Context
January–May 2026 research with Prof. Krishna Narayanan at Texas A&M. This is ongoing exploratory work, not a finished compression result. Public repos: GenDiff and llmzip_extension.
The algorithm
Generative reconstruction compression: instead of encoding the file token-by-token, the sender transmits a prompt, a seed, and a compact diff against a jointly generated candidate. The receiver regenerates that candidate and applies the diff. The cover is the existing GRC flow (gendiff-grc-flow.png): a seed, entropy-coded opcodes, and context-aware content bits go out; the receiver regenerates the raw output and applies the diff. Still an open question, not a finished result.
The same January–May 2026 window also included LLMZip-style arithmetic coding over Hugging Face causal LMs. Those runs belong to the research. They are not a finished comparison.
My role
Sole implementer of the GenDiff prototype and the extension experiments. Literature starting point was the public LLMZip paper/repo.
Tech
Python, PyTorch, Transformers, custom arithmetic coder (core/AC), matplotlib.
Dates
January–May 2026. GenDiff experiments in that window (April 2026).
