We are just scratching the surface of using AI for major discoveries.

And no, LLMs are not dead.

AlphaResearch tackles a fundamental challenge: can LLMs discover truly novel algorithms that push beyond existing human knowledge?

Previous approaches like AlphaEvolve rely solely on code execution (technically correct but potentially uninteresting solutions), while pure idea-generation systems produce creative concepts that may be computationally infeasible.

AlphaResearch solves this by combining execution-based verification with a reward model (AlphaResearch-RM-7B) trained on 24K ICLR peer-review records, creating a "dual research environment" that ensures both feasibility and innovation.

On the circle-packing problem, it discovered an algorithm achieving 2.939 total radii (n=32), surpassing the best human solution (2.936) and AlphaEvolve (2.937).

It might not look like a big deal, but this demonstrates the potential of LLMs that can autonomously produce genuinely novel, state-of-the-art algorithms.

This is significant because it shows AI can actively advance the frontiers of human research, not just reproduce existing knowledge.

Paper: arxiv.org/abs/2511.08522

Track trending papers here: nlp.elvissaravia.com