EPFL’s LinCodeEvolve (Viazovska, Abbe) finds seven record-breaking binary linear codes with LLM-guided program search
Result confirmed
Status
- Claim
Result confirmed
- Our reporting
- High confidence
- Verification
- Expert-checked
- Importance
- 3 of 5
- Last verified
- 30 September 2026
Your AI and this story
- GPT-6.2 Astra121 days after its cutoff
- Claude Opus 5.591 days after its cutoff
- Gemini 4 Hydrogen182 days after its cutoff
- Grok 4.7121 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 91 days before it.
Key facts
- arXiv 2609.37056 [cs.IT], 29 Sep 2026; authors Seddas, Shashkov, Viazovska, Abbe (EPFL)
- Seven new record codes: [172,21,66], [173,20,68], [176,21,68], [181,21,70], [184,21,72], [189,22,72], [200,21,77]; six have concise quasi-cyclic descriptions
- Standard code modifications extend these to 22 improved entries of the binary tables at codetables.de; ‘the results have been checked by the maintainer of the code tables and will be incorporated into them’
- Method: an LLM proposes code-construction programs (a chain of oracle, strategist, contract-repair and implementer calls), scored by an exact minimum-distance evaluator; a strategy loop with expert supervision redirects the search when progress plateaus
- Comparison: directly prompting GPT-6 Astra at extra-high reasoning effort gave distance 65 for (172,21) vs 66 from LinCodeEvolve; at (200,21) both reached 77, but LinCodeEvolve’s code has 25 minimum-weight codewords vs 75
- Compute: a single 80 GB H100; each experiment took 3–4 hours on average
What happened
Finding binary linear codes with large minimum distance is a central coding-theory problem, and certifying minimum distance is NP-hard. LinCodeEvolve keeps ShinkaEvolve’s island-based archive, novelty judge and model selection. Candidates enter the archive only with an exact certificate (full weight distribution). Among codes with equal distance it prefers fewer minimum-weight codewords.
Why it matters
Viazovska, who solved sphere packing in dimensions 8 and 24, is now co-authoring LLM-search papers. It is another case of FunSearch-style search improving a long-maintained table of records with modest compute (one H100). The authors note that careful prompting of a frontier model (GPT-6 Astra) came close on some parameters, but systematic search did better across many parameter pairs.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 2 primary
Primary
Changes
- Filed