Thoughts on OpenAI achieving IMO Gold

The fact that OpenAI achieving IMO Gold is not the impressive bit. DeepMind was 1 point below Gold medal threshold in last year IMO. What is truly impressive is OpenAI achieved Gold with a general purpose reasoning model (in contrast to a Lean based specialized system used by DeepMind), no external tool use (e.g. internet, python, Lean), and no verified reward.

Why is it a big deal?

If what OpenAI claims is true, it is using the same reasoning model that powers their Deep Research and ChatGPT agents to achieve Gold in IMO (see https://x.com/millionint/status/1946558130906968330). This is the very definition of AGI, a general purpose AI that is capable of handling tasks in different domains. Moreover, the Gold medal is achieved without external tool use, i.e. the model churned out natural language proof like human does without resorting to obscure methods like writing python or Lean code

How did they do it?

In my opinion, OpenAI is doing one thing extremely well that all other frontier labs seem to be struggling with: getting RL to work in domains with no verified rewards. In other words, they have a general-purpose verifier beyond just rule-based final answer checking. As a result, their reasoning model can work not only in domains with easily verifiable answers like math problems with numeric solutions and coding questions, but can also generalize to hard-to-verify tasks like travel planning and drafting research reports.

Gathering from various sources, the reasoning model is post-trained end-to-end with RL. That means they managed to derive a reward signal that guides RL training on general domains—one that is extremely hard to hack. Compared to other reasoning models, it is exceptionally good at performing long-horizon reasoning (say, >20 minutes) while remaining logically coherent.

Many have postulated that some variant of an LLM-as-a-judge reward model is being used. In fact, DeepSeek published some work a few months ago in that direction, but it doesn’t appear to have gone anywhere. No one has yet figured out how OpenAI is able to pull this off—i.e., end-to-end RL training in domains with no verified rewards

Multi-agent methods?

I personally suspect that the model that achieved IMO Gold is based on the aforementioned o3/o4 reasoning model plus some sort of multi-agent method— i.e. multiple reasoning models cooperate with each other to achieve a collective superior result.

A subtle hint came from @polynoamail, lead of OpenAI multi-agent team:

"A small team took a research idea few believed in and used it to achieve a result fewer thought possible"

Reference: https://x.com/polynoamial/status/1946478249187377206

This suggests the core innovation is likely something that people have thought about but few thought would work out