Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Abstract: The deep learning revolution, kicked off by AlexNet, taught us that end-to-end training beats decomposing a problem into hand-designed stages. Generative modeling, however, has remained the exception-despite generative models being remarkably capable, they are still not trained end-to-end. This is because, at its core, generative modeling is about handling distributions with many modes, and existing scalable approaches handle this the same way, by factoring the generation procedure, which prevents end-to-end generation. In this work, we introduce Explorative Modeling, a new paradigm that instead factors the training loop, exploring K candidate matches between model generations and data, and training on the best, so predictions commit to modes rather than blurring them. We find Explorative Models (XMs) useful in two settings. First, increasing exploration adds a third pretraining axis beyond parameters and data for existing generative models-where scaling exploration monotonically improves performance across both continuous and discrete domains (images, video, and language). Notably, gains from exploration increase with scale, climbing from 7% to 36% as data scales and from 13% to 23% as models grow, with efficiency gains more than doubling at 3x the compute. Concretely, exploration improves FLOP efficiency by 4.1x, sample efficiency by 6.2x, parameter efficiency by 47%, lifts the strongest of image-generation recipes to a near-state-of-the-art 1.43 FID on ImageNet without guidance, enables scaling how end-to-end existing models are, and unlocks scaling generalization. Second, XMs enable end-to-end reconstructive generative modeling, matching diffusion on control tasks with 16-256x fewer inference steps. Together, these results establish XMs as both a new pretraining axis for existing generative models and a standalone end-to-end generative modeling paradigm.

• The paper introduces Explorative Modeling, a novel training strategy that factors the training loop to capture diverse data modes using a best-of-K approach.
• It demonstrates significant improvements including 4.1× FLOP efficiency, 6.2× sample efficiency, and 47% parameter efficiency on large-scale image generation tasks.
• The approach challenges conventional generative models by enabling true end-to-end reconstructive generation and reducing inference compute.

Explorative Modeling: A Third Pretraining Axis for Generative Models

Introduction and Motivation

The current paradigm in deep learning leverages end-to-end training, a strategy that has consistently outperformed fragmented, hand-engineered components across domains such as classification, detection, and segmentation. However, generative modeling, particularly for complex and multimodal data, has historically evaded truly end-to-end protocols. This is due to the necessity of handling multimodal data distributions, for which prevailing methods (autoregressive, diffusion, and flow models) factorize the generation process, making generation at training fundamentally different from inference.

The paper "Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation" (2607.27372) introduces Explorative Modeling (XM), a framework that decouples the necessity of generation factorization by instead factorizing the training loop. This approach adds a new axis—exploration—that directly scales generative expressivity, complementing the established axes of model capacity and dataset size.

Figure 1: A generative model can factor either generation or training. Most models factor generation (right column), preventing end-to-end operation; Explorative Modeling factors the training loop (top row), enabling training on modes a model captures best and allowing end-to-end generation.

Methodology: Explorative Modeling Paradigm

Conventional reconstructive generative models are restricted by the unknown coupling between latent variables and data samples, causing regression objectives to blur distinct modes and undermine expressivity. Explorative Modeling addresses this by, during each training step, generating $K$ candidate samples for each data point and only backpropagating through the candidate best matching the data. Formally, given a data sample $x$ and $K$ generations $\hat{y}_i$, the objective is

$$\mathcal{L}(\theta) = \min_{i \in \{1, \ldots, K\}} J(\hat{y}_i, x)$$

with $J$ commonly being a reconstruction loss (e.g., squared error). This “best-of-$K$” strategy enables the model to assign different latents to distinct data modes, thereby preventing the averaging effect inherent in standard regression. As $K$ increases, the model can capture more complex, multimodal distributions.

Two variants are defined: Forward XM, which matches each data point to the best of $K$ generations (favoring recall and coverage), and Reverse XM, which matches each generation to its nearest data sample (favoring precision). These dual mechanisms allow tuning for desired tradeoffs.

Theoretical and Empirical Findings

Generative Expressivity as a Scaling Axis

Explorative Modeling directly scales what the authors define as generative expressivity—the number of distribution modes a generative model can capture. In traditional setups, generative expressivity is capped by the factored structure of the generative process, and no amount of extra parameters or additional data can compensate for this. Explorative Modeling directly increases expressivity proportional to $K$.

Figure 2: Explorative Modeling explores matches between model generations and data, lifting generative expressivity and capturing distinct modes rather than regressing toward a mean.

Numerical Results and Scaling Behavior

Exploration as a pretraining axis yields robust efficiency and performance gains that empirically increase with scale:

• FLOP efficiency improves by $x$0, sample efficiency by $x$1, and parameter efficiency by $x$2 in large-scale image generation (ImageNet 256$x$3256).
• Performance gains from exploration amplify as model/data scale increases, with improvement rates escalating from $x$4 to $x$5 as data scales, and $x$6 to $x$7 as model size increases.
• State-of-the-art image-generation recipes are lifted by exploration, achieving a near-SOTA FID of $x$8 on ImageNet without guidance.
• Explorative Modeling achieves substantial generalization improvements, displaying reduced overfitting and enabling compute-generalization tradeoffs.

End-to-End Reconstructive Generation

A particularly important result is that, by maximizing exploration ($x$9), XM recovers end-to-end reconstructive generation: sampling and training strictly coincide. Empirical evaluation on robotics control and world modeling tasks demonstrates that Explorative Policies and World Models match diffusion-based baselines, but at $K$0–$K$1 less inference compute, reducing typical diffusion rollouts to a single forward pass.

Contrasts and Bold Claims

• Contradicts the prevailing belief that scaling parameters and data alone suffice for arbitrarily improving generative models; scaling generative expressivity (via explorative training) is necessary to avoid bottlenecks and sample-quality degradation as models and datasets scale.
• Challenges the status quo of factorized-generation-centric protocols (autoregressive, diffusion, etc.) by demonstrating that training-time factorization, via exploration, alone can make generative modeling end-to-end and more generalizable, with no need for recurrent self-conditioning or rollouts.

Practical and Theoretical Implications

The introduction of explorative modeling has direct implications for the design and scaling of generative architectures:

• Training efficiency: With large-scale exploration, training steps become more costly, but the method is FLOP-optimal at scale, ultimately saving computational resources per effective model improvement.
• Model design flexibility: Factors such as the number of generative process steps (e.g., in jumpy models) become tunable alongside exploration, making how “end-to-end” a model is into a parameter that scales with resources.
• Improved generalization: Surplus generative expressivity reduces overfitting, which is particularly beneficial in data-constrained regimes.
• Applicability: The approach is shown to benefit both continuous domains (vision) and discrete domains (language), suggesting wide relevance across modalities.

Future developments are likely to focus on more scalable variants (gradient-based latent search, hybrid forward/reverse XM, memory-efficient approximations), application to sequence models (autoregressive LLMs, transformers), and integration with compositional and energy-based generative frameworks.

Conclusion

Explorative Modeling reframes generative modeling by introducing exploration as a third axis of pretraining, orthogonal to parameters and data. This axis provides control over generative expressivity, enabling scalable, end-to-end training for a broad range of domains. Strong empirical efficiency and performance results, coupled with theoretical clarity, position this paradigm as a fundamental advancement in generative model pretraining, with significant implications for model scaling, generalization, and downstream deployment (2607.27372).