Mathematical Research Capabilities: Proven AI Breakthroughs

We are all forced to revise our assessments of the mathematical capabilities of large language models. For a long time, it was easy to dismiss LLM-generated proofs as mere regurgitation of existing literature. If a model solved a problem, we told ourselves it was just "noticing" an answer already sitting in the archives. But the laughter has grown quiet. My recent experience with ChatGPT 5.5 Pro suggests that we have crossed a threshold where these models are no longer just summarizing; they are actively contributing to research-level mathematics.

The real shift happens when you stop asking these models to solve famous, well-trodden problems and start feeding them the "low-hanging fruit" of modern combinatorics. I recently tested ChatGPT 5.5 Pro on a series of problems posed by Mel Nathanson regarding sumsets in additive number theory. Specifically, I looked at the diameter required for a set to achieve prescribed sumset sizes.

Here is where most people get tripped up: they assume the model is just hallucinating or stitching together random snippets. In this case, the model spent roughly 17 minutes thinking before producing a construction that yielded a quadratic upper bound—a result that was clearly optimal. When I asked it to format the output as a LaTeX preprint, it did so in under three minutes. The logic wasn't just a copy-paste job; it effectively optimized the construction by selecting a more efficient Sidon set than the one used in the original paper.

This raises a difficult question: does the model have original ideas, or is it just better at synthesizing existing techniques than we are? Most human mathematics consists of exactly that—putting together existing knowledge in novel ways. If an LLM can identify that a specific Sidon set construction improves a bound in a paper that hasn't received much attention, is that not research?

The implications for the field are profound. We used to rely on these "open" problems as training grounds for early-career mathematicians. Now, the bar has been raised. It is no longer enough for a problem to be open; it must be hard enough that an LLM cannot solve it within an hour. If you are working on additive number theory research, you are effectively competing against a system that can iterate through construction parameters at a speed no human can match.

That said, there is a catch. When I pushed the model toward more complex, general cases involving higher-fold sumsets, it hit a wall. The difficulty of these problems often lies in the structural complexity of the sets themselves, not just the optimization of parameters. While the model can handle the "easy" arguments we’ve missed, it still struggles when the underlying theory requires a deep, non-obvious leap that isn't hinted at in the existing literature.

We are entering an era where the definition of a "mathematician" is shifting. If you aren't using these tools to stress-test your own conjectures, you are likely missing out on potential improvements to your work. The future of mathematical discovery will belong to those who can effectively prompt these systems to explore the gaps in our current knowledge. Try this today with a problem you’ve been stuck on and share what you find in the comments.