REVIEW 18 cited by
Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
The proliferation of generative models, combined with pretraining on web-scale data, raises a timely question: what happens when these models are trained on their own generated outputs? Recent investigations into model-data feedback loops proposed that such loops would lead to a phenomenon termed model collapse, under which performance progressively degrades with each model-data feedback iteration until fitted models become useless. However, those studies largely assumed that new data replace old data over time, where an arguably more realistic assumption is that data accumulate over time. In this paper, we ask: what effect does accumulating data have on model collapse? We empirically study this question by pretraining sequences of language models on text corpora. We confirm that replacing the original real data by each generation's synthetic data does indeed tend towards model collapse, then demonstrate that accumulating the successive generations of synthetic data alongside the original real data avoids model collapse; these results hold across a range of model sizes, architectures, and hyperparameters. We obtain similar results for deep generative models on other types of real data: diffusion models for molecule conformation generation and variational autoencoders for image generation. To understand why accumulating data can avoid model collapse, we use an analytically tractable framework introduced by prior work in which a sequence of linear models are fit to the previous models' outputs. Previous work used this framework to show that if data are replaced, the test error increases with the number of model-fitting iterations; we extend this argument to prove that if data instead accumulate, the test error has a finite upper bound independent of the number of iterations, meaning model collapse no longer occurs.
Forward citations
Cited by 18 Pith papers
-
Can Generative Artificial Intelligence Survive Data Contamination? Theoretical Guarantees under Contaminated Recursive Training
Contaminated recursive training converges to the true distribution at rate t^{-min(p, α)} — the slower of the model's baseline rate p and the real-data fraction α.
-
DataOrchestra: Learning to Orchestrate Per-Example Curation of Pretraining Data
A learned orchestrator builds per-example drop/untouch/clean pipelines over noise pruning and instruction-conditioned rewriting, improving from-scratch and math continued pretraining over fixed curation methods.
-
Self-Poisoning in Adaptive Out-of-Distribution Detection: A Sharp-Threshold Theory and Certified Label-Free Calibration
In adaptive OOD detection, bank impurity follows a mean-field urn law whose kernel slope acts as a reproduction number; a frozen-reserve gate removes the supercritical collapse, and a two-world theorem caps label-free...
-
Epidemiology of Model Collapse: Modeling Synthetic Data Contamination via Bilayer SIR Dynamics
A bilayer SIR/SIRS model of data-model cross-contamination yields R0 as the geometric mean of layer rates, shows illustrative supercritical dynamics, and finds detection the highest-leverage lever, with GPT-2 chains m...
-
A Task-Centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula
Iterative self-improvement provably keeps improving only when initial performance lies in a moderate difficulty interval, and easy-to-hard curricula beat fixed mixtures under moderate difficulty separation and suffici...
-
What Matters in LLM-generated Data: Diversity and Its Effect on Model Fine-Tuning
Moderately diverse LLM-generated data can improve fine-tuned model performance in low-data settings when distribution shift is minimal, while high diversity or large distribution shift hurts.
-
Privacy Amplification Through Synthetic Data: Insights from Linear Regression
For linear regression with output perturbation, releasing synthetic points generated from hidden random inputs amplifies differential privacy by roughly the inverse input dimension, unless the adversary controls the seed.
-
SeDi-Instruct: Enhancing Alignment of Language Models through Self-Directed Instruction Generation
SeDi-Instruct generates instruction data by relaxing duplicate filtering, sampling cluster-balanced batches, and replacing low-scoring seeds with instructions from high-gradient-norm batches.
-
DIVE: Diversified Iterative Self-Improvement
DIVE combines global sample pooling with diversity-aware data selection to counter output diversity collapse in iterative preference learning for LLMs.
-
How to Synthesize Text Data without Model Collapse?
Token-level editing of human text with a high-confidence threshold yields modest performance gains over the original data and avoids the sharp degradation seen with purely synthetic data.
-
Analyzing and Mitigating Model Collapse in Rectified Flow Models
Rectified Flow's Reflow training degrades over iterations on self-generated data, and the paper's Real-data Augmented Reflow mixes real reverse-sampled pairs to prevent this collapse and enable fewer-step sampling.
-
Epistemic diversity across language models mitigates knowledge collapse
In repeated self-training loops on Wikitext2, ecosystems of four small language models show lower average perplexity than one, two, or sixteen models, but the paper's broader claims about monotonic optima, robustness,...
-
Sealing The Backdoor: Unlearning Adversarial Text Triggers In Diffusion Models Using Knowledge Distillation
SKD-CAG erases adversarial text triggers from diffusion models by distilling the model's own clean outputs through cross-attention guidance, claiming 100% and 93% removal for pixel and style backdoors.
-
Position: Machine Learning Conferences Should Establish a "Refutations and Critiques" Track
ML conferences should create an official peer-reviewed track dedicated to refuting and critiquing previously published work.
-
Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges
Iterative self-training on a model's own correct outputs, with simple length and voting filters, lets transformers generalize to far longer arithmetic and path-finding problems than they saw in training.
-
LLM Web Dynamics: Tracing Model Collapse in a Network of LLMs
Under a shared retrieval-augmented memory, multiple LLMs' outputs converge to near-identical semantic answers, and the analogous Gaussian mixture system is proven to collapse.
-
The Anti-Ouroboros Effect: Emergent Resilience in Large Language Models from Recursive Selective Feedback
A recursive fine-tuning study claims quality filtering reverses model collapse, but the paper's own data show the filtered model only matched its starting score.
-
Theoretical Proof that Auto-regressive Language Models Collapse when Real-world Data is a Finite Set
The paper's proof of inevitable language-model collapse reduces to a definitional identity, since the error terms whose accumulation drives the result are chosen to fit the model outputs rather than derived from train...
Discussion (0). Continue with ORCID to comment.