How Transparent Is DiffusionGemma? Comparing Reasoning Transparency in Diffusion vs. Autoregressive LLMs

Paper Overview

• Field: Machine Learning
• Authors: Joshua Engels, Callum McDougall, Bilal Chughtai
• Published: 2025-06-20
• arXiv: 2506.16807

Summary

LLM reasoning transparency is a critical affordance for understanding model decisions, mitigating misuse and misalignment, and debugging surprising model behaviors. However, DiffusionGemma performs a larger fraction of its computation in a continuous latent space; does this make its reasoning less transparent?

The authors decompose transparency into two components:

• Variable transparency: whether we understand intermediate snapshots of a model's computational state.
• Algorithmic transparency: whether we can use these snapshots to reconstruct the process by which the model arrived at its outputs.

Key Findings

• Naively, DiffusionGemma has poor variable transparency: its opaque serial depth (the amount of serial computation between interpretable model states) appears to be 28.6x higher than autoregressive Gemma 4.
• However, the information flowing between denoising steps can be mapped through an interpretable token bottleneck without degrading downstream performance. Treating these intermediate states as interpretable reduces the opaque serial depth to only 1.1x that of Gemma 4.
• Algorithmic transparency is harder for diffusion models than autoregressive ones, because all token predictions in the canvas can change at every denoising step, allowing the model to implement complex distributed algorithms during denoising.
• A series of interpretability case studies found preliminary evidence of diffusion-specific phenomena, such as non-chronological reasoning, token and sequence smearing, and intermediate context reasoning.
• Finally, testing monitorability showed that DiffusionGemma and Gemma 4 have comparable monitorability.

Original Abstract

LLM reasoning transparency is a critical affordance for understanding model decisions, mitigating misuse and misalignment, and debugging surprising model behaviors. However, DiffusionGemma performs a larger fraction of its computation in a continuous latent space; does this make its reasoning less transparent? We study this question by decomposing transparency into two components: variable transparency, whether we understand intermediate snapshots of a model's computational state; and algorithmic transparency, whether we can use these snapshots to reconstruct the process by which the model arrived at its outputs. Naively, DiffusionGemma has poor variable transparency: its opaque serial depth, the amount of serial computation that occurs in between interpretable model states, seems at first...

*Auto-collected on 2026-06-20*