AI for autonomous science fails in long experiments: the bottleneck of autonomy of ultralong horizon

Large language models (LLMs) — the backbone of systems like ChatGPT — still cannot act as reliable autonomous scientists when experiments extend for long periods. An article featured in Hugging Face points out that the autonomy of ultralong horizon is the main bottleneck, since LLMs lose strategic coherence over prolonged experimental cycles, compromising automated scientific research.

What is ultralong horizon autonomy in AI?

Ultralong horizon autonomy refers to the ability of an AI system to perform tasks that require cohesive planning and execution for long periods — hours, days or even weeks. In science, this means that an AI would need to formulate hypotheses, design experiments, collect data, adjust variables and iterate without human intervention. The [Hugging Face] (https://huggingface.co/papers?q=ultra-long-horizon+autonomy) article highlights that current LLMs show "strategic bias", i.e. gradually deviate from the initial objectives.

Why does autonomous AI fail in long experiments?

The main difficulty is in maintaining coherence. Language models process information in local contexts — one piece of text at a time — and do not have a robust long-term memory mechanism to retain initial strategic reasoning. In experiments that require dozens of steps, the model tends to get lost, repeat actions or ignore intermediate results. This severely limits application in areas such as chemistry, biology and materials science, where long experimental cycles are the norm.

What is the impact of the autonomy bottleneck on automated scientific research?

Independent AI scientists — such as systems that try to discover new drugs or catalysts — depend on the ability to adapt strategies over time. Without autonomy of ultralong horizon, these tools produce inconsistent results or require constant human supervision, nullifying the benefit of automation. The article suggests that as long as there are no advances in long-term memory and hierarchical planning, autonomous science will remain limited to short-term problems. For Brazilian developers who work with AI agents, understanding this limitation is crucial — as we discussed in Anthropic's MCP: how to create tools that AI agents really use, agent architecture needs to evolve to support more complex tasks.

Frequently Asked Questions about autonomous AI and long experiments

How does ultralong horizon autonomy differ from short AI tasks?

While short tasks (such as answering questions) require local cohesion, long autonomy demands strategic planning and persistent memory, something that current LLMs do not dominate.

Which areas of science would benefit most if this bottleneck is resolved?

Areas such as drug discovery, chemical synthesis and materials science, which depend on long sequential experiments, would be the most positively impacted.

Are there promising approaches to improving the strategic coherence of LLMs?

Yes, techniques such as long-term memory based on vector banks, hierarchical planning and feedback loops are explored, but still do not solve the problem completely for very long horizons.