Agentic-Ideation: Sample Efficient Agentic Trajectories Synthesis for Scientific Ideation Agents
TLDR
Proposes Agentic-Ideation, a framework for efficient trajectory synthesis to train agentic LLMs for scientific ideation using oracle-guided multi-agent generation.
Reasoning
The paper addresses the high cost of data synthesis for training agentic LLMs in scientific ideation, introducing a novel oracle-guided strategy. However, the abstract lacks empirical results or real-world validation, and the scope is limited to ideation rather than broader scientific discovery tasks.
Read-first score
Read-first score 52.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 56.
Field roles
Rank sensitivity
Stability: volatile; rank range: 42.
Keyword Scores
Deep Analysis
Innovations
- Oracle-Guided Data Synthesis strategy using reference ideas to steer multi-agent trajectory generation
- Masking strategy on tool execution results during training to focus on decision-making logic
- Comprehensive tool space incorporating external and cognitive tools for scientific ideation
- Agentic-Ideation framework combining automated trajectory synthesis and specialized agentic LLM training
Methodology
The framework defines a tool space of three external and three cognitive tools, then uses an Oracle-Guided Data Synthesis strategy where a reference idea steers a multi-agent system to reconstruct logical reasoning and tool invocation paths, generating directed trajectories. An agentic LLM is trained on these trajectories with a masking strategy on tool execution results to focus on decision-making logic.
Key Results
The method outperforms the state-of-the-art workflow-based baseline by 11.91% in overall quality and improves sample efficiency of high-quality data synthesis by over 10x.