Mar 23, 2026
Frankenstein ducks the linear probes
Abhinav Chand, Balaji R
Recent work has shown that linear probes are very effective in
detecting deceptive behavior of large language models, both
pre-generation (intent) and post-generation (harmful content) of
LLM’s response . We investigate fine-tuning methods to
suppress detection by deception probes. Our results show that we
can reduce the detection rate of linear probe monitors with our
fine-tuning method.
This seems interesting to me as a useful component for doing the red team iteration for a control game where the protocol involves linear probes. This is potentially useful for this.
One core issue is that this paper's threat model is unrealistic in a way that matters. They train against the exact probe direction they then evaluate on. But in a real deployment, the blue team constructs probes after receiving the model, possibly using different contrast pairs, different layers, or ensembles. The red team doesn't get to see those probes during training. So the paper demonstrates something closer to "you can overfit against a known classifier" rather than "you can make a model that's fundamentally harder to probe."
I also think this paper could have been condensed.
This project seems very similar to Obfuscated Activations (https://arxiv.org/pdf/2412.09565), with a topical shift to sleeper agents / AI control. The training method degrading sleeper agent behavior as a condition to evade the probe is a weakness, as is the lack of results showing generalization beyond evading the specific probe that was trained against.
Overall it's good and ambitious work for the scope of a hackathon, but would've benefitted from some more careful literature review + more accurate signposting of the weaknesses of the results. The continuation I'd be most excited for is a demonstration like the iterative monitor evasion result in figure 8 from Obfuscated Activations.
Cite this work
@misc {
title={
(HckPrj) Frankenstein ducks the linear probes
},
author={
Abhinav Chand, Balaji R
},
date={
3/23/26
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
