AIM: Agentic Idea Management for Automated Research
Hyeong Kyu Choi, Bhavana Dalvi Mishra, Jiefeng Chen, Mihir Parmar, Rui Meng, Chun-Liang Li, Xiangru Tang, Sharon Li and 2 others
Google Cloud AI Research · University of Wisconsin-Madison
Published September 29, 2026 · on the feed October 1, 2026 · ▲ 50 · arXiv 2609.38445
The abstract
Frontier LLMs are increasingly used to automate scientific research through iterative search. We distinguish idea-driven search from solution-driven search and identify three core challenges: organizing evolving research ideas, selecting promising directions, and maintaining alignment between ideas and their implementations. To address these challenges, we introduce the Agentic Idea Manager (AIM), a fully autonomous framework for managing and exploring research directions in idea-driven automated research. Inspired by Bayesian optimization, AIM uses an Agentic Surrogate and an Agentic Acquisition mechanism to organize discovered ideas and guide their selection. A Solution Auditor maintains idea-solution integrity, while a Resource Planner adaptively allocates the remaining experimental budget across parallel search branches. Experiments on 10 AutoLab benchmark tasks show that AIM surpasses the strongest baseline by 1.6 percentage points on System Optimization tasks and 4.9 percentage points on long-horizon Model Development & CUDA tasks. Notably, AIM reaches the best baseline performance up to 3.1x faster in wall-clock time. We further provide a theoretical analysis of when searching over ideas becomes beneficial. Our analysis shows that explicit idea-level allocation makes semantic coverage directly controllable, and that broader coverage becomes increasingly valuable when competitive research directions are sparse among many plausible alternatives. Project Page: https://imhgchoi.github.io/agentic-idea-manager/
The Brief
2:26A short podcast summary of the paper. Written and voiced by AI.
Transcript
AIM. Agentic Idea Management for Automated Research. By Hyeong Kyu Choi and nine others, from Google Cloud AI Research and the University of Wisconsin–Madison. Published on the twenty-ninth of September. No code repository is linked to it.
Frontier language models are increasingly used to automate scientific research, by searching iteratively for something that works. The authors draw a line between two kinds of search, one driven by solutions and one driven by ideas, and the work is about the second. Idea-driven search brings three challenges of its own: organizing ideas that keep evolving, choosing which directions look promising, and keeping each idea aligned with the implementation that is supposed to embody it.
The Agentic Idea Manager, AIM, is a fully autonomous framework built to meet those three challenges, managing and exploring research directions in idea-driven automated research. Its organizing inspiration is Bayesian optimization, and the names of its two central parts, a surrogate and an acquisition mechanism, come from that tradition, here rebuilt as agents.
An Agentic Surrogate and an Agentic Acquisition mechanism together organize the ideas discovered so far and guide which ones are selected. Two further components handle the other challenges. A Solution Auditor maintains the integrity between an idea and its solution, so that what gets built matches what was proposed. And a Resource Planner allocates the remaining experimental budget adaptively, across search branches running in parallel.
The experiments use ten tasks from the AutoLab benchmark. On System Optimization tasks, AIM beats the strongest baseline by one point six percentage points. On the long-horizon Model Development and CUDA tasks the margin is four point nine points. AIM also reaches the best baseline's performance up to three point one times faster in wall-clock time.
Alongside the experiments comes a theoretical analysis of when searching over ideas pays off at all. Its conclusions are two. Allocating effort explicitly at the level of ideas makes semantic coverage directly controllable. And broader coverage becomes increasingly valuable when competitive research directions are sparse, hidden among many plausible alternatives.
The case for AIM rests on treating ideas, not solutions, as the unit being searched, and then borrowing the structure of Bayesian optimization to manage them. The analysis says when that choice matters most: when the good directions are rare and the plausible ones are many, controlling how widely ideas are covered is exactly what makes a difference.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
An agent can now run its own experiments. Keeping its ideas in order is the hard part.
You are running a language model agent that does research on its own: it proposes approaches, implements them, runs experiments and reads the results. Every experiment costs time from a fixed budget.
The road ahead
- 1 · The problem ~7 min A research agent that searches over ideas must keep a growing pile of ideas organized, spend a fixed budget well, and check that results really test each idea.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Treat idea search like Bayesian optimization, with one agent estimating which ideas look promising and another choosing what to try, plus an auditor and a budget planner.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Each iteration organizes ideas, selects some, implements them in parallel, audits the results, then reallocates the remaining budget before the next round.
- 6 · The experiment ~6 min You’ll guess: In wall-clock time, how much faster did AIM reach the best baseline's performance?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.