Qwen3.6-35B-A3B Feature Circuits
Paper: "Feature-Level Circuits in Hybrid MoE+GDN Reasoning Models: A Multi-Substrate Negative Study with Methodological Post-mortem"
First SAE-based mechanistic interpretability investigation of Qwen3.6-35B-A3B (hybrid MoE + Gated Delta Network + Gated Attention, 40 layers, 256 experts, 9 active per token).
Summary
We train TopK Sparse Autoencoders (n_features=4096, k=32) on four distinct computational substrates at layers L11/L17/L23:
- Residual stream
- MoE sub-block output
- Gated-Attention sub-block output (L11, L23)
- Gated-Delta-Network sub-block output (L17)
Two consistent negative findings and one methodological post-mortem:
1. Residual stream: null cross-layer correlation
Max |Pearson| = 0.009 across approximately 3 × 10⁶ feature pairs. No detectable cross-layer feature dependencies.
2. Sub-block outputs: high correlation, near-zero causality
Cross-layer correlations rise to 0.70–0.77, but single-feature causal ablation (subtract z_i · W_dec[i] from sub-block output) systematically contradicts correlation-implied circuits:
- Pearson(correlation, ATE) = −0.22
- mean |ATE| = 0.0005
- full sub-block ablation: ~50% relative norm change downstream, but distributed across 89% of features at L23 — top-10 carry only 1.1% of total delta
3. Sign-consistent microcircuits at 100× smaller magnitude
Filtering by per-token sign consistency ≥ 0.80 recovers small mono-directional causal edges (|ATE| ≈ 0.003–0.006). Semantic characterization identifies them as task-format circuits, not reasoning:
- f1137 (L17): option-letter "G" detector silencing three L23 math-unit features (sign consistency 0.85–0.88)
- f1084 (L17): fires on " multiple" task-format word, drives antagonistic pair f3540/f4014 with sign consistency 1.00 — interpretable as math/content mode switch
- f4007 (L11): chat-start
<|im_start|> marker with causal propagation to L17 answer-prep feature
No sign-consistent hub fires on reasoning-specific vocabulary. This pattern persists across all three sub-block substrates tested.
Methodological post-mortem: correctness-predictor collapse
An initial correctness-conditional feature selection appeared promising:
| layer | biased AUROC (n=100) | clean AUROC (n=124) |
|---|
| L11 | 0.715 | 0.552 ± 0.051 |
| L17 | 0.605 | 0.565 ± 0.019 |
| L23 | 0.630 | 0.597 ± 0.040 |
Under proper stratified 5-fold train/test split with features re-selected on each training fold, AUROC dropped to near-chance (L11 95% CI [0.396, 0.685] includes 0.5). The "early-layer dominance" claim also reversed (L23 becomes best under clean protocol).
Root cause: top-10 features were selected using the same labels against which the classifier was evaluated. Cross-validation inside classifier training does not correct for feature-selection leakage that has already touched all labels. Magnitude of bias: ~15 pp AUROC inflation.
Recommendation: feature-predictor studies in SAE interpretability require 3-way split (selection / train / test) with label blinding at the respective stage. Pre-selection on the same corpus as evaluation systematically over-estimates predictive power.
Artifacts
phase_a/ — residual stream SAEs + cross-layer nullphase_b/ — MoE sub-block SAEs + correlation summary (max 0.71) + ATE control dataphase_c/ — single-feature ATE + full-sub-block ablation controls + SVD + microcircuit hub characterizationsphase_d/ — GA/GDN sub-block SAEs + correctness-predictor replication data (showing the collapse)
All training notebooks and replication scripts at
https://github.com/caiovicentinoReproducibility
- Stage B corpus:
caiovicentino1/Qwen3.6-35B-A3B-mcr-stage-b (624 rollouts on SuperGPQA 10-option MCQ, 50/50 balanced correctness)
- Compute: single Google Colab RTX PRO 6000 Blackwell (96 GB), ~4 h wall-clock total, ~$50
- SAE training: TopK with n=4096, k=32, 25 epochs, dead-feature revival, decoder row-norm constraint
- Var explained: 77–82% residual, 79–82% MoE sub-block, 57–74% GA/GDN sub-block
Limitations
- Single architecture (Qwen3.6-35B-A3B). Dense baseline comparison (Qwen2.5-32B on same task) not yet completed.
- Single SAE architecture (TopK); JumpReLU/Gated SAE may change results.
- Sample size for causal ATE: ~10–40 firing tokens per edge on 5-prompt runs. Aggregate claims hold; single-edge CIs are wide.
- Dead-feature rate: residual-stream SAEs produced ~75% dead features even with revival.
- SuperGPQA is dense with domain vocabulary; purely linguistic reasoning may behave differently.
Open directions
- GDN recurrent hidden state SAE (never tested — potential reasoning substrate)
- MoE router-logit circuits (256-dim expert selection per token, cross-layer trajectories)
- Medium-rank PCA probes on the 400–550-dim intervention subspace identified by SVD
- Aux-loss SAE retraining (OpenAI scaling-monosemanticity recipe) to address 75% dead feature rate
