@inproceedings{baghel-etal-2026-gear,
title = "{GEAR}: Training-Free Rule Distillation for Advanced and Efficient Tool-Augmented Reasoning",
author = "Baghel, Bhiman Kumar and
Wu, Fanyou and
Maharjan, Suraj and
Yao, Peiran and
Sengamedu, Srinivasan H.",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.214/",
pages = "4781--4820",
ISBN = "979-8-89176-439-2",
abstract = "Improving tool use in large reasoning models (LRMs) typically demands expensive test-time compute, training resources unavailable to API-only practitioners, or expensive iterative rollouts with execution feedback required by training-free context-engineering methods, all incompatible with resource-constrained deployment. We introduce **GEAR**, a training-free framework that extracts compositional rules from tool-use failures via single-pass meta-reasoning and distills them into inference prompts. We hypothesize that combining explicit rules with examples outperforms either alone. Across 78 API domains, **GEAR**+2-shot outperforms 3-shot by 24.9{\%} on Qwen3-30B, validating our complementarity hypothesis. On Qwen3-4B, **GEAR** exceeds LoRA SFT by 24.3{\%} and RAG-3-shot by 38.5{\%}, while **GEAR**+2-shot beats the state-of-the-art context-engineering method ACE by 38.8{\%} at 4.1$\times$ lower cost. Remarkably, Qwen3-30B with **GEAR**+2-shot nearly doubles zero-shot accuracy while remaining under 1 s latency, demonstrating rollout-free adaptation for resource-constrained deployment."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="baghel-etal-2026-gear">
<titleInfo>
<title>GEAR: Training-Free Rule Distillation for Advanced and Efficient Tool-Augmented Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Bhiman</namePart>
<namePart type="given">Kumar</namePart>
<namePart type="family">Baghel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fanyou</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Suraj</namePart>
<namePart type="family">Maharjan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peiran</namePart>
<namePart type="family">Yao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Srinivasan</namePart>
<namePart type="given">H</namePart>
<namePart type="family">Sengamedu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Improving tool use in large reasoning models (LRMs) typically demands expensive test-time compute, training resources unavailable to API-only practitioners, or expensive iterative rollouts with execution feedback required by training-free context-engineering methods, all incompatible with resource-constrained deployment. We introduce **GEAR**, a training-free framework that extracts compositional rules from tool-use failures via single-pass meta-reasoning and distills them into inference prompts. We hypothesize that combining explicit rules with examples outperforms either alone. Across 78 API domains, **GEAR**+2-shot outperforms 3-shot by 24.9% on Qwen3-30B, validating our complementarity hypothesis. On Qwen3-4B, **GEAR** exceeds LoRA SFT by 24.3% and RAG-3-shot by 38.5%, while **GEAR**+2-shot beats the state-of-the-art context-engineering method ACE by 38.8% at 4.1\times lower cost. Remarkably, Qwen3-30B with **GEAR**+2-shot nearly doubles zero-shot accuracy while remaining under 1 s latency, demonstrating rollout-free adaptation for resource-constrained deployment.</abstract>
<identifier type="citekey">baghel-etal-2026-gear</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.214/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>4781</start>
<end>4820</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T GEAR: Training-Free Rule Distillation for Advanced and Efficient Tool-Augmented Reasoning
%A Baghel, Bhiman Kumar
%A Wu, Fanyou
%A Maharjan, Suraj
%A Yao, Peiran
%A Sengamedu, Srinivasan H.
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F baghel-etal-2026-gear
%X Improving tool use in large reasoning models (LRMs) typically demands expensive test-time compute, training resources unavailable to API-only practitioners, or expensive iterative rollouts with execution feedback required by training-free context-engineering methods, all incompatible with resource-constrained deployment. We introduce **GEAR**, a training-free framework that extracts compositional rules from tool-use failures via single-pass meta-reasoning and distills them into inference prompts. We hypothesize that combining explicit rules with examples outperforms either alone. Across 78 API domains, **GEAR**+2-shot outperforms 3-shot by 24.9% on Qwen3-30B, validating our complementarity hypothesis. On Qwen3-4B, **GEAR** exceeds LoRA SFT by 24.3% and RAG-3-shot by 38.5%, while **GEAR**+2-shot beats the state-of-the-art context-engineering method ACE by 38.8% at 4.1\times lower cost. Remarkably, Qwen3-30B with **GEAR**+2-shot nearly doubles zero-shot accuracy while remaining under 1 s latency, demonstrating rollout-free adaptation for resource-constrained deployment.
%U https://aclanthology.org/2026.aacl-long.214/
%P 4781-4820
Markdown (Informal)
[GEAR: Training-Free Rule Distillation for Advanced and Efficient Tool-Augmented Reasoning](https://aclanthology.org/2026.aacl-long.214/) (Baghel et al., AACL-IJCNLP 2026)
ACL
- Bhiman Kumar Baghel, Fanyou Wu, Suraj Maharjan, Peiran Yao, and Srinivasan H. Sengamedu. 2026. GEAR: Training-Free Rule Distillation for Advanced and Efficient Tool-Augmented Reasoning. In Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4781–4820, Hengqin, China. Association for Computational Linguistics.