CheckingNemotron 3 Super and UltraNebius Token FactoryarXiv, last six months
What's still unsolved in your field?
80 papers. About a minute. Every quote verified against its source. Swarm reads this year's research in parallel on NVIDIA Nemotron and returns the problems its authors couldn't solve, ranked by how many independent labs raise them.
- 80
- papers read in parallel for one question
- 63 s
- from question to ranked open problems
- 0
- quotes shown that aren't in the paper word for word
From one real run on October 29, 2026, on Nemotron through Nebius Token Factory: 90 parallel Super calls, 167 of 178 proposed quotes kept. Record in reports/seventh-run.json.
Methods
Language models invent research ideas that sound good and hold up badly: in a 2025 study, model ideas scored worse than expert ideas once people carried them out (arXiv 2506.20803 >). So Swarm generates nothing. It extracts what experts already wrote, and proves it.
- 01.TranslateNemotron 3 Super maps your words to the field's own vocabulary and writes the searches.1 call
- 02.RetrieveExact-phrase searches of arXiv, through its API or OpenAlex, last six months, by relevance, top 80 kept.up to 7 searches
- 03.ReadFull text from arXiv HTML. Only limitations, future work, discussion and conclusions go to the model.78 of 80 in full
- 04.ExtractOne Nemotron 3 Super call per paper, all at once, copying sentences that state an unsolved problem.64 in flight
- 05.VerifyEach quote must appear in its paper verbatim after Unicode NFKC normalization, or it's discarded.167 of 178 kept
- 06.RankProblems rank by independent groups: connected components of papers linked by a shared author. A group that spans papers counts only after a Super call names the shared claim and each paper's quotes share a specific word with it.1 check per group, top 6 shown
- 07.FalsifyA fresh search per problem for newer work. A claimed fix is shown only with a quote checked against that paper's abstract.1 search each
- 08.ExplainNemotron 3 Ultra writes for your level and hardware: what it means, what it needs, one first experiment.1 call
Definitions
norm(s) = collapse(fold(lower(NFKC(s)))) keep(q, P) ⇔ |words(q)| ≥ 5 ∧ norm(q) ⊑ norm(P) groups(G) = |components(G)|, (a, b) ∈ E(G) ⇔ authors(a) ∩ authors(b) ≠ ∅ rank = sort by (groups, |statements|) descending, top 6, possibly solved last merge(R) = round robin over R₁ … Rₙ by rank, deduplicated, first 80 Δt(arXiv) ≥ 3 s, server-wide
q is a proposed quote, P its paper's text, ⊑ means "is a substring of". G is the graph of a problem's papers, joined where they share an author.
| Automated tests | 102 backend (pytest), 17 frontend (node --test), all passing |
|---|---|
| Covered | Invented quotes dropped, shared authors merged, unfound terms dropped, search fallbacks, arXiv retries |
| Offline stand-in | The full pipeline with no network and no keys, so every test is reproducible |
| Continuous integration | Lint, format, types, tests and a Docker build on every push |
| Sentence links | 16 of 16 quotes in two real runs located on their arXiv page |
| Field runs | 13 real runs, each recorded in reports/, each failure fixed and tested |
| Models | NVIDIA Nemotron 3 Super and Ultra, on Nebius Token Factory |
|---|---|
| Throughput | 1,000 parallel Super calls in 25.7 s, benchmarked |
| One run | 80 papers, 92 model calls, 128,218 tokens, 63 s |
| Fabricated quotes shown | 0, by construction |
| Prompt injection | Paper text is fenced as data; instructions inside it are ignored |
| Source discipline | One arXiv request every 3 s, server-wide, as arXiv asks |
| Provenance | Every quote opens its paper at the exact sentence, highlighted |
| Exports | BibTeX for every paper, and the full run as a Markdown report |
| Reproducibility | Every run saved and replayable from its link |
Limits: arXiv only, the last six months, and only problems authors chose to write down. Explanations are written by a model and labeled as such.