Multi-Novelty: Improve the Diversity and Novelty of Contents Generated by Large Language Models via inference-time Multi-Views Brainstorming

TL;DR

This work tackles the challenge that large language models often produce less diverse and novel content due to training-data limitations. It introduces inference-time multi-view brainstorming, an architecture-agnostic approach that augments prompts with diverse textual and visual views to boost diversity and novelty in generated outputs. A comprehensive evaluation framework assesses diversity, novelty, and correctness across multiple models on a large-scale 909k-response dataset, revealing significant gains in diversity and novelty with some trade-offs in correctness. The method enables richer, more exploratory AI applications without requiring architectural changes, with potential impact on creative AI agents and AI-driven scientific exploration.

Abstract

Large Language Models (LLMs) demonstrate remarkable proficiency in generating accurate and fluent text. However, they often struggle with diversity and novelty, leading to repetitive or overly deterministic responses. These limitations stem from constraints in training data, including gaps in specific knowledge domains, outdated information, and an over-reliance on textual sources. Such shortcomings reduce their effectiveness in tasks requiring creativity, multi-perspective reasoning, and exploratory thinking, such as LLM based AI scientist agents and creative artist agents . To address this challenge, we introduce inference-time multi-view brainstorming method, a novel approach that enriches input prompts with diverse perspectives derived from both textual and visual sources, which we refere to as "Multi-Novelty". By incorporating additional contextual information as diverse starting point for chain of thoughts, this method enhances the variety and creativity of generated outputs. Importantly, our approach is model-agnostic, requiring no architectural modifications and being compatible with both open-source and proprietary LLMs.

Figures (8)

• Figure 1: Overview of our proposed methodology for enriching the diversity and novelty of LLM-generated content using multi-view embeddings. Our approach starts with diverse input prompts (e.g., philosophical questions, problem-solving scenarios, imaginative tasks) and generates multiple textual and visual views (n=50 per each prompt) through multi-view generators. These multi-view embeddings are then fed into various open-source and closed-source LLMs, such as Qwen, DeepSeek-R1, and GPT-4o, to generate responses with enriched diversity and novelty. We created the 909kPR dataset consisting of 909,500 generated answers across different models. Finally, the DNC framework evaluates the generated responses using three measures: diversity, novelty, and correctness.
• Figure 2: Text Multi-View Embedding.
• Figure 3: This figure illustrates the process of preparing image view embeddings and provides an example for 10 input prompts. Row 1 displays 10 prompts from various subjects. Row 2 shows images crawled based on each input prompt. Row 3 presents the descriptions generated for the images, and Row 4 contains the rewritten descriptions, which serve as our image view embeddings.
• Figure 4: Novelty detection method by using GPT-4o and SBERT.
• Figure 5: Diversity, Novelty, and Correctness (All answers) vs (Correct answers) For $num\_samples = 100$. Diversity measure is $Self\_BLUE$, Novelty measure is based on $SBERT$, and Correctness measure is based on $GPT-4o$.