Research
Legend: Title — Posts on this site Title — Linkposts
- Aug 2026 Interpreting Prefill Awareness ↗
- Jun 2026 Out of Context Obfuscation: What Facts Matter for Probe Evasion? ↗
- Dec 2025 Open Source Replication of the Auditing Game Model Organism ↗ Replicated Marks et al. (Auditing language models for hidden objectives) on Llama 3.3 70B
- Jun 2025 RL-Obfuscation: Can Language Models Learn to Evade Latent-Space Monitors? Models can learn to evade (some) harmfulness probes if we reward them to do so, without needing to pass direct gradients through the probe.
- Apr 2025 MIB: A Mechanistic Interpretability Benchmark ↗
- Dec 2024 InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques We provided a new way to train models to follow predefined circuits, and then made a benchmark out of those models to evaluate circuit discovery methods.