Robert T. Lange

Open-Ended Evolution as an In-Silico Design Principle

Moodern AI systems are human designed and inspired by our own central nervous system. While this has proven powerful, it limits our ambition to our own cognitive creativity and constraints. Biological evolution, by contrast, has produced general intelligence in many diverse forms and under extreme environmental conditions. My work asks: what happens when we take evolution seriously as first-class design principles for developing in-silico intelligence? How can we safely build systems that endlessly improve their own discovery processes in a fully autonomous way?

The Scientific Method as Generalized Evolutionary Tree Search

Science and biological evolution share fundamental mechanisms: generate variants, test against reality, propagate survivors. Hypotheses mutate through recombination; experiments select for explanatory power; cultural diffusion — papers, code, conversation — serves as the inheritance medium. While Darwinian evolution is often dismissed as simple, its power arises from co-adaptation of the search process itself — gene regulation, adaptive mutation rates, and developmental plasticity evolve alongside the solutions they produce. Science is equally reflexive: new instruments reshape what questions we ask, and new theories reshape how we experiment. Framing invention as generalized evolutionary tree search unifies hypothesis generation, experimentation, and peer review into a single distributed loop — one that can, in principle, be dreamt of being automated.

A Paradigm Shift: Foundation Models as Generalized Evolutionary Engines

Numerical	Text/LLM-based
LES (2022) LGA (2023) LQD (2025) DiscoEnv (2024) DiscoRL (2024)	EvoLLM (2024) DiscoPOP (2024) The AI Scientist (2024/5,v2) Darwin Goedel Machine (2025) ShinkaEvolve (2025)

I deeply believe in Sara Hooker's hardware lottery thesis — that hardware progress shapes the manifold of possible research directions. Plateaus are often broken not by incremental refinement but by orthogonal innovation that creatively destroys the dominant consensus view. At the start of my PhD, the work of Jürgen Schmidhuber, Luke Metz, and David Ha inspired me to build hardware-accelerated evolutionary optimization for meta-evolution. My work since traces a shift from numerical to LLM-based evolutionary search. While numerical optimization can incorporate non-differentiable components and navigate ill-conditioned landscapes, it requires a fixed meta-substrate — typically a neural network — to parametrize and constrain the algorithm of interest. Foundation models break this constraint. An LLM can mutate hypotheses and program implementations directly — no fixed substrate needed — enabling open-ended variation beyond what any numerical encoding can express. LLMs bring inductive biases distilled from human knowledge, but code plus automated verification turns that into a "Darwin complete" search procedure.

On Interpolation and Extrapolation: The Hero Cult Fallacy of Invention

A common objection to LLM-driven discovery is that these models merely interpolate — they cannot truly create. But the force of this criticism depends entirely on the reference frame. Consider Picasso. He did not wake up with strokes of genius. He absorbed Miró, Dalí, Juan Gris, and African sculpture over decades — then recombined what he stole. Not plagiarism, but the iterative development of taste through trial and error: combinatorial variation filtered by aesthetic judgment. LLMs operate the same way, interpolating across a knowledge base far wider than any individual's. What looks like extrapolation from one person's vantage point is interpolation from a sufficiently broad one. The real frontier is not raw novelty but guided recombination and verification — and that is precisely what evolutionary search over LLM-generated variations may provide. Ultimately, AI will serve as a cognitive telescope — an instrument that extends our reach into spaces of knowledge beyond what any human mind can compute or comprehend alone.

The Vision: From Vibe Coding to Fully-Automated Scientific Discovery

The most ambitious interpretation of these ideas is the automation of the entire scientific research stack itself. The AI Scientist was the first system to demonstrate an LLM-powered pipeline for generating ML research papers end-to-end: ideation, experimentation, writing, and reviewing. The AI Scientist-v2 extended this to agentic tree search and produced the first AI-generated manuscript to pass peer review at an ML conference workshop. Beyond fully autonomous research, I am interested in how these tools can serve as a collaborative tool for human experts. E.g., in 2025, ShinkaEvolve played a crucial role in Team Unagi's victory at the ICFP Programming Contest. By evolving and refining SAT solver encodings against a suite of hard problems, it aided human researchers to tackle complex real-world problems.

Looking Ahead

My current research focuses on three directions:

1. Self-improving agents — systems that use evolutionary and meta-learning principles to continuously improve their own capabilities, including code optimization.
2. Open-ended agentic hypothesis iteration — moving beyond linear implementation toward systems that integrate evidence across long-lived trajectories.
3. UI/UX for automated science — tooling for human-AI collaboration, process augmentation and smooth steering of the search process.

The common thread: working towards evolving the scientific research stack itself.

A Human-Centered Perspective on Automated Discovery

The combination of evolutionary search, meta-optimization, and foundation models provides a concrete toolkit for building automated invention systems. Simultaneously, the capabilities of frontier models are improving at a staggering pace — we are literally just getting started. This opens up many fundamental questions regarding the future of every scientific discipline and the role of humans. How will we define a 'PhD student' in 2030? How will we assign credit to contributions? How do we circumvent a 'rich-get-richer' cycle of private research?

Personally, I do not believe that human researcher will entirely be replaced by AI in the near future. But I do believe that AI is already, and will increasingly play a significant role in how we search of idea-implementation spaces. Code is already to cheap to meter. In-silico research will follow. Current state-of-the-art systems for science are scratching GPT-3.5 level performance. As this continues to improve, humans will become 'good shepherds' steering herds of agents into a benevolent world. But this will require rewriting and implementing the social contract of research.

The hardest problems facing society are human-made. They demand knowledge discovery at a scale and speed that human researchers alone cannot sustain. Automated scientific discovery is our best bet. And we collectively have to start discussing andrethinking research credit assignment, democratization and safe-guarding — now.

Collaboration & Mentoring

I was fortunate to have people take a chance on me early on. If you have a chip on your shoulder and want to prove to the world that there's an adjacent way of thinking about evolution and automated discovery — reach out. Take convicted action.