🤖 AI

Agentic-Ideation: Sample Efficient Agentic Trajectories Synthesis for Scientific Ideation Agents

The paper introduces Agentic-Ideation, a novel framework that leverages an Oracle-Guided Data Synthesis strategy to efficiently generate high-quality training trajectories for scientific ideation agents, thereby overcoming the high costs of previous methods and achieving a significant improvement in both data synthesis efficiency and overall ideation quality compared to state-of-the-art baselines.

Original authors: Keyu Zhao, Lingyan Kong, Fengli Xu, Yong Li

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Keyu Zhao, Lingyan Kong, Fengli Xu, Yong Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Teaching a Robot to Be a "Scientist"

Imagine you want to build a robot that can come up with brilliant new scientific ideas on its own. This is the dream of an "AI Scientist."

Currently, most attempts to build this robot are like giving it a strict, pre-written recipe. The robot has to follow steps 1, 2, and 3 exactly as written by a human. If the recipe says "look up a book," it looks up a book. If it says "write a sentence," it writes a sentence. The problem is that real science isn't a straight line; it's a messy, winding path where you have to change direction, try new tools, and rethink your strategy constantly. The rigid recipes are too stiff to handle this.

This paper proposes a new way: instead of giving the robot a recipe, they train the robot to think and act like a flexible, independent researcher.

The Problem: The "Needle in a Haystack" Issue

To teach a robot to be flexible, you need to show it examples of good thinking (called "trajectories"). You might think, "Let's just let the robot try to come up with ideas, and if it gets it right, we save that example."

But in science, there is no single "right answer" like in a math quiz. There are millions of possible ideas, and most are bad. If you let the robot guess randomly, it might try 12 times just to get one good idea. This is like trying to find a specific needle in a haystack by blindly grabbing handfuls of hay. It takes forever and wastes a massive amount of computer power.

The Solution: The "Oracle" (The Magic Compass)

The authors invented a clever trick called Oracle-Guided Synthesis.

Imagine you are trying to teach a student how to solve a mystery.

The Old Way (Rejection Sampling): You let the student wander around the house randomly. They might knock over a vase, open the wrong drawer, and get lost. You only keep the times they accidentally find the hidden clue. This is slow and frustrating.
The New Way (Oracle-Guided): You give the student a magic compass (the "Oracle") that points directly to the hidden clue. The student still has to walk the path and figure out how to get there, but the compass ensures they don't wander off into the woods.

In the paper, the "Oracle" is a perfect, pre-existing scientific idea (the "Reference Idea"). The system uses this idea as a guide to help the robot reconstruct the logical path of how to get there.

The robot sees the destination (the Reference Idea).
It figures out the steps: "I need to search for this," "I need to find a gap in the research," "I need to reflect on my idea."
It writes down the path in one go, without wasting time on dead ends.

This makes the data generation process 10 times faster than the old random guessing method.

The Robot's Toolkit: "Thinking" and "Doing"

To make the robot smart, the authors gave it a specific set of tools, split into two categories:

External Tools (The "Eyes and Ears"):
- Search: Looks up papers in a database.
- Get_References: Checks what a paper cited (its roots).
- Get_Cited: Checks who cited the paper (its future impact).
- Analogy: These are like the robot's library card and internet connection.
Cognitive Tools (The "Brain"):
- Analyse_Gap: Looks at what it knows and asks, "What is missing?"
- Ideation: Comes up with a new idea based on that missing piece.
- Reflection: Acts as a strict editor. It asks, "Is this idea actually new? Or is it just a copy of something old?" If it's a copy, the robot throws it away and tries again.
- Analogy: These are the robot's critical thinking skills.

The Training: Hiding the Answers

When they train the robot, they use a special technique. They show the robot the steps it took (the search, the gap analysis, the idea), but they hide the actual results of the searches during the learning process.

Why? If the robot sees the answer immediately, it might just memorize the answer instead of learning how to find it.
The Result: By hiding the results, the robot is forced to learn the logic of decision-making. It learns why to search for something, not just what it found. This makes the robot robust and able to handle new, unseen problems.

The Results: A Smarter, Faster Scientist

The authors tested their new robot against the best existing "recipe-following" robots and a standard AI model.

Quality: The new robot came up with ideas that were 11.9% better overall. Experts rated its ideas as more novel (newer), more significant (more important), and more feasible (actually possible to do).
Efficiency: As mentioned, it took 10 times less effort to create the training data.
Real-World Test: In a specific example, the robot was asked to find a problem in "Diffusion Models" (a type of AI image generator).
- It searched for weaknesses.
- It found a gap: current models don't fix errors while they are working.
- It proposed a new idea to fix this.
- Crucially: It used its "Reflection" tool to realize its first draft was too similar to an existing method, rejected it, and refined it into a truly new, high-quality idea.

Summary

This paper introduces a system that teaches AI to be a flexible scientific researcher rather than a rigid script-follower. It solves the problem of "how do we teach AI to be creative?" by using a magic compass (Oracle) to guide the training process efficiently, ensuring the AI learns the logic of discovery rather than just memorizing answers. The result is an AI that generates better, more innovative scientific ideas much faster than previous methods.

Technical Summary: Agentic-Ideation

1. Problem Statement

Scientific ideation is the foundational phase of the research lifecycle, determining the novelty and impact of downstream discoveries. While Large Language Models (LLMs) have shown potential for automating this process, existing approaches face two critical limitations:

Rigidity of Workflow-Based Systems: Current "AI Scientist" systems predominantly rely on pre-defined, linear workflows (e.g., literature retrieval followed by template filling). These static pipelines, driven by complex human-designed prompts, lack the flexibility to adaptively navigate the vast search space of scientific literature or the complex action space of research reasoning.
Inefficiency in Agentic Data Synthesis: Training flexible Agentic LLMs requires high-quality trajectory data. However, standard data synthesis methods (like rejection sampling) used for tasks with definite answers (e.g., QA) fail in scientific ideation. Because ideation involves an open-ended answer space with no single "correct" ground truth, unguided agents struggle to converge on high-quality ideas, leading to prohibitively high computational costs and low sample efficiency.

2. Methodology

The authors propose Agentic-Ideation, a framework designed to synthesize sample-efficient agentic trajectories and train a specialized LLM for scientific ideation. The framework consists of three integrated stages:

A. Agentic Tool Space Definition

The authors define a hybrid tool space ( $T = T_{ext} \cup T_{think}$ ) comprising:

External Tools ( $T_{ext}$ ):
- Search(q): Queries an academic database for comprehensive metadata (abstracts, citations).
- Get_References(did): Retrieves a bibliography (titles only) for a target paper.
- Get_Cited(did): Identifies papers citing a target paper (titles only).
Thinking Tools ( $T_{think}$ ):
- Analyse_Gap(H): Synthesizes retrieved information to identify logical inconsistencies or underexplored intersections ("research voids").
- Ideation(H, ogap): Generates a novel scientific idea based on identified gaps.
- Reflection(H, oidea): Critically evaluates the generated idea to trigger refinement if necessary.

B. Oracle-Guided Agentic Data Synthesis

To overcome data scarcity, the authors introduce a hierarchical multi-agent system guided by an "Oracle":

Architecture: The system includes a Planner (generates a high-level roadmap), a Controller (executes tool selection), and specialized tool agents.
Oracle Guidance: Instead of relying on blind trial-and-error, the system uses a "Reference Idea" ( $I_{ref}$ ), derived from an "Anchor Paper," as a navigational beacon.
Dual-Thought Mechanism: The Controller generates two cognitive streams simultaneously:
1. Private Thought ( $th_{priv}$ ): Uses the Reference Idea to calculate the optimal tool path, effectively collapsing the search space with "foresight."
2. Public Thought ( $th_{pub}$ ): Rationalizes decisions using only the available history (masking the Reference Idea), mimicking natural autonomous exploration.
Synthesis Process: The system reconstructs a valid, logically coherent reasoning path in a single pass by leveraging the Oracle's guidance to bridge the gap between the current state and the target idea. Trajectories are only preserved if the final generated idea semantically aligns with the Reference Idea.

C. Agentic Supervised Fine-Tuning (SFT)

The synthesized trajectories are used to train the target LLM (Qwen3-8B):

Linearization: Trajectories are converted into a structured format with special tokens for tool calls, results, and thoughts.
Masking Strategy: During loss calculation, the results of external tool executions (tokens within <Result>...</Result>) are masked. This forces the model to focus on the decision-making logic and reasoning behind tool invocation rather than memorizing specific return values, ensuring robustness and generalization.

3. Key Contributions

Paradigm Shift to Agentic Ideation: The work pioneers the application of agentic LLMs to scientific ideation, moving away from rigid, workflow-based approaches toward flexible, autonomous reasoning frameworks.
Oracle-Guided Synthesis Framework: A novel multi-agent data synthesis strategy that utilizes a global view (Reference Idea) to guide trajectory generation. This addresses data scarcity in open-ended domains and improves sample efficiency by over 10× compared to rejection sampling.
State-of-the-Art Performance: The trained agent achieves superior results across all evaluation metrics, demonstrating that the proposed training paradigm effectively enhances the model's ability to generate novel, significant, and feasible scientific ideas.

4. Experimental Results

The framework was evaluated on a dataset of accepted papers from ICLR 2025, ICML 2025, and NeurIPS 2025, with a test set of 300 papers.

Performance: Agentic-Ideation outperformed all baselines (including ResearchAgent, SciPIP, SciMON, VirSci, and the backbone Qwen3-8B).
- Overall Score: Achieved a 6.67 score, an 11.91% improvement over the best baseline.
- Metric Breakdown: Significant gains were observed in Novelty (+11.40%), Significance (+11.14%), and Feasibility (+8.70%).
- Human Evaluation: Blind evaluation by human experts confirmed the automated results, with Agentic-Ideation scoring highest in Overall quality (6.45) and Significance (+12.28% over the best baseline).
Efficiency: The Oracle-Guided strategy reduced the average number of rollout attempts required to synthesize a single valid sample from 12 (in rejection sampling) to 1, representing a >10× improvement in synthesis efficiency.
Ablation Studies: Removing any single tool (Search, Get_References, Get_Cited, Analyse_Gap, Reflection) resulted in consistent performance degradation, confirming the necessity of the full hybrid tool space. Specifically, Search was critical for feasibility, and Analyse_Gap was critical for novelty.

5. Significance and Claims

The paper claims that Agentic-Ideation addresses the critical bottleneck of data scarcity in training agentic systems for open-ended scientific tasks. By transforming aimless trial-and-error into directed trajectory reconstruction, the framework makes large-scale agentic data synthesis computationally feasible.

The authors assert that their method enables the development of autonomous AI scientists capable of grounded reasoning and reliable innovation. By training on synthesized trajectories where the model learns to make decisions based on logic rather than memorized feedback, the system produces ideas that are not only creative but also technically sound and scientifically valuable. This advancement is positioned as a step toward democratizing high-level ideation capabilities and accelerating research across diverse scientific disciplines through effective human-AI collaboration.