Daily Digest — August 6, 2026

Applying uniform activation steering to a diffusion LM at every denoising step degrades quality because different attributes commit at fundamentally different points — topic within the first 2% of steps, sentiment over the first 20%. SAEs trained on four DLMs (124M–8B parameters) map these schedules; an adaptive framework that concentrates intervention in each attribute's commitment window achieves strong multi-attribute control at substantially lower perplexity and inter-attribute interference than uniform or non-SAE baselines. This is the first mechanistically informed steering approach designed specifically for diffusion generation.

Standard SAEs implicitly assume each learned feature is one-dimensional. SASA proves this assumption provably induces feature splitting via two distinct mechanisms when the true feature occupies a multi-dimensional subspace. Replacing single-vector decoders with learned subspaces, enforcing Top-s block sparsity, and adapting rank with a nuclear-norm regularizer eliminates these mechanisms on GPT-2 Small and Mistral-7B, improving monosemanticity while training on roughly half the token budget.

Masked DLMs implement a bidirectional induction circuit — previous-token and next-token heads cooperate to write local context into the residual stream, and a later induction head uses both to find and copy from matching source positions anywhere in the sequence, whether the source appeared before or after the query. This is mechanistically impossible in autoregressive models and represents a fundamental architectural difference in how DLMs implement in-context learning, with implications for understanding generalization in bidirectional generation.

Extends transcoder-based circuit tracing — a powerful mech-interp tool for autoregressive LLMs — to multimodal diffusion transformers (FLUX.1[schnell]), training timestep-conditioned transcoders for MLP sublayers in the double-stream MM-DiT architecture. Linearizing the remaining computation yields exact feature-to-feature attribution across text and image streams; the resulting circuits reveal attribute binding mechanisms and cross-stream semantic propagation. Circuit-guided steering interventions are substantially more precise and effective than SAE-based steering.

The first SAE-based interpretability framework for diffusion language models: Top-K SAEs trained on DLM residual streams faithfully extract interpretable features (verified as monosemantic and stable across pre-training and SFT phases) and — uniquely — inserting the SAE into the DLM forward pass does not incur the loss penalty observed in autoregressive LLMs. Establishes the foundational toolkit that Steering Without Breaking (#4) and the broader mech-interp-of-dLLMs research program extends.

SHADOWMASK is the first backdoor attack designed specifically for masked diffusion LMs, modifying the forward corruption process by replacing the all-mask terminal with a trigger-mask mixture prior that creates a dedicated denoising pathway from trigger-corrupted states to an adversarial target sequence. On clean inputs, behavior is fully preserved and indistinguishable from an unattacked MDLM. Backdoor techniques from continuous Gaussian diffusion or autoregressive models do not transfer because MDLM discrete state corruption has no continuous-domain analogue.

Mode connectivity analysis — testing whether two model checkpoints are linearly interpolable through a low-loss path — reveals that naive output-suppression unlearning methods preserve connectivity between the original and unlearned model, meaning "forgotten" knowledge remains accessible via interpolation. Circuit-aware methods that target specific computational pathways introduce a high-loss barrier between the two checkpoints, producing more durable erasure. This provides a mechanistic interpretability account of unlearning robustness that complements capability-based evaluations.