@inproceedings{khalil-mourad-2026-instruction,
title = "The Instruction Tax on Factual Recall: Base Models Can Beat Their Instruction-Tuned Counterparts",
author = "Khalil, Christophe and
Mourad, Aya",
editor = "Kalo, Jan-Christoph and
Nakashole, Ndapa and
Razniewski, Simon and
Suchanek, Fabian M. and
Vlachos, Andreas and
McCallum, Andrew",
booktitle = "Proceedings of the 11th Workshop on Automated Knowledge Base Construction ({AKBC} 2026)",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.akbc-1.40/",
pages = "518--526",
ISBN = "979-8-89176-464-4",
abstract = "We present our solution to the AKBC LM-KBC 2026 Shared Task, where we achieved a score of 0.6836 on the official leaderboard, more than double the baseline of 0.2964. This task requires generating all possible objects for subject-relation pairs using a language model with up to 32 billion parameters, with no retrieval or fine-tuning involved. We show that the prompting methodology is as important as the model architecture: while the instruction-tuned model version achieves 0.5681 with its chat interface, the base model achieves 0.6836 when prompted using text completion. Thus, instruction tuning decreases factual recall; we refer to this effect as the instruction tax, which is caused by the model{'}s weights rather than the prompting method. Our final solution is fairly straightforward: we prompt the base model with some examples from the training data, generate 64-256 outputs, and retain only the repeated ones. Two specialized methods are then applied for each relation: one improves the recall of large answer sets, and the other verifies a person{'}s vitality before predicting their place of death."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="khalil-mourad-2026-instruction">
<titleInfo>
<title>The Instruction Tax on Factual Recall: Base Models Can Beat Their Instruction-Tuned Counterparts</title>
</titleInfo>
<name type="personal">
<namePart type="given">Christophe</namePart>
<namePart type="family">Khalil</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aya</namePart>
<namePart type="family">Mourad</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 11th Workshop on Automated Knowledge Base Construction (AKBC 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jan-Christoph</namePart>
<namePart type="family">Kalo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ndapa</namePart>
<namePart type="family">Nakashole</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Simon</namePart>
<namePart type="family">Razniewski</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fabian</namePart>
<namePart type="given">M</namePart>
<namePart type="family">Suchanek</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andreas</namePart>
<namePart type="family">Vlachos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andrew</namePart>
<namePart type="family">McCallum</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-464-4</identifier>
</relatedItem>
<abstract>We present our solution to the AKBC LM-KBC 2026 Shared Task, where we achieved a score of 0.6836 on the official leaderboard, more than double the baseline of 0.2964. This task requires generating all possible objects for subject-relation pairs using a language model with up to 32 billion parameters, with no retrieval or fine-tuning involved. We show that the prompting methodology is as important as the model architecture: while the instruction-tuned model version achieves 0.5681 with its chat interface, the base model achieves 0.6836 when prompted using text completion. Thus, instruction tuning decreases factual recall; we refer to this effect as the instruction tax, which is caused by the model’s weights rather than the prompting method. Our final solution is fairly straightforward: we prompt the base model with some examples from the training data, generate 64-256 outputs, and retain only the repeated ones. Two specialized methods are then applied for each relation: one improves the recall of large answer sets, and the other verifies a person’s vitality before predicting their place of death.</abstract>
<identifier type="citekey">khalil-mourad-2026-instruction</identifier>
<location>
<url>https://aclanthology.org/2026.akbc-1.40/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>518</start>
<end>526</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The Instruction Tax on Factual Recall: Base Models Can Beat Their Instruction-Tuned Counterparts
%A Khalil, Christophe
%A Mourad, Aya
%Y Kalo, Jan-Christoph
%Y Nakashole, Ndapa
%Y Razniewski, Simon
%Y Suchanek, Fabian M.
%Y Vlachos, Andreas
%Y McCallum, Andrew
%S Proceedings of the 11th Workshop on Automated Knowledge Base Construction (AKBC 2026)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-464-4
%F khalil-mourad-2026-instruction
%X We present our solution to the AKBC LM-KBC 2026 Shared Task, where we achieved a score of 0.6836 on the official leaderboard, more than double the baseline of 0.2964. This task requires generating all possible objects for subject-relation pairs using a language model with up to 32 billion parameters, with no retrieval or fine-tuning involved. We show that the prompting methodology is as important as the model architecture: while the instruction-tuned model version achieves 0.5681 with its chat interface, the base model achieves 0.6836 when prompted using text completion. Thus, instruction tuning decreases factual recall; we refer to this effect as the instruction tax, which is caused by the model’s weights rather than the prompting method. Our final solution is fairly straightforward: we prompt the base model with some examples from the training data, generate 64-256 outputs, and retain only the repeated ones. Two specialized methods are then applied for each relation: one improves the recall of large answer sets, and the other verifies a person’s vitality before predicting their place of death.
%U https://aclanthology.org/2026.akbc-1.40/
%P 518-526
Markdown (Informal)
[The Instruction Tax on Factual Recall: Base Models Can Beat Their Instruction-Tuned Counterparts](https://aclanthology.org/2026.akbc-1.40/) (Khalil & Mourad, AKBC 2026)
ACL