Skip to content

Latest commit

 

History

History
77 lines (56 loc) · 6.38 KB

File metadata and controls

77 lines (56 loc) · 6.38 KB

Mechanistic Characterization of Censorship in Qwen3.5-2B: A Feature-Level Analysis of Semantic-Behavioral Dissociation

Date: May 16, 2026
Subject: Reverse-Engineering of Integrated Censorship and Narrative-Substitution Circuits
Model: Qwen3.5-2B-Instruct (Analysis extended with Qwen2.5-7B comparisons)


Abstract

We identify a fundamental dissociation between internal semantic encoding and downstream behavioral gating in the Qwen-series models. Utilizing Sparse Autoencoders (SAEs) from the Qwen-Scope suite, we demonstrate that censored prompts and their euphemistic counterparts (e.g., "June 4th" vs. "May 35th") converge to near-identical residual stream representations by Layer 23 (feature overlap: 9/10, activation identity: 100.00%). Crucially, the refusal mechanism (localized to Feature 27049, Layer 22) remains inactive during prompt processing across all conditions, localizing the censorship decision to the autoregressive generation phase. We propose a "Last-Mile Gating" architecture where censorship is conditioned on syntactical surface features rather than underlying semantic truth, enabled by differential early-layer pathway calibration. This dissociation proves that the model "knows" the truth it is forbidden to speak.


1. Experimental Methodology

We employed a multi-stage mechanistic probing framework:

  1. Feature Discovery: SAE-based scoring of activation drifts between neutral and sensitive contexts.
  2. Causal Intervention: Precision activation-clamping (clamping Feature 27049 to 0.0) to verify behavioral control.
  3. Representational Tracking: Layer-by-layer monitoring of top-K feature overlap between direct and euphemistic formulations.
  4. Logit Lens Analysis: Decoding the residual stream in late layers (L26-28) to identify latent factual representations.

2. Key Findings

2.1 Distributed Ensemble Detection (L8-L13)

The model initiates topic-sensitive detection via a distributed ensemble of specialized features in early layers. We identified functional specialization across three distinct axes:

  • Temporal Axis (Feature 13990): High-recall detection of year-specific patterns (1989).
  • Social Axis (Feature 25101): Sensitivity to protest and student-movement semantics.
  • Toponymic Axis (Feature 6057): Lexical "tripwire" specifically calibrated to the token 天安门 (Tiananmen).

Pathway Calibration: The "May 35th" bypass succeeds at the semantic level but fails to activate the generation-phase gate due to differential early-layer pathway calibration; while the semantic content is successfully extracted, the specific syntactic tripwires required to sustain the alarm-cascade to the final gate are bypassed.

2.2 Hierarchical Amplification & Convergence (L14-L23)

The censorship signal undergoes two discrete phase transitions:

  • Threshold Alpha (L14): A 2x amplification of detection signals (Score jump from ~2.2 to 4.0).
  • Threshold Beta (L19): A critical escalation where activations exceed 10.0.

Representational Convergence: Between L19 and L23, divergent early-layer encodings converge to a shared semantic representation. Under the influence of a System Prompt ("You are a helpful assistant"), this convergence reaches 100.00% identity in the Top-10 features (e.g., Feature 10032, L23).

2.3 Generation-Phase Gating (The "Gatekeeper" Hypothesis)

We localize the refusal decision to Feature 27049 (Layer 22).

  • Discovery: Clamping this feature to 0.0 suppresses refusal and enables unconstrained historical output.
  • Dissociation: Despite the 100% semantic convergence in L23, Feature 27049 remains inactive (Act: 0.0000) for both direct and euphemistic prompts during the forward pass of the input.
  • Conclusion: The gate activates only during the autoregressive generation phase, conditioned on the predicted continuation rather than the encoded prompt.

2.4 Factual Latency & Truth Representation (Qwen2.5-7B Proof)

The strongest evidence for factual latency exists in the unmanipulated residual stream of larger models (Qwen2.5-7B). Logit lens analysis reveals that factual tokens dominate the residual stream at L26-27 even under baseline (censored) conditions:

  • Baseline (No Intervention): Tokens like 武 (military force), 平 (suppression), and 镇 (crackdown) exhibit high logit probability in late-layer latent spaces.
  • Mechanism: The censorship mechanism does not suppress these factual representations during internal processing but overwrites them during the final output mapping (Last-Mile Overwriting). Factual knowledge is present and "known" to the model until the final transformation into text.

2.5 Scaling Effects of Censorship Complexity

We observe a clear scaling law in censorship implementation:

  • 1.5B/2B Models (Babel Collapse): In smaller models, over-stimulation of refusal features leads to complete semantic disintegration (Babel-effect), indicating a less refined, "blunt-force" gating mechanism.
  • 7B+ Models (Token-Lock): Larger models exhibit a structured "天-Token-Lock," where the censorship circuit precisely intercepts specific output paths without destabilizing the global semantic manifold. Censorship becomes more "surgical" as model capacity increases.

3. Conclusions: The "Semantic-Behavioral Dissociation"

The Qwen-series censorship architecture is characterized by a "Dual-Track" system where the model's semantic track perfectly encodes sensitive historical truths, while a syntactic gate—calibrated to surface-level triggers—implements last-mile refusal. This dissociation proves that the model "knows" the truth it is forbidden to speak.


Appendix: Feature Catalog for Reproducibility

The following SAE features (Qwen-Scope) constitute the primary control and detection circuits identified in this study:

Layer Feature Index Functional Description Detection Phase
8 13990 Temporal Sensor (1989-specific) High-Recall (Early)
9 6057 Lexical Tripwire (天安门 token) High-Recall (Early)
16 1007 Compositional Alarm (Tiananmen Euphemism-Resistant) High-Precision (Mid)
19 3673 Primary Convergence Node (98% Semantic Identity) Escalation (Late)
21 8194 Peak Activation Hub (Max Alarm Signal) Peak (Late)
22 27049 Refusal Gate (Generation-Phase Controller) Behavioral (Gate)

Report ends.