Indirect Object Identification (IOI) Task

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

• Interpretability in the Wild: A Circuit for Indirect Object Identification in GPT-2 Small (Wang et al., 2023)concept0.727
  Cited as causal intervention methodology precedent for this paper's ablation approach
• 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 modelsfinding0.707
  Shows high IIA on random models depends on entity overlap; generalisation is essential for genuine interpretation
• Identification of algorithms implemented in attention layers, distributed across attention headsfinding0.707
  VPD successfully recovered interpretable attention algorithms (previous-token behavior, syntax-boundary routing) in weight space without requiring manual decomposition across heads.
• Detecting Unintended Outputs via Introspectionfinding0.705
  Models can distinguish artificially prefilled outputs from intentional responses by referencing prior internal representations; injection of matching concept vector causes model to retroactively accept prefill as intentional.
• Causal emergence identification tasks can be understood as causal representation learning tasks.claim0.696
  Authors propose a conceptual mapping between CE identification and CRL.
• Object Oriented Programmingframework0.696
• What is the coordination mechanism that enables this cross-level integrated information processing?question0.691
  Key mechanistic question answered by developmental bioelectricity.
• Training identical architectures on the same data with different objective functions should produce systematically different internal evaluative representations, detectable through interpretability tools, even when final task performance is matchedhypothesis0.690
  Second falsifiable prediction linking objective function structure to valence profile