EULER system generates 10 proofs from 120 mathematical predictions

Summary

EULER is a multi-agent system that treats cross-domain mathematical transfer (a 'bridge') as its fundamental search unit; it generated 10 proofs, 3 refutations, and 45 partial results from 120 frozen combinatorial conjectures.

Why it matters

EULER centers on the verification pipeline—six sequential stress tests and a verified result return path—acting as a guardian of correctness rather than relying on brute-force theorem proving; ablation studies show this is the element that actually reduces errors. The finding that domain distance does not predict success, whereas executable operation gain does, offers a more principled metric for when AI can add value in mathematics.

Highlights

• 10 proofs + 3 refutations from 120 contamination-screened conjectures by JCTA authors
• Stress tests reduced erroneous results from 9 to 3; bridge + local operation interaction added +4.2 solved tasks
• Domain distance is insignificant; executable operation gain and valid return path are the true predictors of success

Critical take

A full proof rate of 8.3% in a single-domain (combinatorics) benchmark is modest, and the entire evaluation was drawn from the recent author pool of a single journal; therefore, generalizability to other mathematical fields remains untested. The statistically noted +4.2 interaction effect is small enough to cast doubt on conjecture difficulty and cross-domain robustness.