SafetyPage

Research papers · Paper

Where Animacy Lives in Large Language Models: Tracing the Circuits of the Animacy Concept

Samuele Punzo and Giovanni Cin\`a and Sandro Pezzelle · 13 Sep 2026

ReadSource at arxiv.org ↗Read hereCC BY 4.0

Outline · 33 pages

  1. Introductionp. 1
  2. Related Workp. 2
  3. Animacy in Linguistics and Cognitionp. 2
  4. Animacy in LLMsp. 2
  5. Circuit discoveryp. 3
  6. Methodsp. 3
  7. Datap. 3
  8. Task and Metricp. 4
  9. Behavioral Evaluationp. 5
  10. Circuit discoveryp. 5
  11. Resultsp. 5
  12. Animacy circuitp. 5
  13. Is the circuit robust?p. 6
  14. Is the circuit necessary?p. 6
  15. Analysisp. 7
  16. Functional role of the necessary core componentsp. 7
  17. A task-specific or a general-purpose animacy circuit?p. 7
  18. Discussionp. 8
  19. Dataset Constructionp. 12
  20. GPT 5.4 promptp. 12
  21. Semantic filtering promptp. 13
  22. Semantic score distributionp. 13
  23. Target constructionp. 13
  24. Metric Investigationp. 14
  25. Dataset success filteringp. 14
  26. EAP-IGp. 18
  27. Circuit sufficiency and comprehensiveness analysisp. 18
  28. Circuit necessity analysisp. 18
  29. Conditional ablation controlp. 19
  30. Discovery of Backup Circuits After Original-Circuit Ablationp. 19
  31. Comparison with random initializationsp. 19
  32. Comparison with random circuitsp. 20
  33. Construction of the transfer datasetsp. 20
  34. Setup 2 Circuit Discoveryp. 22
  35. Necessary core subcircuitsp. 24
  36. Induced score of necessary circuit componentsp. 26
  37. Logits contribution of necessary circuit componentsp. 27
  38. Attention patterns of the necessary circuitp. 28
  39. Circuit Discovery on Model-Specific vs Shared Datasetsp. 32