Sparse autoencoders (SAEs) improve interpretability in multimodal models, but it remains unclear whether SAE features form modular, composable units for reasoning-an assumption underlying many intervention-based steering methods. We test this modularity hypothesis and find it often fails: intervening on a task-selective feature set can modestly improve reasoning accuracy, while intervening on the union of two such sets reliably induces output drift (large unintended changes in predictions) and degrades accuracy, even under norm-matched perturbations. This non modular circuit interference is consistent with shared internal pathways where feature unions amplify activation shifts. We develop a reproducible causal pipeline to localize and test these sparse visual thought circuits in Qwen3-VL-8B. On a controlled synthetic benchmark with seven task types and three difficulty levels, linear probes identify a mid decoder locus for task type information. We train SAEs at this layer, construct task-selective sets via an explicit rule, and perform inference time scaling and ablation while quantifying accuracy and drift. Our findings-validated with bootstrapped subsamples and permutation controls, and replicated across multiple VLM families and five diverse datasets clarify the boundaries of SAE feature composability and provide a rigorous diagnostic framework for more reliable VLM control.
Figures & tables
Figure 1: Sparse Visual Thought Circuits (SVTC) in a frozen vision–language model. We (1) localize reasoning-related signals by probing pooled hidden states across layer groups for task_type (7-way) and difficulty (3-way), selecting a predictive site (e.g., decoder Mmiddle_2 ); (2) train a sparse autoencoder at that site to map activations h∈R4096 to sparse codes z∈R32768 and construct task-selective feature sets (Pattern, Global, and their union) via rule-based selectivity; and (3) perform inference-time interventions by modifying selected coordinates of z (scaling/ablation), decoding a reconstruction, and applying the resulting update to obtain h′ . We report accuracy change ΔAcc (pp), output drift Changed (%), and perturbation magnitude E[∥Δ∥/∥h∥] , with negative controls (random and permutation sets), bootstrap over subsampling seeds, and held-out validation on val/test.
Model
Pattern @ s=2.0
Union @ s=0.5
Base
Δ pp
Chg(%)
Δ pp
Chg(%)
Train Set ( n=600 per seed)
Qwen3-VL-8B
0.712
+1.03
10.97
-1.67
14.73
InternVL3.5-8B
0.676
-1.30
13.89
-1.83
17.43
Phi-4-MM
0.609
+0.62
19.85
-2.15
21.78
Gemma-3n-E4B
0.633
+0.63
22.68
-2.65
23.77
Table 1: Cross-model intervention summary across splits. We report base accuracy (Base), intervention effect Δ Acc in percentage points ( Δ pp), and output change rate (Chg%) for Qwen3-VL-8B and baseline models. All numbers are pooled over 5 seeds.
Dataset
N
Base
Pattern ( s=2.0 )
Union ( s=0.5 )
Δ pp
Chg(%)
Δ pp
Chg(%)
Synthetic
1500
0.706
+0.22
10.44
-2.33
14.44
NLVR2
1500
0.895
+2.54
4.83
-6.29
6.72
VSR
1500
0.828
+1.37
11.50
-4.53
14.17
SNLI-VE
1500
0.407
-0.33
5.35
-3.18
11.31
CLEVR
1500
0.532
+1.53
21.50
-2.68
22.36
Table 2: Cross-dataset generalization on Qwen3-VL-8B. SAE weights and feature sets are fixed from Synthetic and applied verbatim to external benchmarks.
Figure 2: The Compositional Paradox: From Circuit Discovery to Mechanical Failure. (A) Circuit Localization: Layerwise probing reveals a clear phase transition: visual reasoning capability ( l⋆ ) emerges and peaks at Layer 21 (99.2% accuracy), distinct from the early Perception stage and the late Output saturation. (B) The Paradox: Across five benchmarks, while individual Pattern steering (circles) generally yields positive gains, the joint Union intervention (squares) consistently drives the model into the “Collapse Regime” (shaded pink), characterized by high output drift and accuracy degradation. (C) Mechanism Validation: Fine-grained analysis on the hypersensitive Layer 10 reveals a non-monotonic U-shaped recovery. Weak signals ( s<0.5 ) act as adversarial noise causing collapse, confirming that a critical signal-to-noise threshold must be overcome for effective steering.
Figure 3: Mechanism of Compositional Collapse. (A) Geometric Antagonism: Although individual Pattern and Global vectors are strong, their antagonistic alignment ( θ>90∘ ) causes the Union vector to shrink significantly inside the unit circle (Signal Collapse). (B) Noise Amplification: This reduced signal magnitude ∥δ∥ pushes the system into the hypersensitive “Collapse Regime,” where the layer normalization operator forces an exponential amplification of latent noise.
Config ( N=600 )
Δ pp
Chg(%)
Pattern (perm, s=2.0 )
-0.77 ± 1.0
15.41 ± 1.8
Union (perm, s=0.5 )
-0.67 ± 0.5
14.73 ± 1.8
Table 3: Robustness controls. (Top) Permutation control rules out sampling noise. (Bottom) Subsample-size N confirms stability of the intervention effect.
Low–Mid Scale
Mid–High Scale
s
Δ PP
Drift (%)
s
Δ PP
Drift (%)
0.20
-25.5 ± 3.8
48.1 ± 2.9
0.65
-15.8 ± 0.5
33.3 ± 1.1
0.33
-19.2 ± 2.8
39.1 ± 3.7
0.75
-16.1 ± 0.8
33.4 ± 2.0
0.35
-18.8 ± 2.2
38.9 ± 2.7
1.00
-16.0 ± 0.7
33.1 ± 0.8
0.40
-18.3 ± 1.4
37.0 ± 2.1
1.25
-17.2 ± 0.7
35.8 ± 1.5
0.45
-16.5 ± 0.9
34.8 ± 1.5
1.50
-17.8 ± 1.6
38.2 ± 2.6
Table 4: Non-monotonic Sensitivity at Layer 10. Results reveal a U-shaped recovery: weak intervention scales ( s<0.5 ) trigger noise-amplified collapse, while higher magnitudes ( s→2.0 ) override the noise threshold to restore performance. Mean ± std over 2 seeds.
Config
Pattern ( s=2.0 )
Union ( s=0.5 )
Δ pp
Chg(%)
Δ pp
Chg(%)
A. Subsample Size Robustness
N=200
+1.40 ± 0.7
13.00 ± 3.4
-0.10 ± 1.1
15.20 ± 3.2
N=400
+1.55 ± 0.6
10.65 ± 1.4
-1.55 ± 0.5
14.15 ± 1.0
N=600
+1.03 ± 0.7
10.97 ± 1.6
-1.67 ± 0.6
14.73 ± 1.9
N=800
+0.63 ± 0.3
10.55 ± 0.8
-1.38 ± 0.8
14.75 ± 0.9
Table 5: Comprehensive Control Experiments on Synthetic. (Top) Subsample Robustness: We vary the evaluation size N to test stability against sampling noise. The Union collapse signature persists across all sizes. (Bottom) Specificity Check: We apply random permutations to the feature indices. The null effect confirms that the observed phenomena are specific to the identified semantic circuits. Data pooled over 5 seeds.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1 : Spatial Grounding of Visual Thought. Activation heatmap of Feature #2398 from the Pattern Set. Despite being discovered via global pooling, the feature functions as a precise semantic detector, selectively activating (red) on the target visual objects while ignoring the background (blue).
Layer Group
Depth
Acc. (%)
Interpretation
Baselines
Chance Level
–
14.28
Random Guess
Shuffled Labels
All
14.31 ± 0.05
No Causal Signal
Qwen3-VL-8B Decoder Layers
Early (Embed)
L0–L5
34.05 ± 1.2
Perceptual Noise
Middle-1
L6–L11
58.60 ± 2.4
Appendix
Table A1 : Full Layer-wise Linear Probing Trajectory on Qwen3-VL-8B. We report the classification accuracy (mean ± std). Layer 21 represents the global maximum ( l⋆ ).
Sparse Autoencoders (SAEs) have shown promise for analyzing language models, but applying them to vision-language models (VLMs) often yields representations that are difficult to use as controllable cross-modal steering directions. We introduce the Joint Sparse Autoencoder (JSAE), which uses an explicit alignment constraint to jointly factorize sequence-pooled vision and language activations into shared, interpretable image/caption-level features. Applied to LLaVA, JSAE recovers cross-modal features for recognizable concepts (e.g., food and animals). Through bidirectional interventions (additive steering and suppression), we observe a layer-dependent asymmetry under our protocol: additive steering peaks at mid-to-late (pre-output) layers and weakens at both ends, whereas suppression scores remain within a comparable range across all probed layers within statistical noise. Experiments on three VLMs, namely LLaVA-v1.6-Mistral-7B, Llama3-LLaVA-8B, and the MoE-based Qwen3-VL-30B, show related layer-localized effects across architectures. Together, these results suggest that explicitly aligned sparse representations support more controllable intervention-based analysis of multimodal features, within an identifiable layer range, than the unconstrained alternatives tested here.
Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it ``sees'', effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.
Javier Ferrando, Enrique Lopez-Cuena, Pablo Agustin Martin-Torres +3
Vision-language models (VLMs) are powerful but remain opaque black boxes. We introduce the first framework for transparent circuit tracing in VLMs to systematically analyze multimodal reasoning. By utilizing transcoders, attribution graphs, and attention-based methods, we uncover how VLMs hierarchically integrate visual and semantic concepts. We reveal that distinct visual feature circuits can handle mathematical reasoning and support cross-modal associations. Validated through feature steering and circuit patching, our framework proves these circuits are causal and controllable, laying the groundwork for more explainable and reliable VLMs.
Jingcheng Yang, Tianhu Xiong, Shengyi Qian +2
University of Illinois Urbana-Champaign · Independent Researcher