Mode Collapse
Originally describing the phenomenon in generative adversarial networks where the generator's output diversity collapses, the term broadly refers to a failure mode in machine learning models where outputs converge toward uniformity during training and the model becomes unable to respond to variations in input.
Timeline (last 90 days)
Mode collapse refers to the generator repeatedly outputting samples that easily fool the discriminator, resulting in severely insufficient diversity in the generated results.
Training models with reinforcement learning for parallel thinking has the fundamental flaw of mode collapse, namely that the multiple parallel paths output by the model degenerate into highly similar trajectories.
Mode collapse was originally a term in generative models (such as GANs), referring to the generator tending to produce only a few kinds of samples and being unable to cover all modes of the data distribution; in the reinforcement learning context it manifests as the policy converging prematurely onto a few high-reward trajectories.
LLMs tend to build things within a fixed set of patterns, and directly accepting their first output leads to homogeneous projects.
Both hypernetwork extensions exhibit mode collapse, and changing the preference weights can hardly cause the model to move systematically along the expected fairness–performance frontier.
All Facts (5)
Mode collapse refers to the generator repeatedly outputting samples that easily fool the discriminator, resulting in severely insufficient diversity in the generated results.
50%UnverifiedTraining models with reinforcement learning for parallel thinking has the fundamental flaw of mode collapse, namely that the multiple parallel paths output by the model degenerate into highly similar trajectories.
50%UnverifiedMode collapse was originally a term in generative models (such as GANs), referring to the generator tending to produce only a few kinds of samples and being unable to cover all modes of the data distribution; in the reinforcement learning context it manifests as the policy converging prematurely onto a few high-reward trajectories.
50%UnverifiedLLMs tend to build things within a fixed set of patterns, and directly accepting their first output leads to homogeneous projects.
50%UnverifiedBoth hypernetwork extensions exhibit mode collapse, and changing the preference weights can hardly cause the model to move systematically along the expected fairness–performance frontier.
50%