“Update: Sadly, after deeper testing, most of the free OpenCode model’s ideas didn’t…”

Update: Sadly, after deeper testing, most of the free OpenCode model’s ideas didn’t pass verification. 😢

Frontier models are much better at verifying their own work first.

However, we did find useful context-sharing strategies across parallel agents that actually do speed up time to the next breakthrough (in the frontier research rewound benchmark and the current best kappa).

I submitted a new best bound this morning within 10 minutes using this method, with Astra high coordinating a fleet of agents.

Strategies to solve the “local minima” problem, like RSI-top, fork and flush, and “forgetting,” are now implemented in our open auto-research agent:

https://github.com/rohanarun/Auto-research-agent

I’m running benchmarks to prove they work for this task, but early results are interesting:

1. Forgetting seems to be one solution to the “local minima” problem: agents stop getting stuck in idea wells.

2. For some tasks, like Integer Mult, we may need the maintainer to specify the scoring mechanism. Most submissions score 0, yet negative research is still helpful.

I’m also working on our new model release, so it feels like we have a big research lab here, running queued experiments across multiple tracks and projects. :)

Quoting @RohanArun: BREAKING: A free model, @opencode Space Bunny Max, has successfully proposed a new best bound for Integer Multiplication!

Conditional κ = 0.001730189061233850760095

Claude Opus 5.5 assisted in the verification and PR push to pull it over the line.

The barrier of entry just dropped to 0.

Previously the expensive max subscriptions prevented some of our best contributors and experts from participating, but now anyone can :)

I’ve merged OpenCode as an optional CLI into our Auto-research agent: https://github.com/rohanarun/Auto-research-agent

https://github.com/CrocSwap/integer-mult-bounds/pull/3184