PeFG-MARA: Personalized feedback generation via multi-agent retrieval-augmented large language models

Generating personalized feedback requires identifying response-specific issues while grounding advice in reliable instructional evidence. We present PeFG-MARA, a modular workflow with three role-specific 7B LLM agents (Planner, Generator, and Critic) and a non-LLM Retrieval Module. The Planner turns a student’s response into a structured feedback plan and optional retrieval query; the retriever supplies instructional evidence; the Generator drafts a comment; and the Critic approves or requests revision. The agents share a base model and use separately trained LoRA adapters. We evaluate the system on four educational-feedback datasets covering science short answers, English writing, and mathematical problem solving, against neural, single-agent RAG, and prompted LLM baselines. On SAF, PeFG-MARA reaches BLEU/ROUGE-L/BERTScore scores of 32.8/37.4/0.921 versus 32.2/36.7/0.918 for GPT-4o Prompted. On the 20-essay EssayF test set, it yields the highest descriptive point estimates for BERTScore, coverage, specificity, and factual consistency; the larger LEAF test provides complementary writing-feedback evidence. Same-data controls and ablations indicate contributions from retrieval and coordinated planning and critique, while additional supervision improves performance further. The results suggest that an explicit feedback plan can help align evidence with a student’s particular errors. The inspectable plan could support instructor review of draft feedback, subject to retrieval quality, latency, and further classroom evaluation.