mlsanityData to 14 Oct 2026

Jet Expansions of Residual Computation

arXiv 2410.06024 ↗First posted 8 Oct 2024cs.LG cs.AI cs.CLcited by 4

Not on the map. Why?

Abstract (arXiv)

We introduce a framework for expanding residual computational graphs using jets, operators that generalize truncated Taylor series. Our method provides a systematic approach to disentangle contributions of different computational paths to model predictions. In contrast to existing techniques such as distillation, probing, or early decoding, our expansions rely solely on the model itself and requires no data, training, or sampling from the model. We demonstrate how our framework grounds and subsumes logit lens, reveals a (super-)exponential path structure in the recursive residual depth and opens up several applications. These include sketching a transformer large language model with n-gram statistics extracted from its computations, and indexing the models' levels of toxicity knowledge. Our approach enables data-free analysis of residual computation for model interpretability, development, and evaluation.

Key facts. Cited by 4 papers in the map (data through 14 Oct 2026). Builds on 39 of its 57 references that are in the map. First posted on arXiv 8 Oct 2024.

Builds on

39 of 57 references.

References · 39

  1. 2017Attention Is All You Needcited by 60,361
  2. 2015Deep Residual Learning for Image Recognitioncited by 52,096
  3. 2023LLaMA: Open and Efficient Foundation Language Modelscited by 14,132
  4. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 13,028
  5. 2017A Unified Approach to Interpreting Model Predictionscited by 6,555
  6. 2021Finetuned Language Models Are Zero-Shot Learnerscited by 3,780
  7. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,720
  8. 2017Axiomatic Attribution for Deep Networkscited by 3,384
  9. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  10. 2020Transformer Feed-Forward Layers Are Key-Value Memoriescited by 1,336
  11. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,255
  12. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
27 more
  1. 2022Toy Models of Superpositioncited by 971
  2. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 724
  3. 2023Eliciting Latent Predictions from Transformers with the Tuned Lenscited by 623
  4. 2024OLMo: Accelerating the Science of Language Modelscited by 613
  5. 2022Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Spacecited by 532
  6. 2022ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detectioncited by 422
  7. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
  8. 2015Highway Networkscited by 260
  9. 2019Depth-Adaptive Transformercited by 236
  10. 2023Localizing Model Behavior with Path Patchingcited by 195
  11. 2024Infini-gram: Scaling Unbounded n-gram Language Models to a Trillion Tokenscited by 153
  12. 2023Language Models Implement Simple Word2Vec-style Vector Arithmeticcited by 98
  13. 2021An Interpretability Illusion for BERTcited by 97
  14. 2024Information Flow Routes: Automatically Interpreting Language Models at Scalecited by 95
  15. 2023Neurons in Large Language Models: Dead, N-gram, Positionalcited by 88
  16. 2024A Primer on the Inner Workings of Transformer-based Language Modelscited by 85
  17. 2017Residual Connections Encourage Iterative Inferencecited by 77
  18. 2015Conditional Deep Learning for Energy-Efficient and Enhanced Pattern Recognitioncited by 67
  19. 2023The Foundation Model Transparency Indexcited by 66
  20. 2023Tractable Control for Autoregressive Language Generationcited by 65
  21. 2024Transformers Can Represent n-gram Language Modelscited by 32
  22. 2024Neural Networks Learn Statistics of Increasing Complexitycited by 31
  23. 2024Understanding Transformers via N-gram Statisticscited by 26
  24. 2024Missed Causes and Ambiguous Effects: Counterfactuals Pose Challenges for Interpreting Neural Networkscited by 23
  25. 2019An Information Theoretic Interpretation to Deep Neural Networkscited by 19
  26. 2024Decomposing and Editing Predictions by Modeling Model Computationcited by 19
  27. 2024Explaining Probabilistic Models with Distributional Valuescited by 3

Built on by

Cited by year

Citing papers per year: 2025: 4.

4 papers.

  1. 2025LLM Unlearning via Neural Activation Redirectioncited by 41
  2. 2025Scaling Embedding Layers in Language Modelscited by 20
  3. 2025Evolution of Concepts in Language Model Pre-Trainingcited by 8
  4. 2025DDEQs: Distributional Deep Equilibrium Models through Wasserstein Gradient Flowscited by 4

Similar, not cited

Close in wording, no citation either way

  1. 2026Multi-Stream Residual Computation in Diffusion Language Models
  2. 2026From Residuals to Reasons: LLM-Guided Mechanism Inference from Tabular Data
  3. 2026A Geometric Perspective on Next-Token Prediction in Large Language Models: Three Emerging Phases
  4. 2025M2R2: Mixture of Multi-Rate Residuals for Efficient Transformer Inference
  5. 2025Transformer Dynamics: A neuroscientific approach to interpretability of large language models
  6. 2026QuantForge: Discovering Residual Decompositions for MXFP4 Post-Training Quantization
  7. 2026Residual Context Diffusion Language Models
  8. 2026One Step Forward and K Steps Back: Better Reasoning with Denoising Recursion Models