Research map / Large language models / AI and society

Self-Consuming Generative Models with Co-Evolving Human Preferences

NeurIPS 2026 · Large language models › AI and society · approach: Data commons and human baselines

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Studies self-consuming generative retraining where user preferences co-evolve with model outputs; shows pure curation leads to singleton lock-in, reference-data mixing gives a unique globally attracting equilibrium, and proposes an LP-based algorithm to pick reference distribution and mixing weight at minimum cost.

Problem

Long-term behavior of self-consuming generative model retraining with human curation when user preferences are not fixed but shift in response to model outputs; how to steer equilibria to preserve attributes at low data-collection cost.

Approach

Models a coupled (model distribution, preference vector) dynamical system with K-way Luce curation and preference drift toward the model's mean feature; proves winner lock-in without reference data and global contraction with sufficient reference mixing, then relaxes the bilevel design problem into a family of linear programs indexed by the mixing weight.

Key results

• Under pure synthetic curation (eta=1, CIFAR-10 OT-CFM), airplane proportion rises from ~10% to >95% over 20 rounds with preference locking to airplane.
• With 90% uniform reference data (eta=0.1), class proportions stay near uniform and runs from different initializations converge to a similar equilibrium.
• Designed reference achieves protected classes at >=0.08 at round 20 with eta=0.67, versus uniform baselines requiring lower synthetic ratios.

Surprising findings

• Pure self-consuming curation with strong selection drives output to a single class within 20 rounds, even from a mild initial bias.

Limitations

• Contraction condition eta(L_p+2L_w)<1 uses worst-case Lipschitz constants and is conservative.
• Assumes exact observability of feature map and preferences and a single aggregate preference vector.
• Experiments use single runs per setting due to compute cost.

Experimental setup

Datasets: CIFAR-10 AG News

Benchmarks: CIFAR-10 AG News

Metrics: per-class generation proportion preference vector trajectory minimum class proportion vs threshold collection cost

Baselines: uniform reference data mixing

Builds on: Performative prediction Bradley-Terry model Self-consuming generative models Repeated risk minimization

Closest papers by meaning

• Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.91)
• Stability and Diversity of Networked Self-Consuming Generative Ecosystems (NeurIPS 2026 · similarity 0.90)
• Self-Consuming Generative Models with Adversarially Curated Data (ICML 2025 · similarity 0.90)
• A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops (ICLR 2025 · similarity 0.88)
• Towards Theoretical Understandings of Self-Consuming Generative Models (ICML 2024 · similarity 0.88)
• The two clocks and the innovation window: When and how generative models learn rules (NeurIPS 2026 · similarity 0.87)