Experiments across code LLMs show no-review collapses fastest, human-gated filters slow collapse, and AI self-gates lose effect over time, degenerating to ungated self-training under self-confirming acceptance as proven via gated distributional reweighting and spectral analysis.
How Bad is Training on Synthetic Data?
7 Pith papers cite this work, alongside 3 external citations. Polarity classification is still indexing.
7
Pith papers citing it
3
external citations · external index
years
2026 7verdicts
UNVERDICTED 7representative citing papers
Activation steering produces synthetic safety-violating data that improves downstream classifiers over prompting on most tested concepts when a harmonic mean of alignment, coherence, and diversity is optimized.
Iterated learning theory predicts and LLM experiments confirm non-monotonic compositionality during self-training, reframing model collapse as cultural transmission with matching human regularization patterns.
Self-training restructures language by amplifying surface markers and collapsing deep syntax according to structural depth rather than frequency, as evidenced by correlations across multiple models and a human fine-tuning control.
Iterative self-finetuning of LLMs mostly fails to amplify seeded behavioral traits, with amplification limited to specific DPO setups and often harming coherence.
A game-theoretic model shows that individually rational adoption of generative AI causes model collapse that reduces collective social welfare for important tasks, with habit formation creating spillovers from low-stakes to high-value domains.
Model collapse threatens AI democratization by disproportionately impacting low-resource and marginalized communities through reduced training efficiency and data distributions skewed away from distribution tails.
citing papers explorer
-
When AI Reviews Its Own Code: Recursive Self-Training Collapse in Code LLMs
Experiments across code LLMs show no-review collapses fastest, human-gated filters slow collapse, and AI self-gates lose effect over time, degenerating to ungated self-training under self-confirming acceptance as proven via gated distributional reweighting and spectral analysis.
-
Activation Steering for Synthetic Data Generation: The Role of Diversity in Downstream Safety Detection
Activation steering produces synthetic safety-violating data that improves downstream classifiers over prompting on most tested concepts when a harmonic mean of alignment, coherence, and diversity is optimized.
-
Model Collapse as Cultural Evolution
Iterated learning theory predicts and LLM experiments confirm non-monotonic compositionality during self-training, reframing model collapse as cultural transmission with matching human regularization patterns.
-
Self-Training Doesn't Flatten Language -- It Restructures It: Surface Markers Amplify While Deep Syntax Dies
Self-training restructures language by amplifying surface markers and collapsing deep syntax according to structural depth rather than frequency, as evidenced by correlations across multiple models and a human fine-tuning control.
-
Iterative Finetuning is Mostly Idempotent
Iterative self-finetuning of LLMs mostly fails to amplify seeded behavioral traits, with amplification limited to specific DPO setups and often harming coherence.
-
Generative artificial intelligence reduces social welfare through model collapse
A game-theoretic model shows that individually rational adoption of generative AI causes model collapse that reduces collective social welfare for important tasks, with habit formation creating spillovers from low-stakes to high-value domains.
-
Position: the Stochastic Parrot in the Coal Mine. Model Collapse is a Threat to Low-Resource Communities
Model collapse threatens AI democratization by disproportionately impacting low-resource and marginalized communities through reduced training efficiency and data distributions skewed away from distribution tails.