Learning by Surprise: Surplexity for Mitigating Model Collapse in Generative AI

• URL: http://arxiv.org/abs/2410.12341v3
• Date: Tue, 02 Sep 2025 13:11:32 GMT
• Title: Learning by Surprise: Surplexity for Mitigating Model Collapse in Generative AI
• Authors: Daniele Gambetta, Gizem Gezici, Fosca Giannotti, Dino Pedreschi, Alistair Knott, Luca Pappalardo,
• Abstract summary: As synthetic content infiltrates the web, generative AI models may be retrained on their own outputs.<n>This leads to model collapse: a progressive loss of performance and diversity across generations.<n>We introduce new measures that characterise collapse directly from a model's next-token probability distributions.
• Score: 1.6545633988217645
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: As synthetic content increasingly infiltrates the web, generative AI models may be retrained on their own outputs: a process termed "autophagy". This leads to model collapse: a progressive loss of performance and diversity across generations. Recent studies have examined the emergence of model collapse across various generative AI models and data types, and have proposed mitigation strategies that rely on incorporating human-authored content. However, current characterizations of model collapse remain limited, and existing mitigation methods assume reliable knowledge of whether training data is human-authored or AI-generated. In this paper, we address these gaps by introducing new measures that characterise collapse directly from a model's next-token probability distributions, rather than from properties of AI-generated text. Using these measures, we show that the degree of collapse depends on the complexity of the initial training set, as well as on the extent of autophagy. Our experiments prompt a new suggestion: that model collapse occurs when a model trains on data that does not "surprise" it. We express this hypothesis in terms of the well-known Free Energy Principle in cognitive science. Building on this insight, we propose a practical mitigation strategy: filtering training items by high surplexity, maximising the surprise of the model. Unlike existing methods, this approach does not require distinguishing between human- and AI-generated data. Experiments across datasets and models demonstrate that our strategy is at least as effective as human-data baselines, and even more effective in reducing distributional skewedness. Our results provide a richer understanding of model collapse and point toward more resilient approaches for training generative AI systems in environments increasingly saturated with synthetic data.

Related papers

• Can Generative Artificial Intelligence Survive Data Contamination? Theoretical Guarantees under Contaminated Recursive Training [11.253812961752958]
  Generative Artificial Intelligence (AI) has become a transformative force across science, industry, and society.<n>As these systems grow in popularity, web data becomes increasingly interwoven with this AI-generated material.<n>As generative models are updated regularly, later models will inevitably be trained on mixtures of human-generated data and AI-generated data from earlier versions.
  arXiv  Detail & Related papers  (2026-02-17T22:38:18Z)
• Escaping Model Collapse via Synthetic Data Verification: Near-term Improvements and Long-term Convergence [31.751930228965467]
  We investigate ways to modify this synthetic retraining process to avoid model collapse.<n>Our key finding is that by injecting information through an external synthetic data verifier, synthetic retraining will not cause model collapse.
  arXiv  Detail & Related papers  (2025-10-18T22:39:39Z)
• ForTIFAI: Fending Off Recursive Training Induced Failure for AI Model Collapse [13.096745830570944]
  We introduce the Truncated-Cross-Entropy (TCE) loss function to mitigate model collapse in synthetic data.<n>TCE mitigates collapse by selectively ignoring high-confidence tokens during training, effectively filtering out likely machine-generated artifacts.<n>Our results demonstrate that confidence-aware training objectives can substantially delay collapse onset, offering a practical and generalizable tool for model robustness under synthetic-data exposure.
  arXiv  Detail & Related papers  (2025-09-10T20:06:51Z)
• Multi-modal Synthetic Data Training and Model Collapse: Insights from VLMs and Diffusion Models [24.73190742678142]
  We study the risk of generative model collapse in multi-modal vision-language generative systems.<n>We find that model collapse exhibits distinct characteristics in the multi-modal context, such as improved vision-language alignment and increased variance in image-captioning task.<n>Our findings provide initial insights and practical guidelines for reducing the risk of model collapse in self-improving multi-agent AI systems.
  arXiv  Detail & Related papers  (2025-05-10T22:42:29Z)
• A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops [55.07063067759609]
  High-quality data is essential for training large generative models, yet the vast reservoir of real data available online has become nearly depleted.<n>Models increasingly generate their own data for further training, forming Self-consuming Training Loops (STLs)<n>Some models degrade or even collapse, while others successfully avoid these failures, leaving a significant gap in theoretical understanding.
  arXiv  Detail & Related papers  (2025-02-26T06:18:13Z)
• Machine-generated text detection prevents language model collapse [16.73923337828054]
  We train a machine-generated text detector and propose an importance resampling approach to prevent model collapse.<n>We demonstrate that it not only prevents model collapse but also improves performance compared to training on purely human data.
  arXiv  Detail & Related papers  (2025-02-21T18:22:36Z)
• Harnessing the Intrinsic Knowledge of Pretrained Language Models for Challenging Text Classification Settings [5.257719744958367]
  This thesis explores three challenging settings in text classification by leveraging the intrinsic knowledge of pretrained language models (PLMs) We develop models that utilize features based on contextualized word representations from PLMs, achieving performance that rivals or surpasses human accuracy. Lastly, we tackle the sensitivity of large language models to in-context learning prompts by selecting effective demonstrations.
  arXiv  Detail & Related papers  (2024-08-28T09:07:30Z)
• Creativity Has Left the Chat: The Price of Debiasing Language Models [1.223779595809275]
  We investigate the unintended consequences of Reinforcement Learning from Human Feedback on the creativity of Large Language Models (LLMs) Our findings have significant implications for marketers who rely on LLMs for creative tasks such as copywriting, ad creation, and customer persona generation.
  arXiv  Detail & Related papers  (2024-06-08T22:14:51Z)
• When AI Eats Itself: On the Caveats of AI Autophagy [18.641925577551557]
  The AI autophagy phenomenon suggests a future where generative AI systems may increasingly consume their own outputs without discernment. This study examines the existing literature, delving into the consequences of AI autophagy, analyzing the associated risks, and exploring strategies to mitigate its impact.
  arXiv  Detail & Related papers  (2024-05-15T13:50:23Z)
• Corpus Considerations for Annotator Modeling and Scaling [9.263562546969695]
  We show that the commonly used user token model consistently outperforms more complex models. Our findings shed light on the relationship between corpus statistics and annotator modeling performance.
  arXiv  Detail & Related papers  (2024-04-02T22:27:24Z)
• Heat Death of Generative Models in Closed-Loop Learning [63.83608300361159]
  We study the learning dynamics of generative models that are fed back their own produced content in addition to their original training dataset. We show that, unless a sufficient amount of external data is introduced at each iteration, any non-trivial temperature leads the model to degenerate.
  arXiv  Detail & Related papers  (2024-04-02T21:51:39Z)
• generAItor: Tree-in-the-Loop Text Generation for Language Model Explainability and Adaptation [28.715001906405362]
  Large language models (LLMs) are widely deployed in various downstream tasks, e.g., auto-completion, aided writing, or chat-based text generation. We tackle this shortcoming by proposing a tree-in-the-loop approach, where a visual representation of the beam search tree is the central component for analyzing, explaining, and adapting the generated outputs. We present generAItor, a visual analytics technique, augmenting the central beam search tree with various task-specific widgets, providing targeted visualizations and interaction possibilities.
  arXiv  Detail & Related papers  (2024-03-12T13:09:15Z)
• DPP-Based Adversarial Prompt Searching for Lanugage Models [56.73828162194457]
  Auto-regressive Selective Replacement Ascent (ASRA) is a discrete optimization algorithm that selects prompts based on both quality and similarity with determinantal point process (DPP) Experimental results on six different pre-trained language models demonstrate the efficacy of ASRA for eliciting toxic content.
  arXiv  Detail & Related papers  (2024-03-01T05:28:06Z)
• Towards Theoretical Understandings of Self-Consuming Generative Models [56.84592466204185]
  This paper tackles the emerging challenge of training generative models within a self-consuming loop. We construct a theoretical framework to rigorously evaluate how this training procedure impacts the data distributions learned by future models. We present results for kernel density estimation, delivering nuanced insights such as the impact of mixed data training on error propagation.
  arXiv  Detail & Related papers  (2024-02-19T02:08:09Z)
• From Identifiable Causal Representations to Controllable Counterfactual Generation: A Survey on Causal Generative Modeling [17.074858228123706]
  We focus on fundamental theory, methodology, drawbacks, datasets, and metrics. We cover applications of causal generative models in fairness, privacy, out-of-distribution generalization, precision medicine, and biological sciences.
  arXiv  Detail & Related papers  (2023-10-17T05:45:32Z)
• On the Stability of Iterative Retraining of Generative Models on their own Data [56.153542044045224]
  We study the impact of training generative models on mixed datasets. We first prove the stability of iterative training under the condition that the initial generative models approximate the data distribution well enough. We empirically validate our theory on both synthetic and natural images by iteratively training normalizing flows and state-of-the-art diffusion models.
  arXiv  Detail & Related papers  (2023-09-30T16:41:04Z)
• RenAIssance: A Survey into AI Text-to-Image Generation in the Era of Large Model [93.8067369210696]
  Text-to-image generation (TTI) refers to the usage of models that could process text input and generate high fidelity images based on text descriptions. Diffusion models are one prominent type of generative model used for the generation of images through the systematic introduction of noises with repeating steps. In the era of large models, scaling up model size and the integration with large language models have further improved the performance of TTI models.
  arXiv  Detail & Related papers  (2023-09-02T03:27:20Z)
• Self-Consuming Generative Models Go MAD [21.056900382589266]
  We study how to use synthetic data to train generative AI algorithms for imagery, text, and other data types. Without enough fresh real data in each generation of an autophagous loop, future generative models are doomed to have their quality (precision) or diversity (recall) progressively decrease. We term this condition Model Autophagy Disorder (MAD), making analogy to mad cow disease.
  arXiv  Detail & Related papers  (2023-07-04T17:59:31Z)
• CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing [139.77117915309023]
  CRITIC allows large language models to validate and amend their own outputs in a manner similar to human interaction with tools. Comprehensive evaluations involving free-form question answering, mathematical program synthesis, and toxicity reduction demonstrate that CRITIC consistently enhances the performance of LLMs.
  arXiv  Detail & Related papers  (2023-05-19T15:19:44Z)
• Less is More: Mitigate Spurious Correlations for Open-Domain Dialogue Response Generation Models by Causal Discovery [52.95935278819512]
  We conduct the first study on spurious correlations for open-domain response generation models based on a corpus CGDIALOG curated in our work. Inspired by causal discovery algorithms, we propose a novel model-agnostic method for training and inference of response generation model.
  arXiv  Detail & Related papers  (2023-03-02T06:33:48Z)
• Large Language Models with Controllable Working Memory [64.71038763708161]
  Large language models (LLMs) have led to a series of breakthroughs in natural language processing (NLP) What further sets these models apart is the massive amounts of world knowledge they internalize during pretraining. How the model's world knowledge interacts with the factual information presented in the context remains under explored.
  arXiv  Detail & Related papers  (2022-11-09T18:58:29Z)
• Model Criticism for Long-Form Text Generation [113.13900836015122]
  We apply a statistical tool, model criticism in latent space, to evaluate the high-level structure of generated text. We perform experiments on three representative aspects of high-level discourse -- coherence, coreference, and topicality. We find that transformer-based language models are able to capture topical structures but have a harder time maintaining structural coherence or modeling coreference.
  arXiv  Detail & Related papers  (2022-10-16T04:35:58Z)
• Diversity vs. Recognizability: Human-like generalization in one-shot generative models [5.964436882344729]
  We propose a new framework to evaluate one-shot generative models along two axes: sample recognizability vs. diversity. We first show that GAN-like and VAE-like models fall on opposite ends of the diversity-recognizability space. In contrast, disentanglement transports the model along a parabolic curve that could be used to maximize recognizability.
  arXiv  Detail & Related papers  (2022-05-20T13:17:08Z)
• Closed-form Continuous-Depth Models [99.40335716948101]
  Continuous-depth neural models rely on advanced numerical differential equation solvers. We present a new family of models, termed Closed-form Continuous-depth (CfC) networks, that are simple to describe and at least one order of magnitude faster.
  arXiv  Detail & Related papers  (2021-06-25T22:08:51Z)
• Contrastive Model Inversion for Data-Free Knowledge Distillation [60.08025054715192]
  We propose Contrastive Model Inversion, where the data diversity is explicitly modeled as an optimizable objective. Our main observation is that, under the constraint of the same amount of data, higher data diversity usually indicates stronger instance discrimination. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that CMI achieves significantly superior performance when the generated data are used for knowledge distillation.
  arXiv  Detail & Related papers  (2021-05-18T15:13:00Z)
• Topical Language Generation using Transformers [4.795530213347874]
  This paper presents a novel approach for Topical Language Generation (TLG) by combining a pre-trained LM with topic modeling information. We extend our model by introducing new parameters and functions to influence the quantity of the topical features presented in the generated text. Our experimental results demonstrate that our model outperforms the state-of-the-art results on coherency, diversity, and fluency while being faster in decoding.
  arXiv  Detail & Related papers  (2021-03-11T03:45:24Z)
• How Faithful is your Synthetic Data? Sample-level Metrics for Evaluating and Auditing Generative Models [95.8037674226622]
  We introduce a 3-dimensional evaluation metric that characterizes the fidelity, diversity and generalization performance of any generative model in a domain-agnostic fashion. Our metric unifies statistical divergence measures with precision-recall analysis, enabling sample- and distribution-level diagnoses of model fidelity and diversity.
  arXiv  Detail & Related papers  (2021-02-17T18:25:30Z)
• Anomaly Detection of Time Series with Smoothness-Inducing Sequential Variational Auto-Encoder [59.69303945834122]
  We present a Smoothness-Inducing Sequential Variational Auto-Encoder (SISVAE) model for robust estimation and anomaly detection of time series. Our model parameterizes mean and variance for each time-stamp with flexible neural networks. We show the effectiveness of our model on both synthetic datasets and public real-world benchmarks.
  arXiv  Detail & Related papers  (2021-02-02T06:15:15Z)
• Firearm Detection via Convolutional Neural Networks: Comparing a Semantic Segmentation Model Against End-to-End Solutions [68.8204255655161]
  Threat detection of weapons and aggressive behavior from live video can be used for rapid detection and prevention of potentially deadly incidents. One way for achieving this is through the use of artificial intelligence and, in particular, machine learning for image analysis. We compare a traditional monolithic end-to-end deep learning model and a previously proposed model based on an ensemble of simpler neural networks detecting fire-weapons via semantic segmentation.
  arXiv  Detail & Related papers  (2020-12-17T15:19:29Z)
• On the Transferability of Adversarial Attacksagainst Neural Text Classifier [121.6758865857686]
  We investigate the transferability of adversarial examples for text classification models. We propose a genetic algorithm to find an ensemble of models that can induce adversarial examples to fool almost all existing models. We derive word replacement rules that can be used for model diagnostics from these adversarial examples.
  arXiv  Detail & Related papers  (2020-11-17T10:45:05Z)
• Improving the Reconstruction of Disentangled Representation Learners via Multi-Stage Modeling [54.94763543386523]
  Current autoencoder-based disentangled representation learning methods achieve disentanglement by penalizing the ( aggregate) posterior to encourage statistical independence of the latent factors. We present a novel multi-stage modeling approach where the disentangled factors are first learned using a penalty-based disentangled representation learning method. Then, the low-quality reconstruction is improved with another deep generative model that is trained to model the missing correlated latent variables.
  arXiv  Detail & Related papers  (2020-10-25T18:51:15Z)
• Multi-Fact Correction in Abstractive Text Summarization [98.27031108197944]
  Span-Fact is a suite of two factual correction models that leverages knowledge learned from question answering models to make corrections in system-generated summaries via span selection. Our models employ single or multi-masking strategies to either iteratively or auto-regressively replace entities in order to ensure semantic consistency w.r.t. the source text. Experiments show that our models significantly boost the factual consistency of system-generated summaries without sacrificing summary quality in terms of both automatic metrics and human evaluation.
  arXiv  Detail & Related papers  (2020-10-06T02:51:02Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.