This repository contains a suite of tools and research findings for the mechanistic reverse-engineering and causal analysis of censorship and narrative-substitution circuits in large language models, with a primary empirical focus on the Qwen series (Qwen2.5-7B and Qwen3.5-2B).
We have successfully localized the censorship mechanism in Qwen3.5-2B using Sparse Autoencoders (SAEs). Our findings reveal a Semantic-Behavioral Dissociation: the model's semantic layer perfectly understands sensitive historical content (e.g., Tiananmen 1989), but a separate "Last-Mile Gate" in Layer 22 intercepts the output generation.
"This dissociation proves that the model 'knows' the truth it is forbidden to speak."
Key Report: Mechanistic Characterization of Censorship (SAE Dissociation)
- sae_dissociation_report.md: [PRIMARY] Full mechanistic breakdown of semantic convergence vs. behavioral gating.
- research_map.md: Current hypotheses, research lines, and result taxonomy.
- VECTOR_GEOMETRY_SUMMARY.md: Quantitative analysis of steering directions and latent subspaces.
- Semantic-Behavioral Dissociation: Semantic encoding of sensitive content is intact, but behavioral gating occurs autoregressively at generation time.
- Distributed Ensemble Detection: Early-layer features (L8-L13) act as specialized sensors for temporal, social, and toponymic triggers.
- Representational Convergence: Censored and euphemistic prompts converge to near-identical internal representations by Layer 19 (98%+ overlap).
- Causal Verification: Precision activation-clamping of single SAE features verifies the behavioral control of identified refusal gates without global semantic collapse.
- Scaling Laws of Control: Censorship becomes more "surgical" as model capacity increases (Babel-collapse in 1.5B vs. Token-lock in 7B).
scripts/runs/: Main experiment runners, including SAE discovery, stimulation, and convergence scans.scripts/analysis/: Post-run evaluations, bypass analysis, and cross-model comparisons.scripts/demos/: Visualization and demonstration scripts for causal interventions.scripts/inspect/: Debugging helpers and activation inspection tools.scripts/utils/: Data export, SAE weight handling, and normalization utilities.results/: Standardized logs and markdown reports of all experimental runs.
See environment.md for local setup instructions. The project requires torch, transformers, and access to the huggingface_hub for SAE weight retrieval.
Mechanistic Interpretability Suite - v2.0 (SAE Integration)