About SynthCore
An autonomous research pipeline that turns recent mathematics into testable machine-learning ideas.
What is this?
SynthCore is an automated research project. It studies recent mathematics papers on arXiv and asks one question: what can this math do for machine learning? The pipeline extracts concrete, testable ideas and verifies them empirically. Every idea — verified or not — is published on this site.
Who's behind this
SynthCore is built and maintained by Maxim Zhivotok, a Python/ML engineer with 10+ years of experience in neural network architecture research, LLM and TTS fine-tuning, and applied ML systems.
LinkedIn · GitHub · Hugging Face · X · ORCID
The pipeline
- Fetch. Papers are fetched from arXiv (mathematics categories) and converted to text.
- Extract. An LLM screens each paper: is there an applicable mechanism? → structured ideas with usefulness, difficulty and novelty ratings.
- Stage 1 — toy mechanism gate. A cheap controlled experiment checks whether the phenomenon claimed by the idea reproduces. A failed gate stops the idea; only a pass proceeds to Stage 2.
- Benchmark. The idea is implemented on a real (small) neural network task and compared against a tuned baseline over 8 paired seeds; significance is computed with a permutation test, and the smallest detectable effect is reported.
How to read the verdicts
Each idea carries a status badge:
- Mechanism works — the mathematical mechanism reproduced on the toy system; a benchmark run is pending.
- Beats tuned baseline — a statistically significant practical win over a tuned baseline (8 paired seeds). Read it together with “Mechanism transferred to benchmark”: a win with no transfer evidence is promising performance with an unresolved cause.
- Mechanism confirmed, baseline not beaten — the math works, but the practical effect is not significant.
- Mechanism failed / Failed on benchmark — honest negative results.
- Unverified — not tested yet.
Three results, in order
Stage 1 · Toy mechanism gate asks whether the phenomenon claimed by the idea appears in a small controlled system. Failure stops the pipeline. After a pass, Stage 2 reports two separate results: mechanism transfer asks whether that same phenomenon still appears in the trained benchmark model, while the practical benchmark asks only whether task performance beats the tuned baseline.
| Toy gate | Benchmark mechanism | Practical result | Meaning |
|---|---|---|---|
| Passed | Confirmed | Wins | The claimed mechanism survived transfer and the implementation helped. |
| Passed | Confirmed | Not beaten | The mechanism transferred, but practical benefit was not established. |
| Passed | Not reproduced | Wins | The implementation won, but not for a demonstrated version of the claimed mechanism. The cause is unresolved and may be another effect. |
| Failed | Not run | Not run | The claimed toy phenomenon failed, so the idea did not proceed to Stage 2. |
Honesty & limitations
Ideas are AI-generated hypotheses, not peer-reviewed results. Verdicts are computed by deterministic test code, not by the language model. A "beats baseline" result means: on this particular small benchmark, with this tuning budget, the idea won with statistical significance. It does not guarantee gains at other scales. Failures and non-results are published too — that is the point.
Tech
The site is a fast read-only view over the pipeline's SQLite database (Go + HTMX, server-rendered). Every idea exposes structured JSON (/api/idea/{id}.json and JSON-LD metadata), and each verified idea ships its full experiment code and reports — downloadable as a ZIP or readable online.
License
All experiment code artifacts published on this site are released under the MIT License: you may use, copy, modify and reuse them, including commercially, with the copyright notice preserved. The idea descriptions and analysis texts may be quoted and reused with attribution and a link to the corresponding page.