researchAi2

Ai2 launches Theorizer in Asta and open-sources the AI agent that proposes scientific theories

TL;DR

Ai2 launched Theorizer, an AI agent that combines findings from scientific papers to propose explanations for observations, inside its Asta platform. The code is open-source, and Ai2 covers token costs up to a limit in the web interface. A lung cancer case study shows the workflow.

4 min read
0

Ai2 launched Theorizer in Asta, its platform of AI agents for scientific research, on November 19, 2026, and open-sourced the code. Theorizer is an AI agent that combines findings from scientific papers to propose explanations for observations, such as a tumor's unexplained growth. The release is aimed at researchers who want to examine how it generates theories and adapt the methods to their own work.

Access and cost

There are two ways to use Theorizer:

  • Asta web interface: Ai2 covers token costs up to a limit.
  • Local runs: researchers can run the open library on their own machines and pay for tokens as needed.

Ai2 says scientists have spent nearly a year helping develop Theorizer by testing its theories in their own research. Ai2 will also host a webinar for scientists on Thursday, December 10 at 10am PT, covering how Theorizer connects findings into theories, a real cancer research example, and what it will take for AI to reliably contribute to the scientific process.

How Theorizer works

Theorizer searches the research literature in response to a user's question, then creates a template specifying what to extract from each paper. For a question about how lung cancers grow, that might include which genes a study examined, how researchers measured their activity, and what happened when the researchers altered those genes. It then compares the extracted findings to generate theories and links each claim to the papers behind it. Each proposed explanation also comes with a description of where it might apply.

Researchers can inspect the cited studies to judge how well a theory is supported and whether the studies examined comparable conditions. Theorizer also flags claims it considers new. Ai2 cautions that a search can miss earlier work and the model can misjudge originality, so researchers still need to check a claim against the wider literature.

Case study: driver-negative lung adenocarcinoma

Dr. Sid Devarakonda, medical director of the thoracic oncology clinical and research program at Providence Swedish Cancer Institute, tested Theorizer in collaboration with the Asta team. He studied "driver-negative" lung adenocarcinoma, the most common subtype of lung cancer. Most lung adenocarcinomas carry a genetic alteration that activates the RAS pathway, a chain of signals that promotes cell growth. In roughly 20% to 30% of cases, even comprehensive sequencing does not reveal a recognized driver alteration.

Devarakonda asked Theorizer, "What's happening in these tumors?" After several iterations it helped him formulate three hypotheses:

  1. An undetected genetic change is activating RAS signaling.
  2. A different mechanism allows cells to divide without relying on RAS.
  3. Some tumors have acquired characteristics of another cell type and no longer behave like typical lung adenocarcinomas.

Devarakonda and the Asta team then tested these theories with AutoDiscovery, which writes and runs code to test hypotheses against datasets. He used The Cancer Genome Atlas, a public collection of molecular data from tumors, to compare genetic changes, gene activity, and protein measurements in cancers with and without recognized drivers. He also examined data from laboratory-grown cancer cells to identify genes needed for growth or survival.

AutoDiscovery suggested that disruption of the DREAM complex, a group of proteins that normally keeps cell division in check, might help explain the tumors' growth. Devarakonda had not initially focused on this mechanism. He compared RAS activity with how strongly genes normally controlled by the complex were switched on, and follow-up work with data from laboratory-grown cancer cells gave him preliminary support for the hypothesis.

Of his first set of Theorizer theories, Devarakonda recognized ideas he had developed through years of research. He recalled thinking, "Wait a second, this is exactly what I think. And it took me 15 years to think like this."

What this means

Theorizer's design puts verification in the researcher's hands: every theory links back to published studies that can be inspected, and Ai2 states plainly that its novelty flags can be wrong. The open-source release lets researchers examine and adapt the extraction-and-comparison method rather than treat it as a closed tool. In the lung cancer example, the agent's output was a set of specific hypotheses that could be tested in molecular data. The testing was done with a separate agent, AutoDiscovery, and the DREAM complex result is described as preliminary support, not a confirmed finding. Model names, benchmark scores and context window figures were not part of this announcement.

Related Articles

research

Google builds isolated 'Mirror Dimension' web to train and test Chrome AI agents

Google's Chrome team built the Mirror Dimension, an isolated web ecosystem it fully controls, to train and test Gemini-based browsing agents without exposing them to the open internet. Google is detailing the project in a blog post, according to Wired.

research

Hugging Face releases Qwen3.8-27B-pi-adapted LoRA, lifting SWE-bench Pro hard-51 pass@1 from 39.2% to 64.7%

Hugging Face released Qwen3.8-27B-pi-adapted, an Apache-2.0 LoRA adapter trained with reinforcement learning for the pi coding agent. According to MarkTechPost, it raises SWE-bench Pro hard-51 pass@1 from 39.2% to 64.7% and Terminal-Bench 2.1 from 39.5% to 55.8%.

research

Microsoft Research's TokenEconomizer plugin cuts coding agent costs 18.4% to 30.9% on SWE-Bench

TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research and Microsoft Security, learns from an agent's past sessions in a repository to cut token costs. Across 277 SWE-Bench tasks and five agent and model combinations, mean cost per task fell 18.4% to 30.9%, according to the authors.

research

ACE Robotics' ACE-CaP-0 lets AI agents write reusable robot code, reporting 92.1% success

ACE Robotics has released ACE-CaP-0, a research recipe in which a general-purpose AI agent writes, tests and refines robot control programs, then freezes them into reusable skills that run without the agent. The work reports 92.1% success. The paper was posted to arXiv on November 12.

Comments

Loading...