Next-Concept Modeling on code: tokenizer fixes and a plan/order-level extension (setup summary, 2026-11-19)
Setup summary and code snapshot for an independent study of Next-Concept Modeling (NCM, arXiv 2610.14162) on Python code: tokenizer-training fixes for code reconstruction, and an extension of NCM's generator with a plan level and an order level, compared with parameter-matched and scratchpad autoregressive baselines on synthetic order-inference tasks. Files are restricted; access on request. summary.pdf sha256 881978595e0de53da87dc29e155070f6e9a0eedfd1730ab67b5eb1e42533359dcode_tok_de98cac8faf9.tar.gz sha256 7abc60466c5f264eb0bb68367fbc17f90cfcc8d9a6d01614016f73416e0daae2 (git commit de98cac8faf9560b806557ff3a04c57edfffb67d)
Authors
- Rupam Neogi (ORCID: https://orcid.org/0009-0006-1442-2847)
Institutions
- Sister Nivedita University (IN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24373348
- Primary Topic
- Software Engineering Research
- Type
- preprint