Generalization in Generation: A closer look at Exposure Bias

By Florian Schmidt. 2019

152 citations recorded by the source corpus.

Exposure bias refers to the train-test discrepancy that seemingly arises when an autoregressive generative model uses only ground-truth contexts at training time but generated ones at test time. We separate the contribution of the learning framework and the model to clarify the debate on consequences and review proposed counter-measures. In this light, we argue that generalization is the underlying property to address and propose unconditional generation as its fundamental benchmark. Finally, we combine latent variable modeling with a recent formulation of exploration in reinforcement learning to obtain a rigorous handling of true and generated contexts. Results on language modeling and variational sentence auto-encoding confirm the model’s generalization capability.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

In Generative Models

Citations in the corpus

Learning to summarize from human feedback (2020, background): “are incentivized to place probability mass on all human demonstrations, including those that are low-quality; and distributional shift during sampling can degrade performance [56, 52].”
ImageBART: Bidirectional Context with Multinomial Diffusion for Autoregressive Image Synthesis (2021, background): “Their samples can still exert artifacts resulting from a sequential modeling of components, and, since these models are always trained by “teacher-forcing”, they are susceptible to an exposure bias [3, 54, 23, 60, 40].”
Diffsound: Discrete Diffusion Model for Text-to-Sound Generation (2022, background): “predicted tokens can be affected by the “accumulated errors” in previous predicted tokens in inference [43].”
Why Exposure Bias Matters: An Imitation Learning Perspective of Error Accumulation in Language Generation (2022, background): “Finally, previous works have linked exposure bias to out-of-domain (Wang and Sennrich, 2020) and out of distribution (Schmidt, 2019) generalization, and hallucinations (Wang and Sennrich, 2020) but these claims remain weak in absence of a clear and principled formalization of”
Teacher-Student Training For Robust Tacotron-Based TTS (2019, methodology): “A typical neural TTS system suffers from the exposure bias problem [16,17] in the autoregressive model [18] that is used by the decoder module.”
Relating Neural Text Degeneration to Exposure Bias (2021, methodology): “Since then, there have been many methods proposed to tackle this problem (Bengio et al., 2015; Huszár, 2015; Ranzato et al., 2016; Lamb et al., 2016; Yu et al., 2017; Wiseman and Rush, 2016; Schmidt, 2019; Zhang et al., 2019a; Wang and Sennrich, 2020).”
Exposure Bias versus Self-Recovery: Are Distortions Really Incremental for Autoregressive Text Generation? (2019, background): “Schmidt (2019) provides valuable discussions around the claim of exposure bias, but they did not propose an operational definition or metric.”
Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (2025, unknown): “This issue is more broadly known as exposure bias [60, 71], where a model is trained exclusively on ground-truth context but must rely on its own imperfect predictions at inference time, resulting in a distributional mismatch that compounds errors as generation progresses.”