Tackling Students’ Coding Assignments with LLMs

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F24%3A10485470" target="_blank" >RIV/00216208:11320/24:10485470 - isvavai.cz</a>

Výsledek na webu

<a href="https://doi.org/10.1145/3643795.3648389" target="_blank" >https://doi.org/10.1145/3643795.3648389</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1145/3643795.3648389" target="_blank" >10.1145/3643795.3648389</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Tackling Students’ Coding Assignments with LLMs

Popis výsledku v původním jazyce

State-of-the-art large language models (LLMs) have demonstrated an extraordinary ability to write computer code. This ability can be quite beneficial when integrated into an IDE to assist a programmer with basic coding. On the other hand, it may be misused by computer science students for cheating on coding tests or homework assignments. At present, knowledge about the exact capabilities and limitations of state-of-the-art LLMs is still inadequate. Furthermore, their capabilities have been changing quickly with each new release. In this paper, we present a dataset of 559 programming exercises in 10 programming languages collected from a system for evaluating coding assignments at our university. We have experimented with four well-known LLMs (GPT-3.5, GPT-4, Codey, Code Llama) and asked them to solve these assignments. The evaluation results are intriguing and provide insights into the strengths and weaknesses of the models. In particular, GPT-4 (which performed the best) is currently capable of solving 55% of all our exercises and achieved an average score of 86% on exercises from the introductory programming course (using the best of five generated solutions).

Název v anglickém jazyce

Tackling Students’ Coding Assignments with LLMs

Popis výsledku anglicky

State-of-the-art large language models (LLMs) have demonstrated an extraordinary ability to write computer code. This ability can be quite beneficial when integrated into an IDE to assist a programmer with basic coding. On the other hand, it may be misused by computer science students for cheating on coding tests or homework assignments. At present, knowledge about the exact capabilities and limitations of state-of-the-art LLMs is still inadequate. Furthermore, their capabilities have been changing quickly with each new release. In this paper, we present a dataset of 559 programming exercises in 10 programming languages collected from a system for evaluating coding assignments at our university. We have experimented with four well-known LLMs (GPT-3.5, GPT-4, Codey, Code Llama) and asked them to solve these assignments. The evaluation results are intriguing and provide insights into the strengths and weaknesses of the models. In particular, GPT-4 (which performed the best) is currently capable of solving 55% of all our exercises and achieved an average score of 86% on exercises from the introductory programming course (using the best of five generated solutions).

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

S - Specificky vyzkum na vysokych skolach<br>I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Rok uplatnění

2024

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Proceedings of the 1st International Workshop on Large Language Models for Code

ISBN

979-8-4007-0579-3

ISSN

—

e-ISSN

—

Počet stran výsledku

8

Strana od-do

—

Název nakladatele

Neuveden

Místo vydání

Neuveden

Místo konání akce

Lisbon, Portugal

Datum konání akce

14. 4. 2024

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—