A Better Newspaper

## Mechanistic Interpretability: SAE-Based Feature Control and Steering **Status:** Developing **Signal Strength:** 5 **Summary:** Research continues into using internal LLM and diffusion-model representations for interpretation and control via sparse autoencoders (SAEs) and activation steering. New work extends SAEs to diffusion models, dense retrieval, neuro-symbolic matching, cross-layer feature matching via optimal transport, auditable AI-text detection, text-to-MIDI music generation, and concept erasure in diffusion models, while other studies probe the causal faithfulness of attribution methods, the geometric fragility of steering vectors, and whether standard SAE evaluation metrics mask hidden polysemanticity and redundancy. **Key Developments:** - One paper proposes a Sparse Autoencoder-based framework for retrieving and steering high-order semantic features tied to complex behavior-level attributes, using a contrastive feature method to link internal features to reliable control of language generation (arXiv:2601.02978, October 14). - A related comparative study examines whether an LLM's hidden-state geometry can be recovered from behavioral data in psycholinguistic experiments, running similarity-based forced-choice and free-association tasks across eight instruction-tuned models and 17.5M+ trials, then comparing behavioral geometry to hidden-state similarity via representational similarity analysis (arXiv:2602.00628, October 14). - A new paper, SharedSAE, extends sparse autoencoder methods from language models to diffusion models, proposing one shared sparse dictionary across four U-Net attention blocks and all 50 denoising steps of SDXL 1.0 (rather than separate per-block/per-timestep dictionaries), using a block-specific low-rank input adapter to align residual distributions before a shared encoder (arXiv:2610.13997, October 14). - A separate study evaluates whether SHAP-based attribution, applied to SAE latent features, faithfully reflects actual causal effects, noting that SAE feature interpretation has remained "predominantly correlational"; the study compares SHAP attribution against direct causal intervention using sentiment classification (SST-2) with GPT-2 Small and a 32,768-feature layer-7 SAE (arXiv:2610.17898, October 14). - A third new paper examines why contrastive activation addition (CAA), the standard activation-steering construction, is "fragile" and "on many inputs shifts behavior in the wrong direction"; the authors introduce a "steering Hessian" — the matrix of second-order derivatives of behavioral loss with respect to the steering vector — to analyze the curvature/geometry underlying this fragility (arXiv:2610.16084, October 14). - A fourth new paper proposes using SAEs as a "shared coordinate system" to enable symbolic matching for building neuro-symbolic models at LLM scale, aiming to address the bottleneck of human-constructed symbolic backbones and the domain-specificity of existing neuro-symbolic approaches via a data-driven method (arXiv:2610.19636, October 15). - A fifth new paper presents a pipeline converting a fine-tuned neural AI-text detector into a classifier operating on human-readable concepts via SAEs, aiming to make detector decisions auditable for high-stakes deployments such as peer review, education, admissions, journalism, and content moderation, addressing the gap between opaque high-performing neural detectors and underperforming transparent stylometric alternatives (arXiv:2610.30006, October 22). - A sixth paper, on semantic optimal transport for SAE feature matching and circuit compression, argues that two previously separate, hard-to-scale SAE analysis problems — matching semantically similar features across layers and compressing large feature circuits into interpretable "supernodes" — are both instances of a more fundamental problem: estimating semantic distances between SAE features that lie on different activation manifolds; the paper proposes an optimal-transport-based method to address both tasks (arXiv:2605.28567, October 23, replace-cross version). - A seventh paper, Xetrieval, extends SAE-based explanation beyond language generation to dense retrieval, introducing "Feature Elicitors" to decompose dense retrieval embeddings into sparse semantic features and address the problem that the document semantics underlying a retriever's relevance scores "may not be sufficiently explicit in the embeddings to be recovered as interpretable SAE features" (arXiv:2605.29507, October 23, replace version). - An eighth paper challenges standard SAE evaluation practice itself, arguing that widely used global metrics (sparsity, dead feature rates, reconstruction fidelity, disentanglement) do not specify what a learned feature set should look like as a whole, and proposes an "Atomic Interpretable Unit" framework that reportedly reveals hidden polysemanticity and redundancy obscured by current metrics (arXiv:2610.30474, October 23). - A ninth paper, "MI-MIDI 2," extends the SAE interpretation-vs-control question to music generation, noting that in language models "the evidence used to interpret these features has proven an unreliable guide to which of them steer the output"; the authors develop an end-to-end protocol to discover and evaluate SAE features in text-to-MIDI generation, applying it to the text2midi and MIDI-LLM models and characterizing features through correlations (arXiv:2610.32007, October 26). - A tenth paper, "Concept Acupuncturist" (ConceptAC), proposes an SAE-based concept erasure method for diffusion models aimed at machine unlearning of harmful or policy-violating content, addressing the "distributed and entangled nature of concept representations" that reportedly causes existing unlearning approaches to be inefficient and imprecise when applied via full-trajectory fine-tuning (arXiv:2610.32059, October 26). **Key Entities:** Sparse autoencoders (SAEs), SharedSAE, SDXL 1.0 (diffusion model), SHAP attribution, GPT-2 Small, contrastive activation addition (CAA), representational similarity analysis, neuro-symbolic models, AI-generated text detectors, optimal transport (feature matching/circuit compression), Xetrieval (dense retrieval explanation), Atomic Interpretable Unit framework, text2midi, MIDI-LLM, Concept Acupuncturist (ConceptAC), machine unlearning **Implications:** These findings suggest interpretability tooling is both broadening (from language models to diffusion models, dense retrieval, neuro-symbolic reasoning, cross-layer feature alignment, music generation, and applied auditing of AI-text detectors and content safety) and facing increased scrutiny of its causal and evaluative validity — correlational feature interpretations, steering techniques, and even the standard metrics used to judge SAE quality are now being formally stress-tested for faithfulness, robustness, and completeness. The recurring finding that SAE feature interpretations unreliably predict actual causal control (now replicated in music generation as well as language) is a persistent methodological concern for the field. This matters for AI safety and alignment claims built on these tools, for real-world accountability in deployed detection, retrieval, and content-moderation systems, and for how the field benchmarks interpretability progress going forward. **Connections:** Mechanistic interpretability, AI safety/alignment, diffusion models, activation steering, sparse autoencoders, LLM behavioral control, neuro-symbolic AI, AI-generated text detection, circuit analysis/compression, dense retrieval/search, machine unlearning, generative music models **Section Relevance:** ai-tech: 0.8, science: 0.4