Not evaluated
OEIS Open: general LLMs solve 147 of 492 formalised OEIS conjectures
OEIS Open: general LLMs solve 147 of 492 Lean-formalised OEIS conjectures
The gist
Adamczewski and co-authors (arXiv 2608.11941, 2026-08) report that general language models solve 147 of 492 OEIS conjectures formalised in Lean, about 30%. Status: not yet evaluated.
DiscoverIA status on October 11, 2026: not evaluated, meaning no third-party evaluation found.
AI draft, not reviewedWhat has not been reviewed?
Text written by an AI (DiscoverIA pipeline), on . The status and its sources come from the registry; only the explanatory text has not been reviewed. A reported error is corrected, then the text is reviewed.
Cite this entry
A single status snapshot since Oct 11, 2026: a citation always points to a dated snapshot.
Cite snapshot 1, Oct 11, 2026
APA citation with the dated link
BibTeX @misc, key discoveria-oeis-open-benchmark-r1
@misc{discoveria-oeis-open-benchmark-r1,
author = {{DiscoverIA}},
title = {OEIS Open: general LLMs solve 147 of 492 Lean-formalised OEIS conjectures},
howpublished = {DiscoverIA registry, entry oeis-open-benchmark},
note = {Status “Not evaluated”, snapshot 1, recorded on October 11, 2026. Benchmark},
year = {2026},
url = {https://discoveria.ai/entry/oeis-open-benchmark#releve-1},
urldate = {2026-10-11}
}Dated link opens this snapshot in the history
To understand ≈ 20 s reading time, estimated
What it changes
An open benchmark provides a shared measure for comparing models on conjectures about number sequences, checked by Lean. If the figures held, about 30% of these conjectures would already be within reach of general-purpose models.
What remains unclear
The results are checked by the authors themselves, with no named third party. They depend on a per-attempt compute budget that this entry does not detail. The date is known only to the month.
In the glossary : Benchmark
Read the paper
OEIS Open: How many conjectures can language models turn into theorems?
What it establishes : Adamczewski measures the share of Lean-formalised OEIS conjectures that general-purpose language models can prove with minimal tools: Claude Opus 4.8 solves 30%, GPT-5.5 26% and Gemini 3.5 Flash 22%.
Where to start : The abstract, then Figure 1 (Section 3.1), which compares the models with AlphaProof Nexus. Section 2.3 gives the score recomputed with the Comparator checker (144 out of 492), Section 4.1 the risk of misformalisation.
Written by an AI · reviewed by a human
For experts
Sources, method, key facts, status history, report an error
Status
Not evaluated
Recorded on · arxiv.org
Step 1 of 4 on the scale: not evaluated. Next step: under review.
- Announced
- Recordednot evaluated
- nextTo follow
What this status means
Source reading : Source read word for word (level 3 of 3)
The result is declared neither right nor wrong.
To move to “Under review”: a public review under way. To “Peer reviewed”: a journal publication or a review by named experts.
Sources
- Main sourcearxiv.orgpage read word for word
Method
DiscoverIA status: not evaluated, meaning that no third-party evaluation has been found in the sources read. Set on October 11, 2026 from arxiv.org (page read, quotation checked word for word: “resolve 147 of these conjectures with a budget of”).
Limits
The claimant supplies a Lean proof; as of this date, DiscoverIA has found no independent verification of this proof or of the result.
Key facts
- Claimed by
- Adamczewski et al. (arXiv 2608.11941)
announced on Aug 2026 - Nature
- benchmark
- Field
- Mathematics (mathematics / benchmark)
- Checked by
- authors
- Sources
- 1 : 1 read word for word
- This entry
- AI draft, not reviewed
Status history
Only one observation so far.
An observation is added; none is ever erased.
Report an error or suggest a source
Thank you, your message was sent.
Sending failed. Try again, or write to contact@discoveria.ai.
Is your result listed?What each status requires, correcting this entry, replying to it.