@inproceedings{kucharavy-etal-2025-llms,
title = "{LLM}s Prot{\'e}g{\'e}s: Tutoring {LLM}s with Knowledge Gaps Improves Student Learning Outcome",
author = "Kucharavy, Andrei and
Vallez, Cyril and
Percia David, Dimitri",
editor = {Kochmar, Ekaterina and
Alhafni, Bashar and
Bexte, Marie and
Burstein, Jill and
Horbach, Andrea and
Laarmann-Quante, Ronja and
Tack, Ana{\"i}s and
Yaneva, Victoria and
Yuan, Zheng},
booktitle = "Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025)",
month = jul,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.bea-1.19/",
doi = "10.18653/v1/2025.bea-1.19",
pages = "248--257",
ISBN = "979-8-89176-270-1",
abstract = "Since the release of ChatGPT, Large Langauge Models (LLMs) have been proposed as potential tutors to students in the education outcomes. Such an LLM-as-tutors metaphor is problematic, notably due to the counterfactual generation, perception of learned skills as mastered by an automated system and hence non-valuable, and learning LLM over-reliance.We propose instead the LLM-as-mentee tutoring schema, leveraging the Learning-by-Teaching prot{\'e}g{\'e} effect in peer tutoring - LLM Prot{\'e}g{\'e}s. In this configuration, counterfactual generation is desirable, allowing students to operationalize the learning material and better understand the limitations of LLM-based systems, both a skill in itself and an additional learning motivation. Our preliminary results suggest that LLM Prot{\'e}g{\'e}s are effective. Students in an introductory algorithms class who successfully diagnosed an LLM teachable agent system prompted to err on a course material gained an average of 0.72 points on a 1-6 scale. Remarkably, if fully adopted, this approach would reduce the failure rate in the second midterm from 28{\%} to 8{\%}, mitigating 72{\%} of midterm failure.We publish code for on-premises deployment of LLM Prot{\'e}g{\'e}s on https://github.com/Reliable-Information-Lab-HEVS/LLM{\_}Proteges."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kucharavy-etal-2025-llms">
<titleInfo>
<title>LLMs Protégés: Tutoring LLMs with Knowledge Gaps Improves Student Learning Outcome</title>
</titleInfo>
<name type="personal">
<namePart type="given">Andrei</namePart>
<namePart type="family">Kucharavy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cyril</namePart>
<namePart type="family">Vallez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dimitri</namePart>
<namePart type="family">Percia David</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ekaterina</namePart>
<namePart type="family">Kochmar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bashar</namePart>
<namePart type="family">Alhafni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marie</namePart>
<namePart type="family">Bexte</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jill</namePart>
<namePart type="family">Burstein</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andrea</namePart>
<namePart type="family">Horbach</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ronja</namePart>
<namePart type="family">Laarmann-Quante</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anaïs</namePart>
<namePart type="family">Tack</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Victoria</namePart>
<namePart type="family">Yaneva</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Vienna, Austria</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-270-1</identifier>
</relatedItem>
<abstract>Since the release of ChatGPT, Large Langauge Models (LLMs) have been proposed as potential tutors to students in the education outcomes. Such an LLM-as-tutors metaphor is problematic, notably due to the counterfactual generation, perception of learned skills as mastered by an automated system and hence non-valuable, and learning LLM over-reliance.We propose instead the LLM-as-mentee tutoring schema, leveraging the Learning-by-Teaching protégé effect in peer tutoring - LLM Protégés. In this configuration, counterfactual generation is desirable, allowing students to operationalize the learning material and better understand the limitations of LLM-based systems, both a skill in itself and an additional learning motivation. Our preliminary results suggest that LLM Protégés are effective. Students in an introductory algorithms class who successfully diagnosed an LLM teachable agent system prompted to err on a course material gained an average of 0.72 points on a 1-6 scale. Remarkably, if fully adopted, this approach would reduce the failure rate in the second midterm from 28% to 8%, mitigating 72% of midterm failure.We publish code for on-premises deployment of LLM Protégés on https://github.com/Reliable-Information-Lab-HEVS/LLM_Proteges.</abstract>
<identifier type="citekey">kucharavy-etal-2025-llms</identifier>
<identifier type="doi">10.18653/v1/2025.bea-1.19</identifier>
<location>
<url>https://aclanthology.org/2025.bea-1.19/</url>
</location>
<part>
<date>2025-07</date>
<extent unit="page">
<start>248</start>
<end>257</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LLMs Protégés: Tutoring LLMs with Knowledge Gaps Improves Student Learning Outcome
%A Kucharavy, Andrei
%A Vallez, Cyril
%A Percia David, Dimitri
%Y Kochmar, Ekaterina
%Y Alhafni, Bashar
%Y Bexte, Marie
%Y Burstein, Jill
%Y Horbach, Andrea
%Y Laarmann-Quante, Ronja
%Y Tack, Anaïs
%Y Yaneva, Victoria
%Y Yuan, Zheng
%S Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025)
%D 2025
%8 July
%I Association for Computational Linguistics
%C Vienna, Austria
%@ 979-8-89176-270-1
%F kucharavy-etal-2025-llms
%X Since the release of ChatGPT, Large Langauge Models (LLMs) have been proposed as potential tutors to students in the education outcomes. Such an LLM-as-tutors metaphor is problematic, notably due to the counterfactual generation, perception of learned skills as mastered by an automated system and hence non-valuable, and learning LLM over-reliance.We propose instead the LLM-as-mentee tutoring schema, leveraging the Learning-by-Teaching protégé effect in peer tutoring - LLM Protégés. In this configuration, counterfactual generation is desirable, allowing students to operationalize the learning material and better understand the limitations of LLM-based systems, both a skill in itself and an additional learning motivation. Our preliminary results suggest that LLM Protégés are effective. Students in an introductory algorithms class who successfully diagnosed an LLM teachable agent system prompted to err on a course material gained an average of 0.72 points on a 1-6 scale. Remarkably, if fully adopted, this approach would reduce the failure rate in the second midterm from 28% to 8%, mitigating 72% of midterm failure.We publish code for on-premises deployment of LLM Protégés on https://github.com/Reliable-Information-Lab-HEVS/LLM_Proteges.
%R 10.18653/v1/2025.bea-1.19
%U https://aclanthology.org/2025.bea-1.19/
%U https://doi.org/10.18653/v1/2025.bea-1.19
%P 248-257
Markdown (Informal)
[LLMs Protégés: Tutoring LLMs with Knowledge Gaps Improves Student Learning Outcome](https://aclanthology.org/2025.bea-1.19/) (Kucharavy et al., BEA 2025)
ACL