AI produces plausible explanations easily. In compression research, that is not enough. A file must be restored byte for byte, and the result can be compared numerically.
Cubrim is a data-compression research project. By my account, several models and an engineering workflow help generate, challenge, and test hypotheses. The most eloquent agent does not decide which idea wins. The benchmark does.
The published result
Cubrim's overall ranking compares aggregate ratio across several corpora. Lower is better within this table.
| Rank | Archiver | Mode | Aggregate ratio |
|---|---|---|---|
| 1 | cubrim | competitive | 0.2204 |
| 2 | ppmd | 7z -m0=PPMd | 0.2286 |
| 3 | xz | -9e | 0.2344 |
| 4 | 7z | -m0=LZMA2 -mx9 | 0.2355 |
| 5 | brotli | -q 11 | 0.2408 |
| 6 | zstd | --ultra -22 | 0.2490 |
| 7 | rar | a -m5 | 0.2573 |
| 8 | bzip2 | -9 | 0.2671 |
| 9 | gzip | -9 | 0.3330 |
| 10 | lz4 | -12 | 0.3822 |
The defensible conclusion is narrow: in the published overall ranking for the stated corpora and settings, Cubrim ranks first with an aggregate ratio of 0.2204.
The table does not prove superiority on every kind of data, under every memory or speed constraint, or in an independent benchmark. The project publishes the table itself. A change in corpora, version, or mode requires a new measurement.
Why agents matter here
One person normally tests hypotheses sequentially. Several agent roles can pursue parallel lines: propose transformations, search for counterexamples, analyze failures, and prepare implementations.
This is useful only because the criterion sits outside the model's reasoning. The archiver either restores the original data or it does not. The ratio either improves on the specified set or becomes worse.
Cubrim does not prove that AI can conduct any research independently. It shows something more practical: one person can expand the space of testable hypotheses when an external criterion is defined in advance and explanations cannot replace measurement.
For me, the main result is not simply the first line of the table. It is that one person could organize research as a working system: pursue alternatives in parallel, confront them with counterexamples, implement the strongest candidates, and leave the decision to measurement. The ranking may change. The ability to carry a hypothesis through a complete, verifiable cycle remains.
Source and verification date
- Cubrim Evolution Benchmark, checked on July 21, 2026.
The table, versions, and settings must be checked again before publication.