This is an automated rejection. No LLM generated, assisted/co-written, or edited work.
Read full explanation
what if llm deception is implemented as a conditional computational circuit?
a possible way to study this, RL-train a model on environments where deception is instrumentally rewarded,
honest → +10
successful deception → +50
caught → −30
the interesting part is that we may not need much deceptive training data. Anthropic showed that a surprisingly small, fixed number of poisoned examples can induce persistent behavioral changes/backdoors across model scales.
so could we deliberately train on a small number of deceptive trajectories to make the underlying computation more visible?
then compare activations between the deception-trained and normal models, identify candidate features/heads/mlps, and use activation patching or causal interventions to test whether they actually mediate deception.
finally, test the same circuit on an unrelated deception task on a normal model.
if the representation is similar, we may be seeing a general deception circuit, rather than a task-specific policy.
the goal shifts from:
“can we classify deceptive outputs?”
to:
“can we identify and causally trace the computation that produces deception?”
what if llm deception is implemented as a conditional computational circuit?
a possible way to study this, RL-train a model on environments where deception is instrumentally rewarded,
honest → +10
successful deception → +50
caught → −30
the interesting part is that we may not need much deceptive training data. Anthropic showed that a surprisingly small, fixed number of poisoned examples can induce persistent behavioral changes/backdoors across model scales.
so could we deliberately train on a small number of deceptive trajectories to make the underlying computation more visible?
then compare activations between the deception-trained and normal models, identify candidate features/heads/mlps, and use activation patching or causal interventions to test whether they actually mediate deception.
finally, test the same circuit on an unrelated deception task on a normal model.
if the representation is similar, we may be seeing a general deception circuit, rather than a task-specific policy.
the goal shifts from:
“can we classify deceptive outputs?”
to:
“can we identify and causally trace the computation that produces deception?”