💬 NLP

A Universal Avoidance Method for Diverse Multi-branch Generation

The paper introduces UAG (Universal Avoidance Generation), a model-agnostic and computationally efficient strategy that significantly enhances multi-branch diversity in both diffusion and transformer models by penalizing output similarity, achieving up to 1.9 times higher diversity with substantially lower computational costs compared to state-of-the-art methods.

Original authors: Kyeongman Park, Minha Jhang, Kyomin Jung

Published 2026-04-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kyeongman Park, Minha Jhang, Kyomin Jung

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

Imagine you have a magical storyteller (a computer program) that can write stories or draw pictures based on a single prompt, like "A cat sitting on a fence."

The problem is that if you ask this storyteller to tell you five different stories about that cat, it often gets lazy. It might just change the cat's name from "Whiskers" to "Mittens" and swap "fence" for "wall," but the plot, the tone, and the structure remain exactly the same. It's like asking a friend to tell you five different jokes, and they just tell you the same joke with a different punchline every time. This is what the paper calls a lack of "multi-branch diversity."

The authors of this paper, Kyeongman Park and his team, created a new tool called UAG (Universal Avoidance Generation) to fix this. Here is how it works, using some simple analogies:

1. The "Anti-Cloning" Force Field

Think of the storyteller as a painter. Usually, if you ask for five paintings, the painter might unconsciously reuse the same brushstrokes or color palette because it's the path of least resistance.

UAG acts like a magnetic force field that pushes the painter away from anything they've already drawn.

The Setup: The painter starts drawing the first picture.
The Push: When they start the second picture, UAG whispers, "Hey, don't use those same blue skies or that specific tree shape you used in the first one! Go somewhere else!"
The Result: The second picture ends up looking completely different—maybe it's a sunset instead of a sunrise, or a cartoon style instead of a realistic one.

2. The "Two-Stage" Strategy (The Road Trip Analogy)

The clever part of UAG is that it knows when to push the painter. It uses a smart schedule, like a road trip with two distinct phases:

Phase 1: The Local Detour (Early Stage)
At the very beginning of the story or image, UAG focuses on small details. It says, "Don't use the same first word or the same starting color." This ensures the stories start with different flavors. It's like telling two friends to start a road trip; one starts by driving north, and the other starts by driving south.
Phase 2: The Global Detour (Late Stage)
As the story or image gets bigger, UAG switches to big-picture thinking. It says, "Don't end up with the same plot twist or the same final scene." This ensures the overall meaning and theme are totally different. It's like ensuring one friend ends up at a beach party and the other ends up at a mountain cabin, even if they started in the same city.

3. Why It's a Game-Changer

Previous methods tried to fix this by making the computer "think" about every single word or pixel individually, which is like trying to solve a maze by checking every single brick in the wall. It's slow, expensive, and requires a super-computer.

UAG is like a GPS shortcut. Instead of checking every brick, it just calculates a simple "push" direction once per step.

Speed: It is 4.4 times faster than the best previous methods.
Efficiency: It uses 64 times less computing power (energy).
Versatility: It works on almost any type of AI, whether it's writing text (like a novel) or creating images (like a painting).

The Bottom Line

Before UAG, asking an AI for "diverse" results was like asking a parrot to say five different things; it would just repeat the same phrase with a slightly different accent.

With UAG, you are giving the AI a creative nudge. It forces the AI to break its own habits, ensuring that if you ask for 10 different stories about a cat, you get 10 stories that feel like they were written by 10 different people, rather than one person trying to be different. It's a simple, fast, and universal way to make AI more creative.

1. Problem Statement

Modern generative models (both autoregressive LLMs and diffusion models) excel at producing high-quality outputs but often lack human-level creativity, specifically in multi-branch diversity. When generating multiple outputs from the same prompt (e.g., different story continuations or image variations), models tend to converge on similar patterns, resulting in repetitive content.

Limitations of Prior Art: Existing methods to enforce diversity (e.g., Avoidance Decoding, Procreate) suffer from two major drawbacks:
1. High Computational Cost: They often rely on token-wise or pixel-wise operations, leading to massive FLOP counts and slow inference speeds.
2. Architecture Dependency: Many approaches are tailored to specific model types (e.g., only for transformers or only for diffusion), limiting their generalizability.

2. Methodology: Universal Avoidance Generation (UAG)

The authors propose UAG, a model-agnostic framework designed to penalize similarity among generated outputs without incurring heavy computational overhead.

Core Concept: UAG operates within a "Universal Generation Process" applicable to both autoregressive and diffusion models. It introduces a gradient-based penalty that pushes the current generation away from previously generated outputs (stored in reference banks).
Two-Component Similarity Loss:
1. Local Similarity Loss ( $L_{local}$ ): Measures similarity at the token (for text) or pixel (for images) level.
2. Global Similarity Loss ( $L_{global}$ ): Measures similarity at the hidden-state (latent) level, capturing semantic meaning.
Step-wise Loss Scheduling: To ensure natural progression, UAG employs a logistic schedule to dynamically weight the two losses:
- Early Stages: Emphasizes local similarity to encourage concept-level diversity (e.g., different starting words or visual elements).
- Later Stages: Emphasizes global similarity to foster semantic-level diversity (e.g., different story arcs or scene layouts).
Gradient Adaptation: Instead of re-generating tokens or pixels, UAG adjusts the output representation ( $y_t$ ) by subtracting the gradient of the combined loss:
$\hat{y}_t = y_t - \nabla_{y_t} L_{UAG}$
This creates a "repulsive force" that diversifies the output.
Efficiency Tricks:
- Analytic Gradients: The method uses analytic formulations for gradients rather than automatic differentiation over the entire model graph, significantly reducing computation.
- Heuristic Projections: For autoregressive models where hidden states do not depend on logits, the method projects the global gradient through the output projection matrix ( $W$ ) as a surrogate. For diffusion models, it approximates the Jacobian as an identity matrix.

3. Key Contributions

Model Agnosticism: UAG is the first framework to unify diversity enhancement across both autoregressive (e.g., LLaMA, GPT) and diffusion (e.g., Stable Diffusion, LLada) models using a single mathematical formulation.
Computational Efficiency: By avoiding token/pixel-wise iterative checks and using simple gradient adjustments, UAG drastically reduces computational requirements compared to state-of-the-art (SOTA) baselines.
Dynamic Scheduling: The introduction of a logistic schedule that transitions from local to global diversity penalties addresses the "early vs. late stage" diversity problem effectively.

4. Experimental Results

The authors evaluated UAG on LLaMA-3B, LLada-8B (diffusion language model), and Stable Diffusion v1.5 using datasets like ReedsyPrompts and COCO.

Diversity Performance:
- Achieved 1.43x higher scores on conventional diversity metrics (BLEU, CLIPscore) and 1.19x higher in LLM-based evaluations compared to baselines.
- In human evaluations, UAG ranked best or second-best in diversity, creativity, and coherence across all tested models.
Efficiency Gains:
- Speed: Runs 4.4 times faster than SOTA methods.
- Compute: Requires only 1/64th of the FLOPs compared to previous methods (e.g., Avoidance Decoding requires ~126x more FLOPs for text; Procreate requires ~1.25x more FLOPs for images).
Quality Trade-off: While there is a slight trade-off in "degeneration" (fluency) due to the aggressive penalty, the outputs remain acceptable and readable, with large-scale models (e.g., LLaMA-70B) showing even better suppression of degeneration.

5. Significance

Scalability: UAG makes diverse multi-branch generation feasible for large-scale applications where computational cost is a bottleneck.
Generalizability: By abstracting the generation process into a universal framework, it bridges the gap between different generative model families, suggesting a unified approach to creativity in AI.
Practical Impact: The method enables the generation of highly diverse content (stories, images) from a single prompt with minimal additional hardware requirements, facilitating applications in creative writing, game design, and content creation.

Limitations: The method is currently limited to models fitting the "Universal Generation Process" (autoregressive and diffusion) and does not apply to GANs. Future work aims to extend this to other generative architectures and address ethical safeguards against excessive diversity.