detectorscope
Build AI text detectors whose decisions are mechanistically grounded and human-auditable (EMNLP 2026)
- 495
- Python
- Apache-2.0
- 4 wk ago
Build AI text detectors whose decisions are mechanistically grounded and human-auditable (EMNLP 2026)
Finding Direction of arrival (DOA) of small UAVs using Sparse Denoising Autoencoders and Deep Neural Networks.
Automates attribution-graph analysis via probe prompting: circuit-trace a prompt, auto-generate concept probes, profile feature activations, cluster supernodes.
Official Implementation of 'LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfer'
Inspect, steer & monitor a real LLM on Apple Silicon — a browser-based SAE interpretability lab for Qwen Scope, powered by MLX
Ouro-1.4B SAE-derived loop-boundary causal pilot: matched-control interventions, FP16/FP32 robustness analysis, reproducible figures, and scientific audit.
[ICLR 2026] AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features
An empirical study of activation functions in sparse autoencoders, examining reconstruction, sparsity, feature consistency, and language-model behavioral fidelity.
Do attribution graphs predict what GPT-2 really does? 8,064 feature edits, why graphs mislead (frozen LayerNorm/attention), a fix that keeps them linear, a trust score, a CLI and a browser demo.
Reverse-engineering neural network internals from scratch in NumPy + PyTorch. A 6-week masterclass: linear representation hypothesis, superposition, sparse autoencoders, transformer circuits & induction heads, activation/path patching & causal scrubbing, and steering a real LM. Fully executed notebooks.
Code for the blog post "Do block-sparse featurizers find concept manifolds in sentence embeddings?"
Dose-calibrated block-sparse-dictionary steering for Qwen3.5-2B — interactive demo + preregistered head-to-head vs SAE features
Does a language model's self-explanation actually depend on the activation it explains? Pre-registered controls for introspective verbalization, building on Li et al. (arXiv:2511.08579). Apparatus and frozen pre-registration - no measurement yet.