Noise Augmented Training for Stable Autoregressive Generation

48 concepts37 papers1 proposal

Autoregressive models are exceptionally powerful at generating sequential data, from text and audio to video and physical simulations. By breaking down the joint probability of a sequence into a chain of conditional next-step predictions, these models can capture complex temporal dependencies.

However, they suffer from a fundamental flaw during inference: exposure bias. Noise-augmented training addresses this vulnerability by transforming autoregressive models from brittle predictors into robust, self-correcting dynamical systems.

The Exposure Bias Dilemma

During standard training, autoregressive models are typically optimized using a paradigm called teacher forcing. The model is tasked with predicting the next token or state given a history composed entirely of perfect, ground-truth data.

Training (Teacher Forcing):
[Ground Truth t-2] -> [Ground Truth t-1] -> Model -> Predicts [Clean t]

Inference (Autoregressive Rollout):
[Imperfect Pred t-2] -> [Imperfect Pred t-1] -> Model -> Predicts [Divergent t]

At inference time, the teacher disappears. The model must now condition its predictions on its own previous outputs. If the model makes a minor error at step $t$, the input at step $t+1$ is slightly out-of-distribution. Because the model was never exposed to imperfect histories during training, it does not know how to handle this shift.

As highlighted by Li et al., this creates a severe covariate shift where early-stage generation errors compound over later steps. In physical simulations, this error accumulation causes trajectories to rapidly diverge or explode (Sanchez-Gonzalez et al.). In video or audio synthesis, it manifests as severe blurring, physical illogic, or temporal freezing (Maduabuchi et al.; Hou et al.).

The Intuition of Self-Correction

The core idea behind noise-augmented training is simple: deliberately corrupt the historical context during training so the model learns how to recover from errors. Instead of learning a mapping from a perfect past to a perfect future, the model is trained to predict clean future states from imperfect past states (Hao et al.).

Geometrically, valid data sequences live on a low-dimensional equilibrium manifold embedded within a high-dimensional space. Standard teacher forcing only teaches the model how to navigate exactly on this manifold. If the model steps off the manifold even slightly during inference, it enters a vacuum with no directional guidance.

By injecting noise into the inputs during training, we populate the empty space surrounding the manifold with training examples. The model learns a corrective vector field. When it encounters an off-manifold or perturbed state, its learned dynamics naturally pull the trajectory back toward the stable equilibrium manifold (Thomaszewski et al.; Fragkiadaki et al.).

Mechanisms of Noise Injection

Depending on the data type and model architecture, researchers apply noise in several distinct ways to teach this error-correction behavior.

Gaussian and Random-Walk Noise

In continuous domains like physics simulations, geometry tracking, and motion modeling, adding isotropic Gaussian noise to input coordinates or latent states is highly effective.

For instance, Philion added Gaussian noise to ground-truth coordinates in sequential lane detection, forcing the model to learn trajectories that return to the true path after a deviation. For complex particle simulations, Sanchez-Gonzalez et al. utilized random-walk noise to simulate the exact type of multi-step drift that occurs during long-term rollouts, ensuring physical stability over thousands of frames.

Token-Level and Diffusion-Based Noise

In generative modeling, recent frameworks have begun bridging the gap between autoregressive prediction and diffusion models. Chen et al. introduced Diffusion Forcing, a paradigm that applies independent, variable noise levels to each token in a sequence during training. This allows the model to be trained on highly noisy historical contexts while learning to generate continuous data stably over infinitely long horizons.

Similarly, in long-range video synthesis, Zhang et al. conditioned motion models on noisy versions of previous states, training the network to actively denoise its historical context and eliminating the facial distortions common in standard autoregressive generation.

Structured and Contextual Noise

Rather than completely random static, noise can be structured to match the data's native distribution. Maduabuchi et al. injected data-aligned noise during text-to-video training to make models resilient against semantic drift.

In continuous audio modeling, Rouard et al. proposed a dual-path approach: injecting noise into long-term history while keeping short-term context clean. This stabilizes the continuous autoregressive generation over long horizons without sacrificing high-frequency audio details.

Preventing Exploitation and Drift

Beyond stabilizing generation, noise injection serves as a crucial regularizer when autoregressive models are used downstream.

In model-based reinforcement learning and inverse design, optimization algorithms frequently query a learned dynamics model to find an optimal path. If the model has minor inaccuracies, the optimizer will quickly discover and exploit these "unphysical" shortcuts, steering the agent into unrealistic regions of the state space.

As Allen et al. demonstrated, incorporating stochastic noise into the training of graph network simulators prevents optimizers from exploiting these ungrounded regions. Because the model has learned to map perturbed states back to plausible trajectories, the optimization process is successfully constrained within physically valid boundaries.

Ultimately, noise-augmented training works because it aligns the training distribution with the operational realities of inference. By treating the generation process as an ongoing defense against recursive error amplification, models achieve the robustness required for long-horizon stability (Zhou et al.; Xie and Sabour).

Key Foundational Work

Visual Interaction Networks: Learning a Physics Simulator from Video (2017) — contributed 1 concept elements to this direction.
Recurrent Network Models for Human Dynamics (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_noise_robust_autoregressive_training_1f429d94,
  title = {noise-robust autoregressive training},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/noise-robust-autoregressive-training/rdr_1f429d94299b47d0a622d45de9388703},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_noise_robust_autoregressive_training_1f429d94_v1,
  title = {noise-robust autoregressive training},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/noise-robust-autoregressive-training/rdr_1f429d94299b47d0a622d45de9388703/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake