Open-ended ("free range") research optimization

[BradKML]

There are a lot of papers that are similar to Absolute Zero, and I do think that having the same LLM split between two different roles and self-duel is a good idea. However, they are often stuck on a single task type. Also, not all self-play has to be so adversarial; it can be that one side suggests fun things to do, and the other has to discover something, anything.

List of stuff with the same Self-Play theme

• Absolute Zero Reasoner (AZR) https://github.com/LeapLabTHU/Absolute-Zero-Reasoner
• Self-Play Fine-Tuning (SPIN) https://github.com/uclaml/SPIN
• Self-Play Preference Optimization (SPPO) https://github.com/uclaml/SPPO
• Self-Play Adversarial Language Game (SPAG) https://github.com/Linear95/SPAG
• Self-Play on Zero-Sum Games (SPIRAL) https://github.com/spiral-rl/spiral
• Self-Play Reinforcement Learning (SeRL) https://github.com/wantbook-book/SeRL
• Self-play Critic (SPC) https://github.com/chen-judge/SPC
• Some stuff under peer-review https://openreview.net/forum?id=ed75tWzgt0

In the context of OpenEvolve, there are a few things to consider:

1. Since AI-Scientist kind of already exists, instead of reinventing the wheel multiple times, it should also be able to pass existing novel research to something like OpenEvolve to enhance the corresponding algorithms, before passing back to AI-Scientist for further remixing and diversification of goals Something new to mix in: Research/Paper Database #96 Adding PKMs into the system SakanaAI/AI-Scientist-v2#48
2. DGM and ASI-ARCH exist alongside OpenEvolve, and there has to be a way of unifying all three under a single meta-architecture. The optimization metric would be acceleration per compute-hours (instead of research per compute-hours, which is more of an ASI-ARCH thing) [feature] Implement ideas from the Darwin-Godel Machine paper to evolve the openevolve agent itself #53 AI Scientist 3 with SEAL + DGM/AlphaEvolve? SakanaAI/AI-Scientist-v2#53
3. Multi-task reinforcement learning by self-play is still a hot problem to solve, since given the same skill list and the same environment, there have to be ways that RL itself can be accelerated and made more effective.

Cross-reference to simulated open world training with a focus on "interestingness" jennyzzt/omni#2 (comment)

[BradKML]

We kinda need human-supervised research programs since not all data on Kaggle is clean, so it can have GIGO not at a "bad idea" level but a "bad data" level, coding tasks are easier to self-correct than data science in this regad SakanaAI/AI-Scientist-v2#2 (comment)

[BradKML]

To lower friction, this tool seemed useful https://github.com/SWE-agent/mini-SWE-agent