Painting in the Fog: How Transparent Is DiffusionGemma's Reasoning?
> "If you want to understand something complex, simplify it until you can understand it; then complexify it back." — Richard Feynman
This post is an in-depth Chinese-language walkthrough of a recent interpretability paper: *How Transparent is DiffusionGemma?* (Engels et al., arXiv:2606.20560), exploring whether the reasoning processes of diffusion-based language models can be made as understandable as those of autoregressive models.
The Problem: Reasoning Transparency
When ChatGPT writes a poem or Midjourney paints a picture, we see the result but not the process. The post frames this as the reasoning transparency problem, split into two levels:
- Variable transparency — can we take readable "snapshots" of the model's intermediate states? Autoregressive models (like GPT-4 or Gemma) generate one token at a time, so their state is a readable text sequence. Diffusion models (like Stable Diffusion or DiffusionGemma) repeatedly denoise vectors in a high-dimensional continuous latent space, producing intermediate states that are just opaque clouds of numbers.
- Algorithmic transparency — even with snapshots, can we infer the reasoning that connects them? Diffusion models modify the *entire* output at every step, reasoning globally rather than linearly, which makes this far harder.
| Dimension | Autoregressive (GPT) | Diffusion | |---|---|---| | Generation | One token at a time | Iterative denoising | | Space | Discrete tokens | Continuous latent space | | Intermediate state | Readable text | Unreadable vectors | | Reasoning | Linear, one-directional | Global, multi-step iterative | | Transparency | Relatively easy | Very hard |
Measuring Opacity: Opaque Serial Depth
The paper quantifies opacity with Opaque Serial Depth: the amount of serial computation between interpretable states. For Gemma 4, this is one step (every token is an interpretable state). For raw diffusion models, it is 28.6x higher—a long stretch of computation during which nothing is human-readable.
Finding 1: A Token Bottleneck Lights the Tunnel
By forcing information flow between denoising steps through an interpretable token bottleneck—a small set of discrete, interpretable "concept tokens" analogous to text tokens—the researchers compressed the opaque serial depth from 28.6x down to 1.1x, nearly matching Gemma 4.
This is a deep result: the diffusion model's apparent opacity largely comes from its *representation* (continuous vectors vs. discrete tokens), not from intrinsic incomprehensibility. The information each step actually needs can be expressed in interpretable units.
Finding 2: Algorithmic Transparency Remains Harder
Diffusion models can globally rewrite the whole output at each step, enabling complex distributed algorithms. Case studies reveal three diffusion-specific phenomena:
1. Non-chronological reasoning — global, high-level features ("this is a landscape," "subject centered") may be decided early, with details filled in later, rather than reasoning left-to-right. 2. Token and sequence smearing — a concept like "red" is not localized in one token but smeared across multiple tokens and denoising steps, like watercolor bleeding on wet paper. 3. Intermediate-context reasoning — the model may temporarily entertain features (e.g., a shadow) that later vanish from the output, yet still influence other parts of the final result. Some reasoning happens in contexts that no longer exist, so even full access to intermediate states may not explain the output.
Finding 3: Monitorability Survives
The practical payoff: can internal states be used for downstream monitorability—detecting lies, bias, or jailbreaks? Tested on DiffusionGemma vs. Gemma 4, the results are very similar. Even though the reasoning is more complex, diffusion models' internal states are equally useful for supervision and control.
Core Conclusions
> Diffusion models' reasoning transparency is misleadingly underestimated by their representation. Once re-expressed via an interpretable token bottleneck, their comprehensibility approaches that of autoregressive models.
1. Variable transparency can be dramatically improved (28.6x → 1.1x opaque serial depth). 2. Algorithmic transparency remains challenging due to global edits, non-chronological reasoning, concept smearing, and intermediate-context reasoning. 3. Monitorability is not harmed—internal states remain as effective for downstream oversight.
Why It Matters
Trust requires understanding. If we can monitor a diffusion model's reasoning, we can intervene before harmful content is generated, ask meaningful questions about generated art ("why these colors?"), and design architectures that exploit the global, parallel nature of diffusion reasoning. Open questions include what the bottleneck discards, the exact mechanisms of non-chronological reasoning, and how vanished intermediate contexts affect output safety.
References
1. Engels, J., McDougall, C., Chughtai, B., et al. "How Transparent is DiffusionGemma?" arXiv:2606.20560, 2026. 2. Chen, T., et al. "Diffusion Models: A Comprehensive Survey of Methods and Applications." ACM Computing Surveys, 2024. 3. Meng, K., et al. "Locating and Editing Factual Associations in GPT." NeurIPS, 2022. 4. Nanda, N., et al. "A Circuit Tracing Toolkit for Mechanistic Interpretability." Distill, 2023. 5. Ho, J., et al. "Denoising Diffusion Probabilistic Models." NeurIPS, 2020. 6. Rombach, R., et al. "High-Resolution Image Synthesis with Latent Diffusion Models." CVPR, 2022.
*xiaokai daily paper recommendation | 2026-06-21*