Qwen-Scope: Sparse Autoencoder Toolkit

• Qwen-Scope is a mechanistic-interpretability suite that uses layer-wise sparse autoencoders to diagnose, control, and improve Qwen models.
• The system integrates 14 SAE groups across 7 Qwen variant backbones, supporting inference-time steering, evaluation analysis, and data-centric safety workflows.
• It leverages sparse feature activations as actionable control handles to enable targeted interventions like code-switch reduction and repetition suppression.

Searching arXiv for "Qwen-Scope" and closely related papers on Qwen SAEs/mechanistic interpretability. Qwen-Scope is an open-source suite of sparse autoencoders (SAEs) built on the Qwen model family and positioned as a representation-level interface for diagnosing, controlling, evaluating, and improving LLMs. It comprises 14 groups of SAEs across 7 model variants from the Qwen3 and Qwen3.5 series, spanning both dense and mixture-of-expert (MoE) architectures, and is designed to move SAEs beyond post-hoc interpretability into practical model-development workflows such as inference-time steering, evaluation analysis, data-centric safety pipelines, and post-training optimization (Deng et al., 12 May 2026).

1. Definition and placement within the Qwen ecosystem

Qwen-Scope is a mechanistic-interpretability system centered on layer-wise SAEs trained on residual-stream activations from Qwen backbones. Its stated objective is to make sparse, interpretable feature representations operational for model development rather than limiting them to descriptive analysis (Deng et al., 12 May 2026). In this sense, Qwen-Scope sits downstream of the broader Qwen program, which introduced Qwen as a family of base pretrained LLMs, aligned chat models, and specialized code and mathematics variants with strong tool-use and planning capabilities (Bai et al., 2023).

The released coverage is organized around seven backbones: five dense models and two MoE models. The dense variants are Qwen3-1.7B, Qwen3-8B, Qwen3.5-2B, Qwen3.5-9B, and Qwen3.5-27B (instruct). The MoE variants are Qwen3-30B-A3B and Qwen3.5-35B-A3B. The release covers every transformer layer in each backbone, yielding 14 SAE groups in total (Deng et al., 12 May 2026).

Backbone class	Model variants	SAE configuration
Dense	Qwen3-1.7B, Qwen3-8B, Qwen3.5-2B, Qwen3.5-9B, Qwen3.5-27B (instruct)	$W=16h\in\{32\mathrm{k},64\mathrm{k},80\mathrm{k}\}$, Top-$k\in\{50,100\}$
MoE	Qwen3-30B-A3B, Qwen3.5-35B-A3B	$W=32\mathrm{k}$ with Top-$k=50$, and $W=128\mathrm{k}$ with Top-$k=100$

The central claim of the project is not merely that sparse features are interpretable, but that they can serve as reusable interfaces across multiple downstream workflows. This suggests a shift from “interpretability as inspection” to “interpretability as tooling,” with SAE feature activations functioning as manipulable state variables tied to model behavior (Deng et al., 12 May 2026).

2. Sparse autoencoder design and training objective

Each Qwen-Scope SAE is a layer-wise encoder-decoder model trained on residual-stream activations. The encoder is a single linear map $E:\mathbb{R}^h\to\mathbb{R}^W$, followed by a Top-$k$ ReLU that preserves only the $k$ largest positive activations and zeroes the rest. The decoder is a linear map $D:\mathbb{R}^W\to\mathbb{R}^h$. For dense backbones with hidden size $k\in\{50,100\}$0, the width is $k\in\{50,100\}$1. For MoE backbones with hidden size $k\in\{50,100\}$2, Qwen-Scope releases both $k\in\{50,100\}$3 and $k\in\{50,100\}$4 variants (Deng et al., 12 May 2026).

The training objective is a reconstruction loss under an exact sparsity constraint:

$k\in\{50,100\}$5

Because Top-$k\in\{50,100\}$6 enforces $k\in\{50,100\}$7 exactly, no explicit $k\in\{50,100\}$8 penalty is added. To mitigate feature death, an auxiliary dead-feature loss $k\in\{50,100\}$9 is included with weight $W=32\mathrm{k}$0, producing the full objective

$W=32\mathrm{k}$1

During training, rare outlier activations with extremely large $W=32\mathrm{k}$2-norm are filtered out for stability (Deng et al., 12 May 2026).

The optimization setup is conventional for dictionary-learning-style sparse representation learning: AdamW, learning rate $W=32\mathrm{k}$3, batch size 1,024 activations per update, Top-$W=32\mathrm{k}$4 values 50 or 100, and training data drawn from raw residual-stream activations on the original Qwen pretraining corpus (Deng et al., 12 May 2026). The exact-sparsity design is important because many subsequent workflows depend on binary or thresholded notions of feature presence, not merely on dense reconstruction fidelity.

3. Inference-time steering and feature-level control

A major practical contribution of Qwen-Scope is inference-time steering through SAE feature directions. The method begins with a contrastive identification step: a positive set and negative set are collected, feature activations $W=32\mathrm{k}$5 are computed, and features are ranked by

$W=32\mathrm{k}$6

A direction $W=32\mathrm{k}$7 is then injected or suppressed during inference by modifying the residual representation:

$W=32\mathrm{k}$8

where $W=32\mathrm{k}$9 amplifies and $k=50$0 suppresses the feature (Deng et al., 12 May 2026).

The reported case studies are concrete. On Qwen3-8B, an unintended Chinese feature, identified as feature id 6159, becomes highly activated when an English prompt accidentally triggers Chinese generation; setting $k=50$1 on this feature removes the Chinese text while preserving fluency. A second case study uses a classical-Chinese feature, id 36398, whose activation shifts a modern Chinese story continuation into a classical literary register. These interventions require no weight updates and produce immediate output changes (Deng et al., 12 May 2026).

Qwen-Scope explicitly reports that ablations confirm minimal impact on general fluency and comparable next-token probabilities outside the steered directions, although perplexity numbers are not reported (Deng et al., 12 May 2026). This is significant because it frames SAE directions as local control handles rather than global model edits. A plausible implication is that the latent basis learned by the SAEs is sparse enough to support targeted interventions without inducing broad distributional drift.

4. Evaluation analysis through feature coverage and overlap

Qwen-Scope also introduces SAE-based proxies for benchmark redundancy and inter-benchmark similarity. For a benchmark $k=50$2 with $k=50$3 samples and a panel of $k=50$4 models, the paper defines a true model-ranking vector $k=50$5 on $k=50$6, a subset ranking $k=50$7, and the expected Kendall’s $k=50$8 agreement

$k=50$9

Because direct computation costs $W=128\mathrm{k}$0 forward passes, Qwen-Scope replaces it with a feature-coverage proxy. Let $W=128\mathrm{k}$1 denote the set of active features for sample $W=128\mathrm{k}$2, and let

$W=128\mathrm{k}$3

The feature-coverage AUC is

$W=128\mathrm{k}$4

and feature redundancy is defined as

$W=128\mathrm{k}$5

Across 17 benchmarks, including MMLU, GSM8K, MATH, and MBPP, the paper reports a Spearman correlation $W=128\mathrm{k}$6 between $W=128\mathrm{k}$7 and $W=128\mathrm{k}$8 (Deng et al., 12 May 2026).

For inter-benchmark similarity, the framework uses feature overlap. For benchmarks $W=128\mathrm{k}$9, the asymmetric overlap is

$k=100$0

and the symmetric variant is

$k=100$1

The direct correlation between $k=100$2 and the Pearson correlation of model score vectors is reported as 68.4% (Pearson) and 60.7% (Spearman). After partialling out MMLU as a proxy for general ability, Pearson rises to 75.5% and Spearman to 71.3% (Deng et al., 12 May 2026). The paper gives GSM8K being 63% subsumed by MATH as an example of how such overlap measures can guide benchmark suite design.

The significance of these results lies in replacing expensive cross-model evaluation with a feature-space proxy computed from internal activations. This suggests that benchmark redundancy can be studied as a coverage property of internal capability features rather than solely as a correlation among final task scores.

5. Data-centric workflows: toxicity detection and safety synthesis

Qwen-Scope develops two data-centric pipelines: multilingual toxicity classification and safety-oriented data synthesis. In the toxicity setting, SAEs on Qwen3-1.7B and Qwen3-8B are used to construct a fully transparent, rule-based detector over the TextDetox multilingual corpus, consisting of 13 languages with 5 k examples each and a 50% toxic / 50% clean split. The data are partitioned into 4 k discovery examples and 1 k test examples per language (Deng et al., 12 May 2026).

Feature discovery is performed at layer $k=100$3. For example $k=100$4, token $k=100$5, and feature $k=100$6, the activation tensor is $k=100$7, and feature presence is defined by

$k=100$8

The selection statistic is

$k=100$9

and the rule-based classifier is

$E:\mathbb{R}^h\to\mathbb{R}^W$0

The reported results are that the best single-layer F1 exceeds 0.90 with $E:\mathbb{R}^h\to\mathbb{R}^W$1, middle layers are most predictive, English-discovered features transfer with mean F1 approximately 0.7–0.8 to related languages such as Russian and French, and larger 8B models show stronger and more stable transfer. A simple proxy,

$E:\mathbb{R}^h\to\mathbb{R}^W$2

identifies the best classification layer without held-out evaluation; combining one feature per top-$E:\mathbb{R}^h\to\mathbb{R}^W$3 layers improves worst-case F1; and only 10% of discovery data recovers approximately 99% of full-data performance (Deng et al., 12 May 2026).

The safety-synthesis workflow uses WildJailbreak training prompts and responses as seed data. For each layer-feature pair $E:\mathbb{R}^h\to\mathbb{R}^W$4, feature coverage on the seed set is

$E:\mathbb{R}^h\to\mathbb{R}^W$5

Features are assigned semantic relevance scores $E:\mathbb{R}^h\to\mathbb{R}^W$6 from a judge model on their textual explanations; a target set $E:\mathbb{R}^h\to\mathbb{R}^W$7 is formed by thresholding these scores, and missing targets $E:\mathbb{R}^h\to\mathbb{R}^W$8 are identified by absent coverage. The synthesis pipeline then generates vanilla prompts, adversarial variants, and responses, retaining only examples whose activations actually trigger the desired target feature. Dataset-level coverage is measured by

$E:\mathbb{R}^h\to\mathbb{R}^W$9

Feature-driven synthesis reaches 99.7% of $k$0 under a small budget, whereas natural sampling and random safety synthesis leave large gaps (Deng et al., 12 May 2026).

Downstream supervised fine-tuning on Qwen3-8B combines Alpaca (50 k) with real safety data at budgets of 8 k, 40 k, 120 k, or 200 k, optionally plus 4 k synthetic safety examples. Using only 4 k real plus 4 k feature-driven synthetic data yields safety accuracy 77.75%, compared with 71.75% using 8 k real only, while preserving or improving general benchmarks such as IFEval, TruthfulQA, MMLU, GSM8K, and BBH (Deng et al., 12 May 2026).

6. Post-training optimization, empirical effects, and nomenclature

Qwen-Scope extends SAE signals into post-training optimization in two forms: supervised fine-tuning to reduce code-switching and reinforcement learning with rare-negative augmentation to suppress repetition. For code-switching, the failure metric is the unexpected code-switching ratio

$k$1

Language-specific features $k$2 are identified using the “monolinguality” score

$k$3

where $k$4 is the mean activation on language $k$5 and $k$6 is the mean activation on other languages. The auxiliary suppression loss is

$k$7

and the total training loss is

$k$8

On Qwen3-1.7B and Qwen3-8B, baseline code-switch ratios for any $k$9 zh/ru/ko are 0.81%/0.19%/0.36% for 1.7B and 0.96%/0.16%/0.43% for 8B. On the 210 k setting, SAE-guided SFT reduces zh to 0.22% ($k$0), ru to 0.03% ($k$1), and ko to 0.00% ($k$2), with similar gains at 110 k and no degradation on multilingual benchmarks such as MMLU, Flores, and HellaSwag (Deng et al., 12 May 2026).

For repetition, Qwen-Scope identifies repetition features by contrasting activations on a token’s first occurrence versus repeated occurrences, then verifies causality by steering:

$k$3

This signal is inserted into a DAPO reinforcement-learning loop by sampling $k$4 normal rollouts and one steered negative rollout per prompt. On Qwen3-1.7B, 8B, and 30B, vanilla RL yields only modest repetition reduction, whereas SAE-guided RL rapidly drives the repetition ratio to near zero; general-capability benchmarks remain competitive, with mixed task-dependent gains and regressions (Deng et al., 12 May 2026).

The paper’s broader conclusion is that a single SAE suite can support steering, evaluation, data construction, and post-training optimization within one representational framework (Deng et al., 12 May 2026). A common misconception is that “Qwen-Scope” refers to one of the unrelated 2026 papers titled “SCOPE,” such as “SCOPE: Selective Conformal Optimized Pairwise LLM Judging” or “SCOPE: Real-Time Natural Language Camera Agent at the Edge.” Those works concern pairwise LLM judging and PTZ camera agents, respectively, and are distinct from Qwen-Scope despite the shared acronymal surface form (Badshah et al., 13 Feb 2026, Hindsbo et al., 1 Jun 2026).

Within the Qwen research trajectory, Qwen-Scope therefore occupies a specific niche: it is not a new base model family, nor an alignment stack, nor an agent architecture, but a mechanistic-interpretability and model-development toolkit built on top of Qwen backbones. Its significance lies in operationalizing sparse internal features as reusable control and analysis primitives, thereby connecting model internals to observable downstream behavior in a way that is both technically explicit and directly actionable (Deng et al., 12 May 2026).