AI NewsModels & agentsAnnouncement
Ai2 released Theorizer, an agent that drafts scientific theories from papers
Ai2 released Theorizer, an AI agent that reads research papers and drafts scientific theories, with links back to its sources.
Image: Ai2 on GitHub
Why it mattersA researcher can ask one question and get candidate theories, each tied to the papers that support it. The code is on GitHub for teams that want to run it.
A scientist with a new result can spend weeks reading papers before knowing whether anybody has already explained it. Ai2 released Theorizer, an AI agent that does that first pass automatically. It reads the literature, pulls out findings, and drafts candidate theories tied back to each paper.
Theorizer runs inside Asta, Ai2's platform of AI agents for scientific research, at asta.allen.ai. The code is open source at github.com/allenai/asta-theorizer under Apache 2.0, with 180 stars when we checked. Ai2 is the Allen Institute for Artificial Intelligence, the non-profit lab founded by Paul Allen.
What Theorizer does
Ai2 states that Theorizer searches research papers in response to a user's question. It then builds a template. The template says what to pull from each paper. The agent reads the papers and extracts those findings. It compares the findings and drafts candidate theories. Each theory links back to the papers that support it. The agent flags any claim it believes is new.
The tool pairs with AutoDiscovery, another Asta agent. AutoDiscovery can take a theory from Theorizer and test it against a dataset. Together, the two agents are meant to take a researcher from a broad question to a set of testable hypotheses.
How to use it
Ai2 lists three ways to run Theorizer. The first is the Asta web interface, where Ai2 covers the token cost up to a limit. The second is the Python library asta-theorizer, which installs from PyPI and needs Python 3.12 or later. The third is a standalone server for teams that want to host it themselves.
The library mode and the server call paid models with API keys the user supplies. The README warns to start small and set hard limits on the key, because the cost is unpredictable.
What Ai2 says it misses
Ai2 lists two real limits in the post. Theorizer can miss relevant studies. And it can call a theory new when the theory already exists elsewhere. The agent compares only the papers it read. A researcher still has to check the output against the broader field and run the experiments.
Ai2 released the tool with two evaluation modes. One favours accuracy, which cuts false claims of novelty. The other favours novelty, for exploratory work where speed of ideas matters more. The paper that describes the method is cited in the blog post.
Theorizer is the second Asta tool Ai2 has opened in the last two months. The first was Astabrief 8B, a small model for writing scientific summaries. The platform now covers literature search, summaries, theory writing and hypothesis testing in one web interface.
Source
Introducing Theorizer: an AI tool that generates scientific theories, Ai2. The code is at allenai/asta-theorizer.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
