Pinned Loading

1. sycophancy-construct-validity sycophancy-construct-validity Public

   Construct validity of sycophancy interventions in Llama-3.1-8B — reading a concept is not controlling it

   Python

2. safety-concept-vectors safety-concept-vectors Public

   Safety-concept directions in Qwen2.5-7B: contrastive extraction and linear probes

   Python 1

3. eval-awareness-detection eval-awareness-detection Public

   Prototype: mechanistic eval-awareness detection in open-weight LLMs (March 2026)

   Python 1

4. does-quantization-kill-interpretability does-quantization-kill-interpretability Public

   Does Quantization Kill Interpretability? Scaling study across 5 models (124M-2.8B): RTN destroys induction heads in small models, GPTQ preserves them at all scales.

   Python 1

5. pcm-bitslicing pcm-bitslicing Public

   Python 1