AICoFe keeps teachers in the feedback loop
AICoFe is a human-centered system for teacher, peer, and self-assessment in higher education. It is organized into four modules: a Visualization Module, a Management Module, a Feedback Generation Module, and a Recording Module. The Management Module connects the platform to SQL and MongoDB databases and a file system, while the Recording Module captures student performances in video and audio using a standard laptop with Microsoft Teams.
The design premise is that peer feedback is often superficial or inconsistent because students lack the feedback literacy to produce constructive comments. Rather than replacing peer judgment, AICoFe adds generative AI on top of rubric-based evaluation and keeps teachers as active mediators who review and curate the final feedback.
Three dashboards split evaluation, reflection, and curation
The Evaluator Dashboard supports teachers and students during evaluation. Evaluators select a rubric and a target student, assign Likert scores per item, and add written comments, with rubric level descriptions always visible to align interpretations. The system also captures interaction metadata such as score timing for later analysis.
The Student Dashboard promotes reflective learning. Students can revisit a video recording of their performance and complete a self-evaluation using the same rubric applied by peers and instructors. The dashboard then visualizes differences and alignments between self-assessment scores and the aggregated evaluations received, alongside AI-supported collaborative feedback that students can rate for coherence and usefulness directly in the dashboard.
The Teacher Dashboard supports review and curation. Teachers inspect quantitative scores and qualitative comments from peers and instructors, access performance videos, and review feedback generated by different LLMs. They can select, edit, or combine sentences from AI outputs to compose the final feedback, and a log tracks drafts and sent versions while distinguishing LLM sentences from teacher-modified ones. Text-to-speech via Google Text-to-Speech lets teachers listen to detailed comments.
Three fine-tuned LLMs draft a fixed three-paragraph structure
The Feedback Generation Module combines quantitative assessment results from teachers and peers with qualitative observations. It extends earlier work to a multi-model setting in which three independent feedback instances are generated, one per LLM. Each instance follows a predefined three-paragraph structure: strengths, areas for improvement, and an action plan with concrete recommendations.
Before generation, textual comments are normalized to remove invalid characters and automatically screened for relevance to the evaluated rubric item. Filtered data are combined into course-specific prompts that integrate average rubric scores, validated qualitative observations, and rubric level descriptions. All data sent to the generative models are anonymized.
The module supports OpenAI GPT-4.1-mini, Google Gemini 2.5 Flash, and Meta Llama 3.1. All three were fine-tuned with supervised learning on the SOPHIAS dataset using input-output pairs structured to the three-paragraph schema. GPT-4.1-mini was fine-tuned via the OpenAI API for 3 epochs (lr multiplier 2, final loss 0.934); Gemini 2.5 Flash via Google Vertex AI for 40 cycles with the optimal checkpoint at step 45 (final loss 0.759, accuracy 0.744); Llama 3.1 locally via LoRA (r=16, alpha=32) with 4-bit NF4 quantization for 3 epochs (final loss 1.210).
A post-generation validation stage checks teacher-defined constraints on length, coherence, and restricted terms, triggering automatic regeneration if needed. Prompts contain exclusively anonymized rubric scores and evaluator comments, and iterative prompt refinement is used to keep output grounded in observations and mitigate hallucination risk.
The evaluation covered 65 students across two degree levels
AICoFe was deployed during 2024-2025 with 46 undergraduate and 19 master's students at the Universidad Autonoma de Madrid. Students delivered oral presentations evaluated through peer, self, and teacher assessments using rubric-based criteria. The instructor designed the rubrics based on the specific competency criteria of each subject to align with each course's learning objectives.
The paper reports that the reliability of this multi-source design was validated through inter-rater agreement analyses and scoring bias checks, consistent with previous work. This matters for interpreting the satisfaction results: the feedback students rated was produced within a design the authors treat as reliable, not from an unvalidated scoring setup.
Coherence and usefulness were rated high in both groups
Students rated feedback coherence and feedback usefulness using numerical scales inside the Student Dashboard. Master's students (17 of 19, 89%) reported mean coherence of 9.24 (sigma 0.83) and usefulness of 9.00 (sigma 1.22). Undergraduate students (46 of 46, 100%) reported mean coherence of 8.83 (sigma 0.92) and usefulness of 9.20 (sigma 1.02).
The two groups differ in which dimension scored higher: master's students rated coherence above usefulness, while undergraduates rated usefulness above coherence. Both remain high, and the authors interpret the scores as suggesting that teacher-curated AI feedback was perceived as aligned with student performance and actionable. These are perception measures collected in the dashboard, not independent assessments of feedback accuracy or of learning outcomes.
Usability scored excellent, but on a smaller sample
Usability was measured with the System Usability Scale. A total of 31 participants completed the questionnaire, yielding a mean SUS score of 79.91 (sigma 11.92), which the paper describes as corresponding to an excellent usability level.
The denominator here is 31 participants, not the 65 students in the deployment, so the usability result rests on a smaller subset than the satisfaction ratings. The paper does not report which degree levels those 31 participants came from, so the SUS score should not be read as a per-group result.
The evidence supports satisfaction, not learning gains
The paper's own conclusion is that AICoFe supports coherent, useful, and pedagogically meaningful feedback while preserving strong teacher oversight, and that the findings highlight the importance of the teacher's role as an active mediator in AI-assisted assessment. The authors also frame AICoFe as a practical, research-informed model for scaling peer assessment without compromising feedback quality.
That framing should be read against what was actually measured. The reported outcomes are student-rated coherence and usefulness plus a SUS usability score. The paper does not report a control condition, a pre-post learning measure, or a comparison of AI-assisted feedback against feedback without AI, so the results cannot establish that AICoFe improves learning or that it outperforms standard peer feedback.
The deployment is also bounded: 65 students at one university across 2024-2025, oral presentations, and rubrics designed per subject by the instructor. The authors state that future work includes a larger deployment and the integration of multimodal data and biometrics into the feedback process, which signals that the current evidence is an initial authentic-setting evaluation rather than a scaled one.
Teacher curation is the variable worth watching
The most decision-relevant feature of AICoFe is not that three LLMs generate feedback, but that teachers select, edit, or combine sentences before the final feedback reaches students, with a log distinguishing LLM sentences from teacher-modified ones. The satisfaction ratings therefore describe teacher-curated AI feedback, not raw model output.
For readers weighing whether to adopt a similar system, that distinction sets the practical threshold. The paper's evidence supports the combination of rubric-based scoring, multi-model drafting, and teacher mediation in this deployment; it does not tell you how the same students would have rated uncurated LLM feedback, nor how the system would perform with different rubrics, subjects, or larger cohorts. The next useful test is whether the curation step, rather than the generation step, is what carries the perceived coherence and usefulness.
Sources used in this answer
AICoFe Demo: AI-based Collaborative Feedback System
In a 2024-2025 deployment with 65 undergraduate and master's students, AICoFe combined rubric-based peer, self, and teacher assessment with three fine-tuned LLMs and teacher curation; students rated feedback coherence and usefulness highly and the System Usability Scale averaged 79.91 across 31 respondents.