Technical Summary: Concurrence of Symmetry Breaking and Nonlocality Phase Transitions in Diffusion Models

Problem Statement

Diffusion models generate high-quality data, yet the precise temporal dynamics of when specific capabilities (such as semantic class selection or global context integration) activate during the generation trajectory remain poorly understood. Two distinct theoretical frameworks have been proposed to describe critical moments in diffusion dynamics:

1. Symmetry Breaking: Rooted in statistical mechanics, this view posits that generation involves a phase transition where trajectories bifurcate from a single high-entropy basin into multiple semantic minima (e.g., distinguishing a dog from a cat). This suggests a "critical window" where conditioning is essential to select a specific semantic class.
2. Nonlocality: Derived from many-body physics and quantum error correction, this view focuses on locality. It defines a phase transition as the point where local denoising fails, requiring global information (a diverging "Markov length") to recover a patch. This suggests a critical window where nonlocal computation is necessary.

While both frameworks predict a critical time window, it is unclear whether these two notions of phase transitions are concurrent in modern, high-performance diffusion transformers (DiTs). Specifically, does the moment a model requires semantic forcing (symmetry breaking) align with the moment it requires global context (nonlocality)?

Methodology

The authors investigate this concurrence using two modern diffusion transformer systems: a class-conditional Facebook DiT-XL (trained on ImageNet) and a text-conditioned SD3 Medium (MMDiT). They employ two complementary families of probes to measure critical times:

1. Instantaneous Probes (Score-Level Analysis)

These probes measure the behavior of the score function $s_\theta(x_t, t)$ along the trajectory without integrating over time.

• Conditioning Gap: Defined as the $L_2$ norm between the conditional score $s_\theta(x_t, t | y)$ and the unconditional score $s_\theta(x_t, t)$. A large gap indicates that the conditioning signal significantly alters the denoising vector, implying a symmetry-breaking event.
• Locality Gap: To measure nonlocality, the authors construct a "local denoiser" by truncating the attention mechanism of the pre-trained model to a local window of radius $r$ (Figure 2). The locality gap is the $L_2$ norm between the global score and this local approximation. A large gap indicates that local context is insufficient for accurate denoising, implying a nonlocality phase transition.

2. Integrated Probes (Sample-Level Analysis)

These probes evaluate the impact of interventions on the final generated samples.

• Forward-Backward Experiments: Starting from a clean image, noise is added up to time $t$, and the image is then denoised unconditionally. If the unconditional denoiser recovers a different semantic class, $t$ has crossed the symmetry-breaking threshold. The authors also test local denoisers in this setup to find the threshold for nonlocality.
• Windowed Conditioning/Local Denoising: The authors intervene in the sampling process by applying conditioning or global denoising only within specific time windows $[t_i, t_i + \Delta]$. They measure classification error and Fréchet Inception Distance (FID) to determine which time windows are critical for maintaining semantic correctness and image quality.

Key Results

Concurrence in Instantaneous Probes

• Facebook DiT-XL: Both the conditioning gap and the locality gap peak in an early window centered around $t \approx 0.2$. This indicates that the model simultaneously requires strong semantic guidance and global context at this specific time.
• SD3 Medium: Both gaps concentrate very close to the noisy endpoint ($t \approx 1$).
• Alignment: Despite being functionally independent (one compares conditional vs. unconditional, the other global vs. local), the two gaps align remarkably well in both models, suggesting the two phase transitions occur simultaneously.

Concurrence in Integrated Probes

• Facebook DiT-XL: The forward-backward experiment and windowed interventions identify a critical window for both symmetry breaking and nonlocality near $t \approx 0.5$.
  • Observation of Suboptimality: The instantaneous probes identified the critical window at $t \approx 0.2$, while integrated probes found it at $t \approx 0.5$. The authors interpret this as the model performing unnecessary conditioning and nonlocal computation earlier than strictly required by the data.
• SD3 Medium: Both instantaneous and integrated probes show a concurrent transition near $t \approx 1$. This suggests the higher-performance model is more efficient, aligning its computational needs with the actual critical window.

Windowed Interventions

• Applying conditioning or global denoising only during the identified critical windows (e.g., $[0.2, 0.7]$ for DiT-XL) yields recognizable, high-quality samples.
• Conversely, removing these signals during the critical window (even if applied elsewhere) results in severe semantic degradation (e.g., generating a "dog" with the wrong class or incoherent global structure), while local textures may remain plausible.

Key Contributions

1. Unification of Phase Transition Notions: This is the first work to empirically unify the symmetry-breaking and nonlocality perspectives in practice, demonstrating that the time a model needs semantic forcing aligns with the time it needs global computation.
2. Operational Diagnostics: The paper provides concrete, measurable diagnostics (score gaps and windowed sampling) to identify when and why diffusion models rely on conditioning and global attention.
3. Efficiency Insights: By revealing that some models (like DiT-XL) compute global information earlier than necessary, the work highlights potential inefficiencies in current architectures.

Significance and Claims

The authors claim that their work provides a principled route to improving diffusion models by reducing complex generation behavior to time-dependent phase diagrams. The significance lies in:

• Architectural Design: Motivating designs that explicitly gate global communication (e.g., cross-token attention) to activate only during the critical window, rather than assuming global attention is needed throughout.
• Training and Sampling: Suggesting that losses or noise schedules could be reweighted to concentrate learning signals near the critical interval. It also supports adaptive sampling schemes that switch between local and global denoisers based on the identified critical time.
• Theoretical Bridge: Connecting the "symmetry breaking" literature (often focused on class selection) with the "nonlocality/Markov length" literature (focused on recoverability), showing they describe the same underlying operational reality in modern transformers.

The paper remains modest, noting that these findings are currently empirical and limited to two specific transformer architectures. It acknowledges that constructing strictly local denoisers is challenging and that the width of the critical window is difficult to quantify without a unified thermodynamic limit theory. However, the observed concurrence offers a robust operational framework for analyzing and optimizing diffusion dynamics.