mlsanityData to 7 Oct 2026

Compressed Sensing for Capability Localization in Large Language Models

arXiv 2603.03335 ↗First posted 11 Feb 2026cs.CLcited by 2

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

Large language models (LLMs) exhibit a wide range of capabilities, including mathematical reasoning, code generation, and linguistic behaviors. We show that Transformer architectures contain small subsets of attention heads that are necessary for certain capabilities. Zeroing out as few as five task-specific heads can degrade performance by up to 60% on standard benchmarks measuring the capability of interest, while largely preserving performance on unrelated tasks. We introduce a compressed sensing-based method that exploits the sparsity of these heads to identify them via strategic knockouts and a small number of model evaluations. We validate these findings across Llama and Qwen models ranging from 1B to 14B parameters and a diverse set of capabilities including mathematical abilities and code generation, revealing a modular organization in which specialized capabilities are dependent on sparse, functionally distinct components. Overall, our results suggest that capability localization is a general organizational principle of Transformer language models, with implications for interpretability, model editing, and AI safety. Code is released at https://github.com/locuslab/llm-components.

Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 47 of its 61 references that are in the map. First posted on arXiv 11 Feb 2026.

Builds on

47 of 61 references.

References · 47

  1. 2020Language Models are Few-Shot Learnerscited by 32,199
  2. 2024The Llama 3 Herd of Modelscited by 14,192
  3. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  4. 2021Evaluating Large Language Models Trained on Codecited by 8,053
  5. 2020Measuring Massive Multitask Language Understandingcited by 7,768
  6. 2017A Unified Approach to Interpreting Model Predictionscited by 6,488
  7. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,040
  8. 2018Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challengecited by 4,710
  9. 2019HellaSwag: Can a Machine Really Finish Your Sentence?cited by 4,018
  10. 2021Program Synthesis with Large Language Modelscited by 3,509
  11. 2017Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layercited by 3,413
  12. 2017Axiomatic Attribution for Deep Networkscited by 3,350
35 more
  1. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,115
  2. 2017Learning Important Features Through Propagating Activation Differencescited by 1,454
  3. 2019Machine Unlearningcited by 1,020
  4. 2019What Does BERT Look At? An Analysis of BERT's Attentioncited by 994
  5. 2022Toy Models of Superpositioncited by 928
  6. 2019Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Prunedcited by 844
  7. 2019Are Sixteen Heads Really Better than One?cited by 803
  8. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 700
  9. 2019Certified Data Removal from Machine Learning Modelscited by 512
  10. 2020Amnesiac Machine Learningcited by 305
  11. 2024DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Headscited by 283
  12. 2023Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Modelscited by 237
  13. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 196
  14. 2023Task-Specific Skill Localization in Fine-tuned Language Modelscited by 104
  15. 2024RazorAttention: Efficient KV Cache Compression Through Retrieval Headscited by 99
  16. 2024Arithmetic Without Algorithms: Language Models Solve Math With a Bag of Heuristicscited by 94
  17. 2024Cutting Off the Head Ends the Conflict: A Mechanism for Interpreting and Mitigating Knowledge Conflicts in Language Modelscited by 87
  18. 2023Copy Suppression: Comprehensively Understanding an Attention Headcited by 73
  19. 2024Harder Tasks Need More Experts: Dynamic Routing in MoE Modelscited by 71
  20. 2024On the Role of Attention Heads in Large Language Model Safetycited by 70
  21. 2023Can Neural Network Memorization Be Localized?cited by 66
  22. 2024Attention Heads of Large Language Models: A Surveycited by 65
  23. 2025Which Attention Heads Matter for In-Context Learning?cited by 61
  24. 2024Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neuronscited by 47
  25. 2024A Closer Look into Mixture-of-Experts in Large Language Modelscited by 47
  26. 2024Interpreting and Improving Large Language Models in Arithmetic Calculationcited by 43
  27. 2025Advancing Expert Specialization for Better MoEcited by 37
  28. 2024Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inferencecited by 28
  29. 2023Circuit Breaking: Removing Model Behaviors with Targeted Ablationcited by 27
  30. 2025ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMscited by 16
  31. 2025Mixture of Sparse Attention: Content-Based Learnable Sparse Attention via Expert-Choice Routingcited by 13
  32. 2025SPEX: Scaling Feature Interaction Explanations for LLMscited by 12
  33. 2024Whispers that Shake Foundations: Analyzing and Mitigating False Premise Hallucinations in Large Language Modelscited by 9
  34. 2025Beyond Components: Singular Vector-Based Interpretability of Transformer Circuitscited by 6
  35. 2025ALPS: Attention Localization and Pruning Strategy for Efficient Alignment of Large Language Modelscited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretabilitycited by 2
  2. 2026From Latents to Wires: Surgical Post-Editing on Large Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Capability-Guided Compression: Toward Interpretability-Aware Budget Allocation for Large Language Models
  2. 2025Speed Always Wins: A Survey on Efficient Architectures for Large Language Models
  3. 2026Compressed-Sensing-Guided, Inference-Aware Structured Reduction for Large Language Models
  4. 2025Optimizing LLMs for Resource-Constrained Environments: A Survey of Model Compression Techniques
  5. 2025Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression
  6. 2024Compressing Large Language Models with Automated Sub-Network Search
  7. 2025Pimba: A Processing-in-Memory Acceleration for Post-Transformer Large Language Model Serving
  8. 2026Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models