SIGNAL//DESK
AI securitysrc: Roost FA glossary v2

Mode collapse

When a model gives near-identical outputs across varied prompts. Suggests over-fine-tuning. In your FA practice, this looks like your internal 'expert' bot giving the same generic, useless boilerplate answer for every client query.

Mode collapse occurs when a model's output distribution narrows significantly, causing it to produce near-identical responses across diverse input prompts. In a security or support context, this is a common symptom of over-fine-tuning, where the model prioritizes a rigid, boilerplate response pattern over contextual relevance, effectively rendering the output useless for nuanced queries.

Mode collapse is a failure state in generative modeling characterized by a collapse of the output entropy, where the model maps a wide variety of input prompts to a singular or highly restricted subset of the target distribution. In the context of fine-tuning, this is typically an artifact of over-optimization, where the loss function penalizes deviation from a narrow training objective so heavily that the model loses its capacity for conditional variance, resulting in repetitive, low-utility boilerplate outputs.

evolution

  1. 2014-06 · history
    GAN Introduction

    Ian Goodfellow et al. introduce Generative Adversarial Networks, where mode collapse is first identified as a primary failure mode during training.

  2. 2016-01 · history
    Unrolled GANs

    Researchers propose unrolled GANs to mitigate mode collapse by allowing the generator to anticipate the discriminator's response.

  3. 2017-01 · history
    Wasserstein GAN (WGAN)

    The introduction of the Wasserstein loss function provides a more stable training objective that significantly reduces the occurrence of mode collapse.

  4. 2023-03 · history
    LLM Fine-tuning Collapse

    As large-scale instruction tuning becomes standard, mode collapse is increasingly recognized as a symptom of over-optimization and data homogeneity in LLMs.


← all terms