# The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content

arXiv: https://arxiv.org/abs/2606.11198 · HTML: https://mlsanity.com/paper/2606.11198/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in RAG robustness and adaptive retrieval, one of 15 directions in Retrieval-augmented generation.
- Builds on 21 of its 23 references that are in the map.
- First posted on arXiv 21 Apr 2026.

## Placed in

- Retrieval-augmented generation / [RAG robustness and adaptive retrieval](https://mlsanity.com/rag/rag-robustness-and-adaptive-retrieval/)

## Builds on (most cited first, 21 of 21)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://mlsanity.com/paper/2201.11903/) (2022)
- [HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering](https://mlsanity.com/paper/1809.09600/) (2018)
- [Mistral 7B](https://mlsanity.com/paper/2310.06825/) (2023)
- [TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension](https://mlsanity.com/paper/1705.03551/) (2017)
- [CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge](https://mlsanity.com/paper/1811.00937/) (2018)
- [Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?](https://mlsanity.com/paper/2202.12837/) (2022)
- [Large Language Models Can Be Easily Distracted by Irrelevant Context](https://mlsanity.com/paper/2302.00093/) (2023)
- [Active Retrieval Augmented Generation](https://mlsanity.com/paper/2305.06983/) (2023)
- [Can We Edit Factual Knowledge by In-Context Learning?](https://mlsanity.com/paper/2305.12740/) (2023)
- [PRODIGY: Enabling In-context Learning Over Graphs](https://mlsanity.com/paper/2305.12600/) (2023)
- [Few-shot In-context Learning for Knowledge Base Question Answering](https://mlsanity.com/paper/2305.01750/) (2023)
- [LAMBADA: Backward Chaining for Automated Reasoning in Natural Language](https://mlsanity.com/paper/2212.13894/) (2022)
- [Skill-Based Few-Shot Selection for In-Context Learning](https://mlsanity.com/paper/2305.14210/) (2023)
- [Disentangling Memory and Reasoning Ability in Large Language Models](https://mlsanity.com/paper/2411.13504/) (2024)
- [How Much Can RAG Help the Reasoning of LLM?](https://mlsanity.com/paper/2410.02338/) (2024)
- [Towards A Unified View of Answer Calibration for Multi-Step Reasoning](https://mlsanity.com/paper/2311.09101/) (2023)
- [The Cost of Down-Scaling Language Models: Fact Recall Deteriorates before In-Context Learning](https://mlsanity.com/paper/2310.04680/) (2023)
- ["In-Context Learning" or: How I learned to stop worrying and love "Applied Information Retrieval"](https://mlsanity.com/paper/2405.01116/) (2024)
- [Schema for In-Context Learning](https://mlsanity.com/paper/2510.13905/) (2025)

## Built on by (most cited first, 1 of 1)

- [The Announcement Carries the Cue: Markup, Boundaries, and the Notation of Pre-Training Corpora](https://mlsanity.com/paper/2608.09093/) (2026)

## Abstract (arXiv)

Retrieval-augmented generation (RAG) systems inject external knowledge to improve LLM outputs, yet the format of injected content -- distinct from its semantic relevance -- can independently distort the model's attention distribution. We identify and formalise a phenomenon we term the structural attention tax: knowledge graph (KG) triples, due to their relational delimiters and repeated slot patterns, capture 2-3x more attention per token than semantically equivalent natural-language text (hato(KG) ≈ 0.70 vs. hato(neutral) ≈ 0.25), compressing demonstration attention by up to 42% -- regardless of whether the triples are relevant or noise. We develop a formal framework decomposing attention scores into semantic and structural components (Eq. 2), derive a compression bound (Proposition 1) connecting token-level format bias to demonstration attention loss, and show that the structural term governs how much attention is diverted while the semantic term governs whether this helps or hurts. This decoupling reveals two orthogonal axes for improving retrieval-augmented ICL: optimising retrieval quality (semantic axis) and reducing format-driven attention capture (structural axis). Empirically, across two model families (Mistral-7B, LLaMA-3-8B) and three QA benchmarks, we observe that source-task alignment dominates: task-matched BM25 retrieval achieves 58-62% on HotpotQA vs. ConceptNet's 25-27%, a >30 pp gap that dwarfs all gating strategies (≤2 pp). We derive five structure-aware mitigation strategies from the framework, ranging from zero-cost prompt modifications to training-time regularisation; format flattening (S3) is validated by both accuracy and attention-level evidence from a verbalized-triple control, while structural dispersal (S1) yields mixed results that illuminate the challenges of format-level intervention.