mlsanityData to 14 Oct 2026

Rethinking Contextualization by Reinterpreting Attention Head Channels

arXiv 2609.33851 ↗First posted 27 Sep 2026cs.LG cs.CLcited by 0

New in LLM reasoning; no direction yet. LLM reasoning

Abstract (arXiv)

Contextualization, the core operation of language modeling, transmits information across words to build sentence-specific word representations. Prior works mainly study contextualization, focusing on individual words and attention heads as a growing discrete dictionary, lacking a global view of their general behavior. Therefore, we propose a general principle: Globally, we find and estimate that different words carry different amounts of information, and less-informative words tend to absorb more contextual information. Specifically, these low-information words do not absorb contextual words uniformly, and finer-grained selectivity enables more precise routing to promote information transmission between matched words. Moreover, to find what mechanism causes such processing, we reinterpret attention heads as channels gated by their singular vectors and find that: (1) these singular vectors point to the hidden states of more informative words, allowing such words to write their information to others more strongly to act as information sources, and vice versa; and (2) these singular vectors can be viewed equally as hidden state features, enabling automated interpretation of attention heads beyond prior heuristic head discovery, also embedding heads into a continuous space rather than treating them as discrete, independent dictionary entries.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 26 of its 57 references that are in the map. First posted on arXiv 27 Sep 2026.

Builds on

26 of 57 references.

References · 26

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 13,028
  3. 2025Qwen3 Technical Reportcited by 8,820
  4. 2017Axiomatic Attribution for Deep Networkscited by 3,384
  5. 2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,411
  6. 2018Deep contextualized word representationscited by 2,290
  7. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  8. 2022In-context Learning and Induction Headscited by 964
  9. 2019Linguistic Knowledge and Transferability of Contextual Representationscited by 383
  10. 2023In-Context Learning Creates Task Vectorscited by 312
  11. 2023Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learningcited by 236
  12. 2024Retrieval Head Mechanistically Explains Long-Context Factualitycited by 232
14 more
  1. 2024How to use and interpret activation patchingcited by 214
  2. 2025Olmo 3cited by 129
  3. 2020Neuron Shapley: Discovering the Responsible Neuronscited by 98
  4. 2023Batch Calibration: Rethinking Calibration for In-Context Learning and Prompt Engineeringcited by 94
  5. 2023Successor Heads: Recurring, Interpretable Attention Heads In The Wildcited by 77
  6. 2024Attention Heads of Large Language Models: A Surveycited by 67
  7. 2024Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformerscited by 50
  8. 2024SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separatorcited by 41
  9. 2024Optimal ablation for interpretabilitycited by 25
  10. 2022Norm of Word Embedding Encodes Information Gaincited by 21
  11. 2024Correcting Negative Bias in Large Language Models through Negative Attention Score Alignmentcited by 12
  12. 2025Eliciting In-context Retrieval and Reasoning for Long-context Large Language Modelscited by 11
  13. 2026Decoupling the Benefits of Subword Tokenization for Language Model Training via Byte-level Simulationcited by 2
  14. 2025Binary Autoencoder for Mechanistic Interpretability of Large Language Modelscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2022Rethinking the Role of Scale for In-Context Learning: An Interpretability-based Case Study at 66 Billion Scale
  2. 2024Probing Context Localization of Polysemous Words in Pre-trained Language Model Sub-Layers
  3. 2025Taming Knowledge Conflicts in Language Models
  4. 2025Tracing Facts or just Copies? A critical investigation of the Competitions of Mechanisms in Large Language Models
  5. 2026Language Models Can Control Their Own Attention
  6. 2025Head Pursuit: Probing Attention Specialization in Multimodal Transformers
  7. 2026Memory Retrieval in Transformers: Insights from The Encoding Specificity Principle
  8. 2026Identifying Crucial Attention Heads for Multilingual Language Models: Retrieval and Retrieval-Transition Heads