1. twuai.com
  2. #papers
#airesearch#papers#short#math#lehmersproblem#mahlermeasure#codingagents#lego#deeplearning#numerics#llm#pretraining#learningrate
#airesearch#papers#short#math#lehmersproblem#mahlermeasure#codingagents#lego#deeplearning#numerics#llm#pretraining#learningrate





hace 24 d铆as

Can a human鈥揂I loop fill every degree of Lehmers problem? | Generalized Boyd 路 2026.10.21

In 1933, Lehmer asked whether the Mahler measure of an integer polynomial can be bigger than 1 yet arbitrarily close to it.

His own degree-10 polynomial, at 1.17628, is still the smallest known.

Systematic searches for the smallest value at each degree stopped at degree 180.

Boyd's one-parameter family keeps factoring into cyclotomic pieces, so it skips nearly half of the target degrees.

In this paper, a human researcher and a Gemini agent built a fast numerical solver and an exact verifier together and ran large searches on cloud clusters.

Five different search methods all landed near the same value, 1.2554.

Multiplying the degree-188 result by a cyclotomic polynomial collapsed 111 terms to 7 and revealed Boyd's two-variable polynomial.

That led to Generalized Boyd, a five-parameter construction that solves for the exponents backward from the target degree.

On the same benchmark, the original family filled 504 degrees.

Generalized Boyd filled all 995 even degrees up to 2,000, each with an irreducible polynomial below Boyd's limit point.

Against 91 published records, the authors' searches matched 85 and beat 6, and Generalized Boyd now holds 99.6% of the best-known records.

Lehmer's question stays open.

Beyond degree 2,000 the behavior is predicted by asymptotic analysis, not computed.

Past degree 250 there are no published records to compare against.

There is no baseline without the AI, and the model is an internal Gemini version.

Worth reading if you are building a human鈥揂I search loop for mathematics.

馃搫 arXiv:2610.30277 (Google DeepMind, Caltech)

#AIresearch #papers #math #LehmersProblem #MahlerMeasure #short

hace 24 d铆as

Can two tools make a coding agents LEGO buildable? | BrickWright 路 2026.10.22

Ask a coding agent for a LEGO model and the render looks fine, but bricks pass through each other, float, or aren't attached. On 512 prompts, the baseline agents the authors ran produced physically valid builds only 9% to...

hace 24 d铆as

Does standard softmax hide a bug? | Softmax Fix 路 2026.10.22

Every major framework computes softmax and cross-entropy by subtracting the largest logit first. Once the correct logit's margin passes 16.6 in fp32, or 5.5 in bf16, its loss and gradient round to exactly zero, and the gradients no longer...

hace 25 d铆as

Can you pick the data schedule without the learning rate? | Data 脳 LR Schedules 路 2026.10.21

Mid-training usually trains on web text and then switches completely to scarce, high-value data like math or code at the very end. Earlier work found that replaying web data can help the target domain, but it compared data schedules...