https://hellogxp.github.io/tags/mechanistic-interpretability/Recent content in Mechanistic-Interpretability on Xueping's BlogHugoen-usMon, 05 Oct 2026 00:00:00 +0000https://hellogxp.github.io/posts/probe-score-not-alarm-policy/Mon, 05 Oct 2026 00:00:00 +0000https://hellogxp.github.io/posts/probe-score-not-alarm-policy/An activation monitor that ranks attack cases above safe cases 92% of the time caught zero of 38 real attacks at the false-alarm budget a deployment could afford. The gap between a probe score and an alarm policy is eight fields wide — a research-informed practice guide to grading what your monitor actually claims, from our DAI 2026 paper.https://hellogxp.github.io/posts/the-unvalidated-instrument/Sun, 13 Sep 2026 00:00:00 +0000https://hellogxp.github.io/posts/the-unvalidated-instrument/Whether an LLM is conscious and whether it is scheming are argued through the same instrument — the model’s report about its own internal states. That instrument has never been calibrated. Ask it to detect an injected thought and it succeeds ~20% of the time; ask it to confess a hidden objective and a 2026 method reads it out at F1 0.98. Same channel, opposite verdicts, no spec sheet. A cited accounting of when self-report is load-bearing and when the reading is void.https://hellogxp.github.io/posts/agent-state-safety/Sun, 12 Jul 2026 00:00:00 +0000https://hellogxp.github.io/posts/agent-state-safety/AgentLens hits 99.15% detection accuracy; Neural Chameleons blinds a perfect AUROC=1.0 probe to 0% TPR with one trigger phrase. The gap between these numbers reveals four levels of claims that are frequently conflated: Decodability ≠ Prediction ≠ Causation ≠ Authority. A research-informed practice guide from internal representations to runtime authority boundaries.https://hellogxp.github.io/posts/trust-stack/Sat, 27 Jun 2026 00:00:00 +0000https://hellogxp.github.io/posts/trust-stack/AI models have two safety switches — ‘knowing’ and ‘acting’ — that are structurally separated in deep layers. Models can perfectly say ’no’ while perfectly doing ‘yes.’ RLHF’s gradient is mathematically zero beyond the harm horizon. LLM judges silently soften verdicts with zero acknowledgment in their reasoning traces. Every audit tool we have checks the wrong axis. This isn’t a bug. It’s structural.https://hellogxp.github.io/posts/cka-delta-universality/Sat, 20 Jun 2026 00:00:00 +0000https://hellogxp.github.io/posts/cka-delta-universality/Different LLM architectures encode high-level concepts in structurally compatible ways—but only up to a point. Geometric similarity between Llama, Qwen, Gemma, and Mistral is moderate at best (CKA_Δ around 0.3–0.7), yet a tiny affine map transfers persona classifiers with 94–99.9% accuracy across six concept domains. This geometric-functional dissociation suggests a weak version of the Platonic Representation Hypothesis: concept representations are universally recoverable, but not universally identical.https://hellogxp.github.io/posts/late-crystallization/Sat, 06 Jun 2026 00:00:00 +0000https://hellogxp.github.io/posts/late-crystallization/85.9% of correct answers never appear in any intermediate layer prediction. We discovered that factual knowledge in LLMs undergoes a sudden phase transition—crystallizing only at the final layer. This phenomenon explains why some interventions work and others don’t.