IDEAgent: Agentic Quality-Diversity Search for Research Idea Generation

Varun Gumma, Navonil Majumder, Soumitra Sinhahajari, Soujanya Poria

DeCLaRe Lab, Nanyang Technological University, Singapore

Abstract

Large Language Models (LLMs) have significantly advanced the automation of scientific discovery. However, existing systems share a core limitation: they optimize ideas independently for either Quality or Diversity. This often results in closely related ideas or large sets of trivial, unsound, or unclear concepts. We instead argue that research ideation should be framed as a conjunction of Quality and Diversity objectives, i.e., a Quality-Diversity (QD) search. Accordingly, we introduce IDEAgent, a multi-agent framework that tracks idea evolution through lineages. We jointly drive Quality using multi-objective feedback for repair and refinement, while Diversity is achieved through lightweight sequential memory and explicit comparison against completed ideas, historical variants, and rejected proposals. To evaluate this joint objective, we develop Yield, a metric that computes the largest mutually diverse set of ideas satisfying predefined quality thresholds. Across 32 topics spanning 8 CS domains, IDEAgent outperforms the best baseline by 3.89× on Yield, while achieving Yield of at least 2 on 8× as many topics. We further analyze quality improvements, showing that repair and refinement improve logical rigor and clarity while preserving non-obviousness. Lastly, we extend IDEAgent to two algorithm discovery tasks for verifiable quality judgment and find that it still consistently outperforms all baselines, and yields the largest set of high-quality and diverse programs that beat a trivial/standard baseline. To encourage future research on QD-search-based ideation, we open-source IDEAgent at https://github.com/declare-lab/IDEAgent.

Method at a Glance

Sequential generation with memory: each seed is repaired and refined before the next one starts.

Multi-agent evaluation — Diversity, quality and soundness judges decide whether an idea is rejected, repaired or accepted.

Quality-Diversity archive — Accepted ideas enter a quality-diversity archive of active, historical and rejected ideas.

From Ideas to Verifiable Programs

IDEAgent adapted to verifiable algorithm discovery: task definition, proposal plus program, sandboxed execution, a single fatal soundness check, and the unchanged diversity judge, critic and archive

Background papers → the task definition and its runtime limits, plus a trivial baseline that every idea must beat.
An idea → a proposal together with the Python / C++ program that implements it.
Non-obviousness and clarity judges → removed. Quality is measured: the program runs in a sandbox (60 s during search on AlphaResearch, the contest time limit on ALE-Bench), and Q_i is its normalized progress from the trivial solution (0) to the best-known result (100).
Soundness panel → one fatal check: the program implements the proposal, with no shortcuts or hard-coded answers.
Diversity judge, archive and critic → unchanged; distinctness now lies in the central mechanism. A valid idea that does not beat the baseline gets one repair; one that does gets up to two refinements, and a higher-Q child replaces its parent.
Models → open models: Qwen3.8-27B as the Ideator; Qwen3.6-35B-A3B as the critic, soundness and diversity judges.

Results: Research Ideation (32 topics)

Method	Yield (NB≥7)	Yield (NB≥6)	Topics with ≥1 qualifying idea (NB≥7)	Topics with ≥1 qualifying idea (NB≥6)
Single-shot	0.000	0.000	0 / 32	0 / 32
Sequential-Memory	0.188	0.531	6 / 32	14 / 32
Stateless	0.250	0.812	7 / 32	24 / 32
NOVA	0.281	1.156	8 / 32	25 / 32
IDEAgent	1.094	2.312	27 / 32	31 / 32

Jury average of two external judges; Yield at D≥7, S≥7, C≥6 with the stated non-obviousness (NB) threshold.

Results: Verifiable Algorithm Discovery

AlphaResearch (Python; Q = 100 is the human best)

Method	Q_max	Q_avg	Y₀	Y₅₀	Y₇₅	Y₁₀₀
Single-shot	53.2	16.9	2.78	1.00	0.44	0.33
Sequential-Memory	92.6	46.1	4.00	3.11	1.67	0.56
Stateless	93.5	47.8	4.11	3.11	1.78	0.56
NOVA	75.4	34.4	4.56	2.67	0.78	0.22
IDEAgent	106.2	54.5	5.67	4.44	3.11	0.67

ALE-Bench (C++ heuristic contests; Q = 100 is the contest winner)

Method	Q_max	Q_avg	Y₀	Y₂₅	Y₅₀	Y₇₅
Single-shot	14.8	6.2	1.14	0.43	0.43	0.00
Sequential-Memory	38.3	18.6	2.29	1.71	1.14	0.29
Stateless	45.1	22.7	2.29	1.71	0.57	0.00
NOVA	37.3	18.0	2.29	2.00	1.00	0.00
IDEAgent	56.4	22.6	3.43	2.57	1.57	0.57

Y_k: size of the largest set of mutually distinct ideas among those that clear the trivial solution with progress Q_i ≥ k (Y₀ counts every idea that clears it). Q_max and Q_avg: the highest and mean Q_i among a problem's ten final ideas. All averaged over a benchmark's problems; a pair is distinct when a majority of three judges (Gemma 4 31B, MiniMax-M2.7, GLM-4.5-Air) say so. Bold marks the best value in each column.

BibTeX

@misc{gumma2026ideagentagenticqualitydiversitysearch,
      title={IDEAgent: Agentic Quality-Diversity Search for Research Idea Generation},
      author={Varun Gumma and Navonil Majumder and Soumitra Sinhahajari and Soujanya Poria},
      year={2026},
      eprint={2607.22375},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2607.22375},
}