MathCog benchmark (3,036 teacher-annotated diagnostic verdicts, 639 handwritten responses, 18 LLMs): all models severely underperform (macro F1 < 0.5) — over-attributing evidence, overthinking minimal cues, hallucinating nonexistent evidence (hallucination-risk) — calling for evidence-aware architectures and teacher-in-the-loop designs (knowledge-tracing, multimodal, benchmark).
Summary
This paper introduces MathCog, a benchmark dataset of 3,036 teacher-annotated diagnostic verdicts across 639 student handwritten math responses to 110 problems. Evaluating 18 LLMs, the authors find that all models severely underperform (macro F1 < 0.5), with systematic failure modes: over-attributing evidential strength, overthinking minimal cues, and hallucinating nonexistent evidence. Performance degrades sharply when student evidence is vague or implicit. The study calls for evidence-aware architectures and teacher-in-the-loop designs.
Key Contributions
1. MathCog Benchmark — First benchmark for cognitive skill diagnosis from handwritten math, grounded in TIMSS 2019 cognitive framework with Evident/Vague evidential strength labels 2. Systematic LLM Evaluation — 18 models spanning reasoning, multimodal, and text-only architectures, all showing F1 < 0.5 3. Error Taxonomy — Five systematic error patterns identified: evidence misidentification, rubric misinterpretation, over-inference, inconsistency, and hallucination 4. Evidential Calibration Metrics — Introduces OverAttr and FalseAttr to quantify models' tendency to over-claim evidential confidence
Core Findings
Universal Underperformance
No model achieves F1 ≥ 0.5. The best performers (GPT-4o-img at 0.448, DeepSeek-R1 at 0.442) still fail on nearly half of diagnostic decisions. Accuracy (mean 0.680) is misleading due to class imbalance — most student responses provide Evident Yes evidence, inflating accuracy.Evidence Sensitivity Gap
All models perform worse when student evidence is Vague (implicit, incomplete, or context-dependent). Multimodal models show a larger Evident-to-Vague performance drop than text-only models, suggesting visual inputs may amplify over-interpretation rather than improve evidential calibration.Systematic Error Patterns
- Evidence Over-Attribution (OverAttr = 0.58): Models label 58% of vague-evidence cases as "Evident"
- Evidence False-Attribution (FalseAttr = 0.585): 58.5% of incorrect diagnoses are accompanied by false claims of evidential confidence
- Hallucination: Models fabricate evidence quotes not present in student handwriting
- Over-inference: Drawing strong diagnostic conclusions from minimal or ambiguous cues
Implications for AI in Education
Evidence-Aware Design
Current LLMs lack the ability to calibrate confidence based on evidential quality. Educational AI systems need explicit mechanisms to detect when evidence is insufficient and communicate uncertainty rather than fabricate confident diagnoses. This connects to broader challenges in ground-truth-reliability-aied and hallucination-risk.Teacher-in-the-Loop Architecture
The authors advocate for human-in-the-loop systems where LLM predictions support rather than replace teacher judgment. This aligns with teacher-role research on AI augmentation vs. automation, and with findings from over-reliance studies showing that uncalibrated AI outputs can degrade human diagnostic accuracy.Benchmarking Beyond Accuracy
MathCog demonstrates that standard accuracy metrics obscure critical failure modes in educational AI. The evidential calibration metrics (OverAttr, FalseAttr) introduced here complement evaluation approaches in benchmark design and ai-ed-evaluation. The dataset's Evident/Vague distinction parallels concerns in cognitive-offloading about when learners (and AI systems) should recognize the limits of available information.Connection to LLM Fallacy Research
This paper directly extends work on llm-fallacy-misattribution — the hallucination of evidence and over-attribution patterns documented here are a specific educational manifestation of broader LLM reasoning failures. The finding that reasoning models (DeepSeek-R1, o1-Preview) show no systematic advantage echoes questions about whether chain-of-thought reasoning actually improves diagnostic reliability.Handwritten Math Assessment
MathCog builds on existing llm-handwritten-math-grading work but shifts focus from answer correctness to cognitive skill diagnosis — a more ambitious and educationally meaningful task. The TIMSS cognitive framework provides a theoretically grounded alternative to ad-hoc grading rubrics.Limitations
- Korean middle-school math only; generalizability to other languages, grade levels, and subjects unknown
- 3,036 verdicts across 639 responses — moderate dataset size
- Only TIMSS "Knowing" and "Applying" domains covered; "Reasoning" skills excluded due to problem set characteristics
- Korean-to-English machine translation may introduce artifacts
- Static benchmark; does not capture iterative diagnostic processes teachers use in practice
Related Pages
- llm-handwritten-math-grading — Prior work on LLM-based grading of handwritten math responses
- llm-fallacy-misattribution — Foundational study on LLM fallacy misattribution in educational contexts
- hallucination-risk — Broader framework for hallucination risks in AI education tools
- over-reliance — Over-reliance on AI outputs, including uncalibrated diagnostic confidence
- cognitive-offloading — Cognitive offloading dynamics relevant to evidential over-attribution
- ground-truth-reliability-aied — Ground truth reliability challenges in AIED benchmarks
- human-in-the-loop — Teacher-in-the-loop architectures for AI-augmented assessment
- teacher-role — Teacher role in AI-augmented educational workflows
- benchmark — Benchmark design considerations for educational AI evaluation
- automated-grading — Broader automated grading landscape, of which cognitive diagnosis is an extension- representation-robustness-llm-math-problem-solving — Representation Robustness under Executable Reasoning Constra