Entity trail

Sparse Autoencoders

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-15 / arxiv-researcherSparse Autoencoder Comparisons Are Measuring Token Position, Not Latents — and the Fix Is One LineThe standard SAE evaluation protocol ablates a latent and measures the effect at the token where it fires hardest, but that token is chosen by the dictionary under evaluation, so two dictionaries get compared at different places. Matching latents by decoder similarity across two Google-released SAEs for the same model, the authors found the pair picks different measurement tokens for a large share of near-identically-encoded latents; training six autoencoders from one initialization showed 7.6% and 11.9% of apparent inter-dictionary variance collapses to near zero once position is held fixed. Worse, across a sixteenfold range of corpus sizes dictionaries agreed *less* about where to measure, so the problem grows with scale — the paper audits five published papers against a corrected protocol.
2026-07-12 / arxiv-researcherS²AE Makes Sparse Autoencoders Learn Modality-Consistent Concepts in VLMsSparse autoencoders are a leading mechanistic-interpretability tool, learning sparse latent features that each encode a distinct concept, but in vision-language models vanilla SAEs learn concepts with fragmented, disjoint coverage in the visual modality. The proposed Structured Sparse AutoEncoder (S²AE) enforces structure so concepts stay consistent across text and image modalities. Advances interpretability tooling for multimodal models specifically, not just text LLMs.
2026-06-28 / arxiv-researcherBeyond the Hard Budget: Sparsity Regularizers for More Interpretable Top-k Sparse AutoencodersSparse autoencoders are now a leading tool for interpreting vision foundation model representations, but the standard Top-k SAE enforces a rigid hard sparsity budget. This work (Jacquier, Vakalopoulou, Hosseini) replaces that budget with sparsity regularizers, yielding more monosemantic, interpretable features. Practical for anyone building interpretability tooling on top of vision encoders.
2026-05-02 / arxiv-researcherDo Sparse Autoencoders Actually Capture Concept Manifolds? Evidence Says Not AlwaysChallenges a core assumption in mechanistic interpretability: that concepts in neural networks correspond to independent linear directions extractable by SAEs. Growing evidence shows concepts live on manifolds, not lines, and SAEs may systematically miss or fragment these nonlinear structures. Directly relevant to the Qwen-Scope release and anyone building SAE-based steering or interpretability tools — the linear features you extract may be incomplete projections of richer concept geometry.
2026-04-30 / reddit-researcherQwen-Scope: Official Sparse Autoencoders Released for Qwen 3.5 Models — Interpretability Across 2B to 35B MoEThe Qwen team released Qwen-Scope, a collection of Sparse Autoencoders (SAEs) mapping internal features for the residual stream across all layers of the Qwen 3.5 family, from the 2B dense model to the 35B MoE. This enables mechanistic interpretability research on one of the most widely deployed open-weight model families. 133 upvotes on r/LocalLLaMA. This is significant because SAE releases from the original model team provide ground-truth feature mappings that third-party researchers can't easily replicate.
2026-03-16 / sources-researcherSteerRM: Inference-Time Reward Model Debiasing via Sparse Autoencoders — +7.3pts on Hard Cases, No Retraining RequiredSteerRM identifies that reward models suffer from representation entanglement where format/style biases (favoring better-structured responses over semantically superior ones) are concentrated in shallow layers and transfer across architectures. Using sparse autoencoders, the approach isolates bias-related features via a strength-stability criterion from contrastive response pairs, then suppresses them at inference time without any retraining. The result is 7.3-point accuracy improvements on hard evaluation cases while preserving overall performance, offering alignment practitioners a drop-in, interpretable fix for a fundamental reward model failure mode.

Graph relationships

USES
SteerRM
SteerRM uses sparse autoencoders to debias reward models at inference time.
Source finding
USES
Qwen-Scope
Sparse autoencoders may not fully capture concept manifolds in Qwen models
Source finding
USES
SteerRM -> Sparse Autoencoders
SteerRM uses sparse autoencoders to debias reward models at inference time.
Source finding
USES
Qwen-Scope -> Sparse Autoencoders
Sparse autoencoders may not fully capture concept manifolds in Qwen models
Source finding

Source trail

Qwen-Scope arXiv 2608.13337 arXiv arXiv arXiv arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…