Research Papers
The future of AI isn’t dependent on humans. Today, we’re announcing ‘SIA,’ an open-source Self-Improving AI, designed to accelerate superintelligence through recursively improving loops. This marks the beginning of a new era of accelerated discovery and technology development.
The limitation isn’t intelligence. It’s the system
Modern research is still fundamentally linear. One hypothesis leads to one experiment, which leads to one result. Even the most advanced labs can only explore a fraction of the possible solution space. Breakthroughs still happen, but they happen slowly and often unpredictably, limited by time, funding, and human capacity.
What changes when research becomes a system that improves itself is the structure of progress itself.
Instead of isolated experiments, you get a continuous loop. Hypotheses are generated, tested, evaluated, and refined, and each cycle improves the next. This loop runs continuously and at scale, turning research into something that compounds over time rather than something that advances step by step.
Introducing SIA
SIA (Self Improving AI) is designed around this idea of recursive self improvement. Today, we’re open sourcing a new repo, which is the core framework behind an agent’s ability to improve itself.
SIA focuses on one problem: how do you design structured feedback loops that allow a system to evaluate its own performance, adapt its strategy, and get better over time?
SIA is available on GitHub today.
Key capabilities
The agent can update its own harness
The agent can update the weights of its own underlying model
The agent can update its own memory layer to support new complexities
Most agents are static frameworks (harness, model, memory layer) designed to execute specific multi-step operations. They plan, act, and use tools. SIA is focused on a different layer: the ability of the agent to evaluate and improve itself (it’s own harness, model and memory layer) based on feedback. It is not a replacement for existing frameworks. It is foundational infrastructure for building agents that get better the more they run.
SIA is the world’s first Self Improving AI system that can update not just it’s own harness but also it’s own weights.
It’s a self-improving agent that runs the full research loop, generating hypotheses, executing experiments, evaluating outcomes, and refining its own capabilities based on results. Each cycle increases its capability, allowing it to take on more complex and open-ended problems over time.
SIA is a framework to help design agents on one core capability, improving their own performance through structured feedback loops. It runs experiments, evaluates outcomes, updates its approach based on what worked, and runs again. This creates the foundation for long-horizon work, where meaningful progress emerges from thousands of iterations rather than a single moment of insight.
Self Improvement is all you need
In line with the bitter lesson, SIA outperforms specialised agents at diverse tasks by self improving. Yes, SIA is a general purpose agent that outperforms an agent trained to specialise on legal work on LawBench, an ML Research agent on MLE Bench and a biology focused agent on Denoising RNA samples.
MLE-bench is OpenAI's benchmark of evaluating agents on their ability to train ML models. SIA was tasked to tackle one of the hardest tasks from MLE Bench: Google Brain's Ventilator Pressure Prediction problem, a medical time-series problem where the model must predict airway pressure during mechanical ventilation from control-signal inputs. SIA did not just beat other agents but repeatedly beat it’s own performance to self improve on the task.
Open by design
We are building SIA in the open because we believe systems that have such powerful capabilities must be available to the world and not confined with a few.
Grounded in real research
SIA is being developed alongside real research partners to ensure it is grounded in practical use.We work with researchers at Stanford, University of Oxford, University of California, Santa Barbara on a range of scientific problems across Quantum research, New energy, Particle Physics, Biotechnology, AI Alignment Research etc.
These collaborations ensure that progress is measured not just in theory, but in real-world frontier outcomes.
What comes next
This is just the beginning. SIA is the first of several open source releases. Over the coming weeks, we will introduce additional components that expand the capabilities of SIA, including systems for evaluation, orchestration, and domain-specific research. Each release strengthens the loop, and each iteration increases the system’s ability to learn and improve. Till Infinity.
Work on humanity’s last invention
Hexo Labs is building the machine that builds every future machine. The machine that turns the wheels of evolution, creates a future of abundance for all and accelerates evolutionary progress.