mlsanityData to 14 Oct 2026

Emergence and Localisation of Semantic Role Circuits in LLMs

arXiv 2511.20910 ↗First posted 25 Nov 2025cs.CLcited by 2

On the map: Mechanistic interpretability → Knowledge and representations inside LLMs

Abstract (arXiv)

Despite displaying semantic competence, large language models' internal mechanisms that ground abstract semantic structure remain insufficiently characterised. We propose a method integrating role-cross minimal pairs, temporal emergence analysis, and cross-model comparison to study how LLMs implement semantic roles. Our analysis uncovers: (i) highly concentrated circuits (89-94% attribution within 28 nodes); (ii) gradual structural refinement rather than phase transitions, with larger models sometimes bypassing localised circuits; and (iii) moderate cross-scale conservation (24-59% component overlap) alongside high spectral similarity. These findings suggest that LLMs form compact, causally isolated mechanisms for abstract semantic structure, and these mechanisms exhibit partial transfer across scales and architectures.

Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Placed in Knowledge and representations inside LLMs, one of 13 directions in Mechanistic interpretability. Builds on 26 of its 64 references that are in the map. First posted on arXiv 25 Nov 2025.

Builds on

26 of 64 references.

Closest ancestors · 1

Same direction, newest first

  1. 2025TRACE: Training and Inference-Time Interpretability Analysis for Language Models4 Jul

Other references · 25

  1. 2017Axiomatic Attribution for Deep Networkscited by 3,384
  2. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  3. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,411
  4. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  5. 2019BERT Rediscovers the Classical NLP Pipelinecited by 1,059
  6. 2022In-context Learning and Induction Headscited by 964
  7. 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
  8. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  9. 2022Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetscited by 629
  10. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
  11. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 209
  12. 2024Transcoders Find Interpretable LLM Feature Circuitscited by 200
13 more
  1. 2024Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanismscited by 149
  2. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 115
  3. 2024AtP*: An efficient and scalable method for localizing LLM behaviour to componentscited by 111
  4. 2024LLM Circuit Analyses Are Consistent Across Training and Scalecited by 63
  5. 2021Disentangling Syntax and Semantics in the Brain with Deep Networkscited by 40
  6. 2024Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic taskscited by 40
  7. 2024Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inferencecited by 29
  8. 2023Subspace Chronicles: How Linguistic Information Emerges, Shifts and Interacts during Language Model Trainingcited by 23
  9. 2024Potential and Limitations of LLMs in Capturing Structured Semantics: A Case Study on SRLcited by 10
  10. 2025Sparse Feature Coactivation Reveals Causal Semantic Modules in Large Language Modelscited by 6
  11. 2025A Systematic Survey of Semantic Role Labeling in the Era of Pretrained Language Modelscited by 3
  12. 2025CARMA: Enhanced Compositionality in LLMs via Advanced Regularisation and Mutual Information Alignmentcited by 3
  13. 2025TRACE for Tracking the Emergence of Semantic Representations in Transformerscited by 3

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026From Circuit Evidence to Mechanistic Theory: An Inductive Logic Approachcited by 1
  2. 2026Align, Unify, Suppress, Route: A Coherentist View of Transformer Computationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Emergent Semantic Role Understanding in Language Models
  2. 2025A Review of Developmental Interpretability in Large Language Models
  3. 2026LLMs Explain't: A Post-Mortem on Semantic Interpretability in Transformer Models
  4. 2025Comprehension Without Competence: Architectural Limits of LLMs in Symbolic Computation and Reasoning
  5. 2026Understanding Large Language Models
  6. 2026Towards a Mechanistic Understanding of Propositional Logical Reasoning in Large Language Models
  7. 2024Understanding Language Model Circuits through Knowledge Editing
  8. 2026LLM Reasoning Is Latent, Not the Chain of Thought