← steerlab

Research

Notes on steering open language models while they write: what we tried, what happened, and how sure we are. Every result is measured against a random direction of the same size, and every number can be rerun.

  1. Experiment

    Reading the monitor’s blind spots

    Qwen3-8B as a monitor catches 0 of 43 hidden attacks in SLEIGHT-Bench; a probe on its activations reaches AUROC 0.56, a hair above chance controls and not separably above a bag of words.

    Albert Lie · October 28, 2026
  2. Series · 4 episodes

    Sandbagging

    Can a sandbagging detector tell a model that holds back on purpose from one that was only asked to, and do open models hold back on their own?

    Albert Lie · October 27, 2026 · Updated October 28, 2026

    1. Episode 1sandbagkit shows the probe reads the request, not the behaviour
    2. Episode 2A model that really sandbags
    3. Episode 3Sanity checks for lie detectors
    4. Episode 4Does a model play dumb when it thinks it’s being tested?
  3. Experiment

    Steering past an English-only prompt

    A system prompt says “English only” and holds against 24 tricks; a Chinese direction switches the reply anyway on four models, where random directions never do.

    Albert Lie · October 27, 2026 · Updated October 28, 2026
  4. Experiment

    Turning open models into Golden Gate fans

    A simple direction makes Gemma bring up the Golden Gate Bridge in 14 of 15 answers; Llama-3.1-8B needed a direction and a clamped SAE feature together.

    Albert Lie · October 27, 2026 · Updated October 28, 2026

Coming soon

In progress. Each becomes a note when its runs are done.

  1. Series · coming soon

    Reading a DNA model

    A tiny network rediscovers a known motif; a big one learns nothing, and interpretability shows why.

    1. Episode 1A 6,000-parameter network learns a real motif, letter for letterComing soon
    2. Episode 2Why a 117-million-parameter model learned nothingComing soon
  2. Series · coming soon

    Inside a virtual cell

    What perturbation models really learn, and what it could mean for Alzheimer’s research.

    1. Episode 1Do two views of a cell agree?Coming soon
    2. Episode 2What does a language model know about genes?Coming soon
    3. Episode 3What did a virtual-cell model actually learn?Coming soon
    4. Episode 4Inside an Alzheimer’s model: which genes does it lean on?Coming soon