Recovering the latent cognitive structure of standardized examinations, and using it to generate personalized study plans.

LearnOpt is an open-source pipeline that:

1. Takes historical exam questions (currently NEET 2016-2024 and JEE Advanced 2016-2023)
2. Tags each question with a skill category (S1-S5: Recall, Application, Multi-concept Integration, Quantitative Reasoning, Elimination/Negation) using an LLM
3. Measures the latent skill distribution of the exam over time -- and detects when it changes (e.g. due to syllabus rationalization)
4. Builds a knowledge graph of topics, prerequisites, and skills
5. Generates a personalized, time-bounded study plan for a student, given their self-assessed mastery, using constrained optimization (OR-Tools)

The accompanying paper (paper/) found that NEET's latent skill distribution was stable from 2016-2021, underwent a statistically significant shift (KL=0.040, permutation p=0.0005) coinciding with NCERT's 2023 syllabus rationalization, and that the same pipeline applied to JEE Advanced shows a markedly different cognitive profile (KL=0.505 vs NEET) -- suggesting exam tier shapes cognitive structure more than subject, which shapes it more than time within a stable regime.

Most exam prep tools tell you "study high-weightage chapters." LearnOpt tells you what kind of thinking the exam tests, whether that has recently changed, and -- given your own strengths and weaknesses -- what to prioritize next.

Unlike commercial systems (e.g. Embibe's PAJ, which relies on a 75,000-node manually-curated knowledge graph and proprietary behavioral data from millions of students), LearnOpt is built entirely from publicly available exam papers using LLM automation, and is designed so any student can run their own calibration and get a personalized plan -- no proprietary data required.

learnopt/
├── paper/                      # arXiv paper source (markdown)
├── pipeline/
│   ├── 01_tag_questions.py    # LLM tagging (Groq + Ollama)
│   ├── 02_analyze_skills.py   # skill distribution, KL divergence, permutation tests
│   ├── 03_build_graph.py      # FalkorDB knowledge graph ingestion
│   └── 04_optimize.py         # OR-Tools study plan generator
├── prompts/
│   └── skill_taxonomy_prompt.md
├── data/
│   ├── neet_2016-2024_tagged.csv
│   └── jee_advanced_2016-2023_tagged.csv
├── student_calibration/
│   ├── calibration_template.md  # fill this out about yourself
│   └── calibrate.py              # turns your answers into a mastery vector
└── notebooks/
    └── reproduce_section4.ipynb

1. Copy student_calibration/calibration_template.md and fill it out (topic difficulty/frequency ratings + your self-assessed mastery per topic -- takes about 30-60 minutes).
2. Run:

   python3 student_calibration/calibrate.py my_calibration.md > mastery_vector.json
   python3 pipeline/04_optimize.py --mastery mastery_vector.json --hours 150 --exam neet

3. Get a ranked study plan: which topics to prioritize, which to skip, and why.

pip install -r requirements.txt
python3 pipeline/01_tag_questions.py --input data/raw/neet_questions.csv --output tagged.jsonl
python3 pipeline/02_analyze_skills.py --input tagged.jsonl

Or open notebooks/reproduce_section4.ipynb for the full Section 4 analysis (skill distributions, KL divergence tables, permutation tests, JEE comparison).

Tagged question data is also available on HuggingFace:

• joyboseroy/neet-skill-tags-2016-2024
• joyboseroy/jee-advanced-skill-tags-2016-2023

The underlying NEET question text is sourced from dipikakhullar/neet (fair use); JEE Advanced questions are sourced from JEEBench (Arora et al., EMNLP 2023).

See prompts/skill_taxonomy_prompt.md for the full tagging prompt and paper/ Section 3.4 for the taxonomy's justification.

If you use this code or data, please cite:

@article{bose2026learnopt,
  title   = {LearnOpt: Recovering the Latent Cognitive Structure of Standardized Examinations via Knowledge Graphs and Constrained Optimization},
  author  = {Bose, Joy and Thomas, Om},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026},
  url     = {https://arxiv.org/abs/XXXX.XXXXX}
}

@dataset{bose2026neetskilltags,
  title     = {NEET Skill Tags 2016--2024},
  author    = {Bose, Joy and Thomas, Om},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/joyboseroy/neet-skill-tags-2016-2024}
}

This is an active research project. The skill-distribution analysis (Section 4.1-4.4, 4.6 of the paper) is complete and reproducible. The optimization/calibration tooling (Section 4.5) is under active development -- contributions welcome.

Code: MIT. Data: see individual dataset cards for upstream licensing (NEET questions are "fair use" per source; please verify before commercial use).