GAI: Generative Agents for Innovation

This study examines whether collective reasoning among generative agents can facilitate novel and coherent thinking that leads to innovation. To achieve this, it proposes GAI, a new LLM-empowered framework designed for reflection and interaction among multiple generative agents to replicate the process of innovation. The core of the GAI framework lies in an architecture that dynamically processes the internal states of agents and a dialogue scheme specifically tailored to facilitate analogy-driven innovation. The framework’s functionality is evaluated using Dyson’s invention of the bladeless fan as a case study, assessing the extent to which the core ideas of the innovation can be replicated through a set of fictional technical documents. The experimental results demonstrate that models with internal states significantly outperformed those without, achieving higher average scores and lower variance. Notably, the model with five heterogeneous agents equipped with internal states successfully replicated the key ideas underlying the Dyson’s invention. This indicates that the internal state enables agents to refine their ideas, resulting in the construction and sharing of more coherent and comprehensive concepts.

💡 Research Summary

The paper introduces GAI (Generative Agents for Innovation), a novel framework that leverages multiple large‑language‑model (LLM) agents equipped with dynamic internal states to simulate the collective reasoning processes underlying human innovation. The authors argue that traditional observational studies of corporate R&D suffer from low reproducibility and limited experimental control, and that an LLM‑based multi‑agent system (LLM‑MAS) could provide a reproducible, manipulable platform for studying how team diversity, communication structures, and intrinsic motivations affect innovative outcomes.

GAI’s architecture consists of two core modules per agent: a memory module and an internal‑state module. The memory module stores timestamped records of technical documents, past dialogues, and the agent’s own reflections, organizing them into a structured format that feeds the internal‑state module. The internal‑state module performs a continuous loop of idea generation and introspection. During idea generation, the agent produces several candidate ideas, each scored on three quantitative criteria—Novelty, Importance, and Consensus—on a 1‑10 scale. An intrinsic reward function combines these scores using agent‑specific weights (α, β, γ), allowing the system to model heterogeneous motivations (e.g., novelty‑seeking vs. consensus‑seeking agents). The highest‑reward idea is then passed to a second LLM acting as an internal critic, which probes for ambiguities, contradictions, or technical gaps. The feedback is incorporated to refine the idea, updating the agent’s “current thoughts” that serve both as context for future reflections and as the basis for dialogue contributions.

The dialogue scheme is explicitly designed for Design‑by‑Analogy (DbA). Agents work on a target domain (household fans) and a source domain (industrial ejectors) and follow a structured template: (1) extract functional similarities (both move fluids), (2) identify mechanical differences, and (3) propose a transfer of solutions. The organizational structure of the agent team is represented as a directed graph, enabling experiments with flat, hierarchical, or mixed communication patterns.

Two experiments evaluate the framework. Experiment 1 uses fictional patent‑style documents describing Dyson’s first‑generation Air Multiplier (the “bladeless” fan) and industrial ejectors. Agents must, without any pre‑trained knowledge of Dyson products, reinterpret the concept of a “driving fluid” for a household fan and propose a feasible mechanism. Nine evaluation criteria (e.g., presence of a nozzle, motor type, fluid‑dynamics explanation) are scored up to eight points. Eight model variants are compared across three dimensions: number of agents (single vs. five), presence of internal state (yes/no), and heterogeneity of intrinsic motivations (homogeneous vs. heterogeneous). Results show that models lacking internal states perform poorly (average ≈ 3/8 points) with high variance, whereas the five‑agent heterogeneous configuration with internal states achieves an average of 7.2/8 points, correctly reproducing key elements such as a nozzle structure, brushless motor, and fluid‑dynamic principles. Experiment 2 extends the evaluation to ten additional source domains, confirming that the framework generalizes beyond the Dyson case.

Key contributions include: (1) a dynamic internal‑state architecture that enables persistent, self‑evolving reasoning across dialogue turns; (2) a DbA‑oriented dialogue protocol that operationalizes functional and relational analogy extraction from technical texts; (3) a flexible graph‑based organizational model for studying the impact of team structure on innovation outcomes.

Limitations are acknowledged: reliance on synthetic documents, evaluation confined to textual scoring without physical simulation, and manually set weighting parameters for novelty, importance, and consensus. Future work is proposed to integrate real patent corpora, couple the system with physics‑based simulators (e.g., CFD) for design validation, and employ meta‑learning to automatically discover optimal motivation weights and communication topologies.

Overall, the study demonstrates that LLM‑driven multi‑agent systems with internal reflective states can meaningfully emulate analogy‑driven innovation processes, opening avenues for both scientific investigation of collaborative creativity and practical AI‑human co‑design in industrial settings.