Research map / Large language models / Language, safety and interpretability

Exploring Iterative Enhancement for Improving Learnersourced Multiple-Choice Question Explanations with Large Language Models

AAAI 2025 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ILearner-LLM, which iteratively refines LLM-generated explanations for learnersourced MCQs using a fine-tuned generator and a fine-tuned evaluator that predicts student quality ratings and feeds scores back into generation.

Problem

Students often struggle to write high-quality explanations for the multiple-choice questions they author in learnersourcing platforms like PeerWise, and explanations are often omitted; automatic generation and evaluation of student-like explanations is needed, with limited high-quality training data

Approach

An explanation generator and an explanation evaluator (both instruction fine-tuned on PeerWise data) interact over K iterations: the predicted quality rating and the previous explanation are injected into the next generation prompt so the output is progressively improved toward student-like explanations.

Key results

• Fine-tuned LLaMA2-13B Merged with ILearner-LLM improved average BLEU on Auckland Law from 27.82 to 34.33 and Sydney Biology from 34.34 to 38.07.
• GPT-4 ILearner-LLM All History gave small gains over GPT-4 alone (e.g., Cardiff BLEU 29.08 to 30.43, UK Year 2 BLEU 30.67 to 31.83) with roughly 1.3-1.7 iterations needed to surpass the first output.
• Fine-tuned LLaMA2-13B Merged evaluator achieved MSE of 0.09-0.22 versus 3.02-3.95 for GPT-4 on most subjects (Auckland Law: 0.12 vs 0.42).

Measured effects reported in the paper

Surprising findings

• Non-fine-tuned LLaMA2-13B and GPT-4 as evaluators tended to inflate scores, with MSE values around 3-4, which the authors attribute possibly to RLHF bias.
• Instruction fine-tuned LLaMA2-13B benefited more from iterative enhancement than GPT-4 did, the authors attribute this to alignment with the instruction format used during iteration.

Limitations

• GPT-4 evaluated on only 100 random test samples due to API cost, and the evaluator and generator reported on small test sets.
• Iterative gains for GPT-4 are small (e.g., BERT changes within about 0.05) and no significant difference between history variants was found.
• Evaluation relies on similarity to student-written explanations (BLEU/BERT) rather than human judgments of explanation quality.

Experimental setup

Datasets: PeerWise Sydney Biology PeerWise Cardiff Biology PeerWise Auckland Law PeerWise UK Medical Year 1 PeerWise UK Medical Year 2

Benchmarks: PeerWise five-subject MCQ explanation test sets

Metrics: BLEU BERT score MSE average quality rating score

Baselines: Vicuna-13B GPT-3.5 GPT-4 fine-tuned LLaMA2-13B LLaMA2-13B without fine-tuning

Builds on: instruction fine-tuning chain-of-thought prompting PeerWise

Closest papers by meaning

• Rethinking Generative Large Language Model Evaluation for Semantic Comprehension (ICML 2024 · similarity 0.87)
• Improving Factuality and Reasoning in Language Models through Multiagent Debate (ICML 2024 · similarity 0.87)
• DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.86)
• Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.86)
• Potemkin Understanding in Large Language Models (ICML 2025 · similarity 0.86)
• Delving into the Reversal Curse: How Far Can Large Language Models Generalize? (NeurIPS 2024 · similarity 0.86)