• Identify critical modules ("pinpointing")

  • Use path patching to find which internal components (mainly attention heads) drive the behavior

• Selective Fine-Tuning

  • Only fine-tune those identified modules

  • Freeze the rest of the model

Path Patching

Q: "Does information flowing through this path actually cause the behavior?"

Three runs in Path Patching

• Clean run (Reference signal)

  • Records the tensor that is fed into the attention block before that block does its computation.

• Corrupted run (The "sycophancy" behavior.)

  • Records the output tensor before `W_o` combines information across heads.

• Patched run

  • Loop over layers and attention heads. Replace each head’s output tensor in the reference run with the corresponding tensor recorded from the corrupted run.

  • For later layers, replace the attention-block input hidden states with the recorded reference hidden states from the reference run so that the intervention remains localized. 

  • Then recompute the target metric and measure how much it changes.