever hack that bridges the gap between what the AI knows and what the tools are called.

1. Problem Statement

Small Language Models (SLMs, typically $\le$ 8B parameters) are increasingly used in multi-agent systems and edge deployments due to their computational efficiency. However, they struggle significantly with tool-augmented tasks, specifically in tool selection (identifying the correct API) and parameter identification.

The primary failure mode identified is Schema Misalignment:

The Phenomenon: When presented with a tool schema (a list of tools and parameters with descriptions), SLMs often hallucinate tool names that are plausible but do not exist in the provided schema.
The Cause: During pretraining, models internalize specific naming conventions (e.g., get_weather_forecast). When the provided schema uses a different convention (e.g., get_weather_info), the model ignores the provided schema and reverts to its internalized, "familiar" naming patterns.
Current Limitations: Existing solutions focus on training models to adapt to arbitrary schemas (via Supervised Fine-Tuning or Reinforcement Learning) or improving tool descriptions. These approaches are computationally expensive, require retraining, or fail to address the root cause: the mismatch between the model's internal knowledge and the external interface.

2. Methodology: PA-Tool

The authors propose PA-Tool (Pretraining-Aligned Tool Schema Generation), a training-free method that adapts the tool schema to align with the model's pretraining knowledge rather than forcing the model to learn new schemas.

The core insight is borrowed from data contamination detection: patterns frequently seen during training exhibit "peakedness" (a concentrated distribution of similar outputs). PA-Tool leverages this signal to identify naming conventions the model is already familiar with.

The Three-Stage Process:

Candidate Generation:
- Given a tool or parameter description, the model is prompted to generate $N$ candidate names (e.g., $N=32$ ) using a sampling temperature $t \in (0, 1]$ .
- This explores the model's learned distribution of naming patterns.
Peakedness Computation:
- For each candidate name $s_i$ , the algorithm calculates a peakedness score $\phi(s_i)$ .
- This score counts how many other generated candidates are sufficiently similar to $s_i$ based on a character-level edit distance (Levenshtein distance).
- A threshold $\tau$ is defined dynamically based on the maximum character length of the candidates: $\tau = \alpha \cdot \ell_{max}$ .
- High peakedness indicates that the model consistently generates variations of this specific name, suggesting it is a deeply internalized pattern from pretraining.
Schema Selection:
- The candidate with the highest peakedness score is selected as the new, pretraining-aligned name.
- If there is a tie, the candidate closest to a greedy-decoding reference name is chosen.
- This process is applied to all tools and parameters in the schema, creating a mapping dictionary that replaces original names with model-aligned names.

3. Key Contributions

Paradigm Shift: The paper challenges the standard approach of adapting models to tools. Instead, it proposes adapting the tool interface (schema) to the model's internalized knowledge.
Training-Free Optimization: PA-Tool requires no model retraining, fine-tuning, or additional computational resources during inference. It only requires a one-time schema mapping generation.
Novel Signal Utilization: It repurposes "peakedness," a metric typically used for detecting data contamination/memorization, as a constructive signal for schema alignment.
Comprehensive Evaluation: The method is validated across multiple benchmarks (MetaTool, RoTBench, API-Bank, $\tau$ -Bench) and diverse model families (Qwen, Llama, Ministral, etc.).

4. Experimental Results

Experiments were conducted on SLMs (3B–8B) and compared against Base models, Greedy decoding, Most-Frequent sampling, Human-designed schemas, and state-of-the-art closed-source models.

Performance Gains:
- MetaTool: PA-Tool achieved up to 17% improvement in accuracy, particularly in the Reliability subtask (identifying when no tool exists), where errors were reduced by 80%.
- RoTBench: Improvements of up to 10% in single-turn tool selection and 6% in multi-turn settings.
- Parameter Identification: Gains extended beyond tool selection to parameter extraction, confirming that schema alignment improves overall tool-use accuracy.
Error Reduction:
- Schema Misalignment Errors: Reduced by 80.0%.
- Functional Confusion: Reduced by 24.0%.
- Context Understanding: Reduced by 18.8%.
Comparison with SFT: PA-Tool alone outperformed Supervised Fine-Tuning (SFT) on several subtasks without any training. Combining PA-Tool with SFT yielded the best results, indicating they address complementary failure modes.
Generalization: The method improved performance across diverse model sizes (from 3B to 70B) and reasoning models, though gains were most pronounced for SLMs where schema misalignment is a dominant failure mode.
Human Evaluation: Human annotators rated PA-Tool generated names as more understandable and functionally descriptive than original names in over 50% of cases, proving the method does not sacrifice human interpretability.

5. Significance and Implications

Resource Efficiency: PA-Tool unlocks the potential of resource-efficient SLMs for complex agentic workflows without the cost of retraining or the risk of catastrophic forgetting.
Scalability: The one-time schema mapping cost is negligible (seconds to minutes per tool set), making it highly scalable for dynamic tool environments.
Robustness: The method is robust to noise in tool names (e.g., typos, reversed strings) because it regenerates names based on descriptions, providing a stable performance floor.
Practical Deployment: It offers a simple, deployable intervention for developers building multi-agent systems, allowing them to bridge the gap between pretraining knowledge and real-world tool interfaces immediately.

In conclusion, the paper demonstrates that schema-level intervention is a highly effective, low-cost strategy to enhance tool-use capabilities in Small Language Models, fundamentally shifting the optimization burden from the model to the interface.