mlsanityData to 7 Oct 2026

Causality neq Invariance: Function and Concept Vectors in LLMs

arXiv 2602.22424 ↗First posted 25 Feb 2026cs.CL cs.LGcited by 7

Not on the map. Why?

Abstract (arXiv)

Do large language models (LLMs) represent concepts abstractly, i.e., independent of input format? We revisit Function Vectors (FVs), compact representations of in-context learning (ICL) tasks that causally drive task performance. Across multiple LLMs, we show that FVs are not fully invariant: FVs are nearly orthogonal when extracted from different input formats (e.g., open-ended vs. multiple-choice), even if both target the same concept. We identify Concept Vectors (CVs), which carry more stable concept representations. Like FVs, CVs are composed of attention head outputs; however, unlike FVs, the constituent heads are selected using Representational Similarity Analysis (RSA) based on whether they encode concepts consistently across input formats. While these heads emerge in similar layers to FV-related heads, the two sets are largely distinct, suggesting different underlying mechanisms. Steering experiments reveal that FVs excel in-distribution, when extraction and application formats match (e.g., both open-ended in English), while CVs generalize better out-of-distribution across both question types (open-ended vs. multiple-choice) and languages. Our results show that LLMs do contain abstract concept representations, but these differ from those that drive ICL performance.

Key facts. Cited by 7 papers in the map (data through 7 Oct 2026). Builds on 21 of its 33 references that are in the map. First posted on arXiv 25 Feb 2026.

Builds on

21 of 33 references.

References · 21

  1. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,130
  2. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 679
  3. 2023In-Context Learning Creates Task Vectorscited by 302
  4. 2023Function Vectors in Large Language Modelscited by 292
  5. 2023Linearity of Relation Decoding in Transformer Language Modelscited by 190
  6. 2023The Hydra Effect: Emergent Self-repair in Language Model Computationscited by 145
  7. 2023How do Language Models Bind Entities in Context?cited by 102
  8. 2024Attention Heads of Large Language Models: A Surveycited by 65
  9. 2025Which Attention Heads Matter for In-Context Learning?cited by 61
  10. 2019Learning to Make Analogies by Contrasting Abstract Relational Structurecited by 53
  11. 2024Identifying Semantic Induction Heads to Understand In-Context Learningcited by 46
  12. 2024Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questionscited by 35
9 more
  1. 2025Emergent Symbolic Mechanisms Support Abstract Reasoning in Large Language Modelscited by 31
  2. 2024Human-like object concept representations emerge naturally in multimodal large language modelscited by 24
  3. 2025On Linear Representations and Pretraining Data Frequency in Language Modelscited by 23
  4. 2024Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Taskscited by 13
  5. 2025Contextualize-then-Aggregate: Circuits for In-Context Learning in Gemma-2 2Bcited by 10
  6. 2024Emergence and Effectiveness of Task Vectors in In-Context Learning: An Encoder Decoder Perspectivecited by 9
  7. 2025Do different prompting methods yield a common task representation in language models?cited by 9
  8. 2025Provable In-Context Vector Arithmetic via Retrieving Task Conceptscited by 7
  9. 2025Whither symbols in the era of advanced neural networks?cited by 6

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 7.

7 papers.

  1. 2026Test, then Route: How Language Models Execute In-Context Conditional Rules Across Models and Languagescited by 1
  2. 2026Necessary, Decodable and Reversible, Yet Not Transferable: A Stress Test for Attention-Head Role Claimscited by 0
  3. 2026A Pre-Specified Construction-Confirmation Test of Operation-Level Causal Transfer Across Finite Isomorphic Symbolic Domainscited by 0
  4. 2026How Do Language Models Choose Between Context and Memory?cited by 0
  5. 2026In-Context Learning Amplifies a Latent Symbolic Circuitcited by 0
  6. 2026Which Attention Heads are like the Human Head? Not the Ones that Computecited by 0
  7. 2026The Geometry of Contextual Relations: Language Models Address Facts by Order of Mentioncited by 0