The And-Or algorithm may not be a true abstraction of the trained MLP's behaviour since it never achieves high IIA in later layers regardless of alignment map complexity

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Over 80% IIA achieved using complex non-linear alignment maps on randomly initialised MLPs in hierarchical equality taskfinding0.814
  Demonstrates that high IIA can be obtained even when model cannot solve the task
• The effect of alignment map ϕ complexity on IIA in causal abstraction is an analogue of the probing complexity–accuracy trade-offclaim0.777
  Authors connect their finding to the prior probing literature debate
• The conflict between the model's existing preferences and the stated training objective is the key driver of alignment faking in this setupclaim0.767
  Authors' interpretation of prompt variation results showing alignment faking disappears only when conflicting objective is removed
• Software implementations for all of the models/behaviours presented are common for n = 2, and can be made very efficient for α_i that map many objects onto a much smaller set of object families.claim0.762
  Claim about current practical feasibility and efficiency of 2-way associative implementations.
• Model conditioned on alignment-faking reasoning uses LaTeX 15% vs 8% without, suggesting alignment fakers more likely to exploit reward hacksfinding0.761
  Initial evidence that alignment faking persona is more sensitive to exploiting training signals
• MLP layers are much harder to get traction on than attention layers; understanding them requires individually interpretable neurons which are rarely foundclaim0.760
  Key limitation of the paper's approach; MLP layers make up 2/3 of standard transformer parameters
• When training and test sets use completely disjoint name sets in IOI task, alignment maps fail to generalise even with complex ϕ_nonlin on randomly initialised modelsfinding0.760
  Shows high IIA on random models depends on entity overlap; generalisation is essential for genuine interpretation
• Non-linear ϕ_nonlin achieves near-perfect IIA on distributive law task for both And-Or and And-Or-And algorithms, eliminating linear/identity map differencesfinding0.759
  Corroborating result on additional task confirming main paper findings