# Rank ten papers. Then rank them again.

A practice set of ten historical papers from ICLR 2017 generative-model research. Working titles and summaries are in plain language and bylines are hidden. The papers carry no TenPeers acceptance, and none implies that its authors took part.

Ballot 1 asks: Which papers will other researchers build on and cite? Order the ten by expected research impact over the next two years, most influential first.

Ballot 2 asks: Which talks would you go to in person? Order the same ten, in a fresh shuffle.

The interactive ballot needs JavaScript. The ten papers and their source records:

1. [Training a generator against a future discriminator](https://arxiv.org/abs/1611.02163) To update the generator, the method simulates several steps that the discriminator would take next. The experiments test whether this makes GAN training more stable.
2. [A different likelihood for image pixels](https://arxiv.org/abs/1701.05517) An autoregressive image model represents pixel values with a discretized mixture of logistic distributions. The authors examine how that change and architectural revisions affect training and image likelihood.
3. [Learning a density by reversing a transformation](https://arxiv.org/abs/1605.08803) Invertible coupling layers transform a simple distribution into a more complex one. The model supports sampling and exact likelihood evaluation in the same architecture.
4. [Differentiating through a categorical choice](https://arxiv.org/abs/1611.01144) A soft version of a categorical sample allows gradients to pass through a discrete choice during training. A temperature parameter controls how closely the sample resembles a one-hot decision.
5. [A continuous distribution for discrete variables](https://arxiv.org/abs/1611.00712) The authors define a family of relaxed discrete random variables with a tractable density. They use it to estimate gradients and test the method on density estimation and structured prediction.
6. [Learning an encoder and a generator together](https://arxiv.org/abs/1606.00704) A discriminator compares data–latent pairs from an encoder to pairs from a generator. Both networks learn through the same adversarial objective.
7. [Giving the discriminator an energy function](https://arxiv.org/abs/1609.03126) The discriminator learns to assign lower energy to regions near the data. The generator searches for examples that the discriminator also assigns low energy.
8. [Measuring the likelihood of a decoder’s samples](https://arxiv.org/abs/1611.04273) The authors use annealed importance sampling to evaluate decoder-based generative models. They study the gap between convincing samples and likelihood estimates.
9. [Why adversarial training can stall](https://arxiv.org/abs/1701.04862) This paper analyzes instability and saturation in GAN training. Targeted experiments investigate when the discriminator leaves the generator with an uninformative gradient.
10. [Separating factors of variation in images](https://mlanthology.org/iclr/2017/higgins2017iclr-beta/) A variational autoencoder changes the weight on its latent constraint. The authors study how that trade-off affects the recovery of separate visual factors.
