What it is
Reworks and scales the optimization behind combination loss analysis, then improves the solver with AlphaEvolve to tighten the matrix multiplication exponent bound. Instead of the prior SQP setup limited to smaller instances, it reformulates combination loss analysis for large-scale gradient optimization, computes max-entropy terms via Sinkhorn, and evolves the optimizer with AlphaEvolve.
Key results
- Certified ω < 2.371177 vs previous best 2.371339
- Gradient method alone improves SOTA by about 0.97×10^-4
- Scales from ℓ*=3 (~25k params) to ℓ*=4 (~7M params)
Shows that modern differentiable optimization and code evolution can improve computer-assisted bounds in fine-grained complexity.