Who's the Best Detective? LLMs vs. MLs in Detecting Incoherent Fourth Grade Math Answers

Felipe Urrutia, Roberto Araya

Abstract

Written answers to open-ended questions can have a higher long-term effect on learning than multiple-choice questions. However, it is critical that teachers immediately review the answers and ask to redo those that are incoherent. This can be a difficult task and can be time-consuming for teachers. A possible solution is to automate the detection of incoherent answers. One option is to automate the review with Large Language Models (LLM). In this paper, we analyze the responses of fourth graders in mathematics using three LLMs: GPT-3, BLOOM, and YOU. We used them with zero, one, two, three and four shots. We compared their performance with the results of various classifiers trained with Machine Learning (ML). We found that LLMs perform worse than MLs in detecting incoherent answers. The difficulty seems to reside in recursive questions that contain both questions and answers, and in responses from students with typical fourth-grader misspellings. Upon closer examination, we have found that the ChatGPT model faces the same challenges.

Introduction

In mathematics education, open-ended questions are superior to multiple-choice for long-term learning, but they create a massive grading bottleneck for teachers. To be effective, teachers must identify "incoherent" answers—responses that aren't just wrong, but nonsensical or indicative of a student giving up—and ask for a redo in real-time. Automating this detection is difficult because fourth-grade writing is a linguistic "wild west" of phonetic spelling and non-standard grammar. This paper investigates whether off-the-shelf Large Language Models (LLMs) can handle this nuanced classification task through prompting, or if custom-trained, supervised Machine Learning (ML) models remain the superior "detectives" in the classroom.

Methodology

The researchers conducted a head-to-head comparison between two distinct AI approaches: supervised ML and LLMs.

Overview of the experimental methods and materials used in the study.
  1. Supervised Machine Learning: The study utilized XGBoost and BETO-mt (a Spanish-language version of BERT). These models were specifically trained on thousands of labeled student responses.
  2. Large Language Models (LLMs): The team tested GPT-3 (text-davinci-003), BLOOM, and YouChat.

Prompt engineering strategies for the LLMs included zero-shot, few-shot (one to four examples), and Chain of Thought (CoT) variants designed to handle "recursivity" and ignore misspellings.

Datasets and Materials

The study utilized real-world data from ConectaIdeas, an online math platform used in Chile, involving 974 fourth graders and 12 teachers. The ML models were trained on 2019 data and tested on a separate 2017 dataset to ensure generalizability. Responses were manually labeled by 10 teachers as either "Coherent" or "Incoherent."

Distribution of coherent and incoherent answers across different question types (Q0-Q5).

As shown in the data distribution above, the researchers categorized questions into six types (Q0-Q5). Q3 (recursive questions) and Q5 represented significant portions of the dataset, providing a robust test for model reasoning.

Key Insights: The Challenges of "Kid-Speak" and Recursion

  • The "Kid-Speak" Barrier: LLMs, trained primarily on adult-centric web data, struggle with phonetic misspellings common in nine-year-olds (e.g., "desenavale" for "decena vale"). While fine-tuned models like BETO recognize these as valid thoughts, LLMs often flag them as incoherent.
  • The Recursion Trap: The biggest failure point for LLMs was "recursive" questions (Q3). When a prompt asks a student to evaluate a fictional character's statement, LLMs often confuse the character's answer with the student's response.
  • High Recall, Low Precision: GPT-3 was highly sensitive, catching most incoherent answers but also incorrectly flagging many valid student responses. This "crying wolf" behavior would be frustrating in a live classroom setting.

Results

The supervised BETO-mt model was the clear winner, achieving an F1-score of 79.15% for detecting incoherence. In contrast, even with four-shot prompting, GPT-3 only reached an F1-score of 63.59%. The precision gap was particularly stark: GPT-3’s precision for the incoherent class was 48.82%, compared to BETO-mt’s 76.19%. Open-source models like BLOOM and YouChat performed significantly worse, sometimes failing to outperform random baselines.

Limitations and Future Work

While newer models like GPT-4 may show improvement, the core limitation remains: LLMs lack specific exposure to the idiosyncratic ways children write and reason.

Future Directions: * Hybrid Systems: Using fine-tuned ML models for detection and LLMs for generating explanations for teachers. * Multimodal Knowledge Transfer: Applying techniques from Computer Vision to help LLMs navigate logical recursion. * Kid-Specific Pre-training: Developing datasets that capture the linguistic development of elementary students to better align LLMs with educational environments.

Summary updated

Abstract

Written answers to open-ended questions can have a higher long-term effect on learning than multiple-choice questions. However, it is critical that teachers immediately review the answers, and ask to redo those that are incoherent. This can be a difficult task and can be time-consuming for teachers. A possible solution is to automate the detection of incoherent answers. One option is to automate the review with Large Language Models (LLM). In this paper, we analyze the responses of fourth graders in mathematics using three LLMs: GPT-3, BLOOM, and YOU. We used them with zero, one, two, three and four shots. We compared their performance with the results of various classifiers trained with Machine Learning (ML). We found that LLMs perform worse than MLs in detecting incoherent answers. The difficulty seems to reside in recursive questions that contain both questions and answers, and in responses from students with typical fourth-grader misspellings. Upon closer examination, we have found that the ChatGPT model faces the same challenges.
BibTeX
@misc{urrutia2023who,
  author = {Felipe Urrutia and Roberto Araya},
  title = {{Who's the Best Detective? LLMs vs. MLs in Detecting Incoherent Fourth Grade Math Answers}},
  year = {2023},
  eprint = {2304.11257},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.11257},
}
Figure