References

Citation edges between papers. Most are from OpenAlex; some from LLM-extracted bibliographies. Showing latest 500 of 8504.

Total citations
8504
Unique citing papers
131
Unique cited papers
5927
By source
s2-reference-walk5431
openalex2089
llm984
Citing paperCited paperDOIStatusSource
Understanding and Steering Llama 3 with Sparse Autoencoders Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 210.48550/arxiv.2408.05147
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders SAE-derived indicators provide a sharp lens into per-sample distribution shift w—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders project-gutenberg-poetry-corpus—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Sparse Crosscoders for Cross-Layer Features and Model Diffing—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Towards monosemanticity: Decomposing language models with dictionary learning—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Prisma: An Open Source Toolkit—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Toy Models of Superposition10.48550/arxiv.2209.10652
active
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders UNDERSTAND-ING DEEP LEARNING REQUIRES RE-THINKING GENERALIZATION—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Pattern recognition and machine learning Information science and statistics—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Announcing open-source saes for llama 3.3 70b and llama 3.1 8b—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Language Models are Unsupervised Multitask Learners—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Dataset Shift in Machine Learning10.7551/mitpress/978026217
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neura—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Spoken SQuAD: A Study of Mitigating the Impact of Speech Recognition Errors on L10.21437/interspeech.2018-
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders SQuAD: 100,000+ Questions for Machine Comprehension of Text10.18653/v1/d16-1264
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Convergent Learning: Do different neural networks learn the same representations10.48550/arxiv.1511.07543
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Deep Anomaly Detection with Outlier Exposure—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversa—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Exploring the Limits of Weakly Supervised Pretraining10.1007/978-3-030-01216-8_
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Invariant Risk Minimization—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Do ImageNet Classifiers Generalize to ImageNet?—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Scaling Laws for Neural Language Models—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Language Models are Few-Shot Learners—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders On Faithfulness and Factuality in Abstractive Summarization10.18653/v1/2020.acl-main.
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Energy-based Out-of-distribution Detection—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Measuring Massive Multitask Language Understanding—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Survey of Hallucination in Natural Language Generation10.1145/3571730
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Learning Transferable Visual Models From Natural Language Supervision—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Training Compute-Optimal Large Language Models10.52202/068431-2176
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Holistic Evaluation of Language Models10.1111/nyas.15007
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders TinyStories: How Small Can Language Models Be and Still Speak Coherent English?—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Emergent Abilities of Large Language Models10.48550/arxiv.2206.07682
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Steering language models with activation engineering—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Sparse autoencoders find highly interpretable features in language models10.48550/arxiv.2309.08600
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Jailbroken: How does LLM safety training fail?10.48550/arxiv.2307.02483
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Universal and Transferable Adversarial Attacks on Aligned Language Models—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders LMSYS-Chat-1M: A large-scale real-world LLM conversation dataset10.48550/arxiv.2309.11998
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders The linear representation hypothesis and the geometry of large language models10.48550/arxiv.2311.03658
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders A StrongREJECT for Empty Jailbreaks10.48550/arxiv.2402.10260
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Not all language model features are one-dimensionally linear—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders LLM Task Interference: An Initial Study on the Impact of Task-Switch in Conversa10.48550/arxiv.2402.18216
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey10.48550/arxiv.2403.14608
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Jailbreak Attacks and Defenses Against Large Language Models: A Survey10.48550/arxiv.2407.04295
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders OR-Bench: An Over-Refusal Benchmark for Large Language Models10.48550/arxiv.2405.20947
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Langu10.48550/arxiv.2406.18510
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Quantifying Feature Space Universality Across Large Language Models via Sparse A—
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Decomposing The Dark Matter of Sparse Autoencoders10.48550/arxiv.2410.14670
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Reasoning Robustness of LLMs to Adversarial Typographical Errors10.18653/v1/2024.emnlp-mai
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders MMLU-CF: A Contamination-free Multi-task Language Understanding Benchmark10.48550/arxiv.2412.15194
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Steering Language Model Refusal with Sparse Autoencoders10.48550/arxiv.2411.11296
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Sparse Autoencoders Trained on the Same Data Learn Different Features10.48550/arxiv.2501.16615
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Which of These Best Describes Multiple Choice Evaluation with LLMs? A) Forced B)10.48550/arxiv.2502.14127
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Sparse Autoencoders Do Not Find Canonical Units of Analysis10.48550/arxiv.2502.04878
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders The origins of representation manifolds in large language models10.48550/arxiv.2505.18235
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Steering CLIP's vision transformer with sparse autoencoders10.48550/arxiv.2504.08729
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders From Noise to Narrative: Tracing the Origins of Hallucinations in Transformers10.48550/arxiv.2509.06938
referenced-only
s2-reference-walk
Understanding and Steering Llama 3 with Sparse Autoencoders Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet—
active
s2-reference-walk
Stochastic Parameter DecompositionSparse Crosscoders for Cross-Layer Features and Model Diffing—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionSkeletonization: A Technique for Trimming the Fat from a Network via Relevance A—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionCausal scrubbing: a method for rigorously testing interpretability hypotheses—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionAlgebraic geometry and statistical learning theory—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionTanh penalty in dictionary learning—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionTowards monosemanticity: Decomposing language models with dictionary learning—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionTaking features out of superposition with sparse autoencoders—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionThe Quest for the Right Mediator: A History, Survey, and Theoretical Grounding o10.48550/arxiv.2408.01416
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionAddressing feature suppression in saes—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionInvestigating Gender Bias in Language Models Using Causal Mediation Analysis—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionAUTO-ENCODING VARIATIONAL BAYES—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionAxiomatic Attribution for Deep Networks—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionCounterfactuals10.1007/bf02123381
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionDeep Inside Convolutional Networks: Visualising Image Classification Models and —
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionSanity Checks for Saliency Maps—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionConceptual challenges for interpretable machine learning10.1007/s11229-022-03485-5
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionDecoupled Weight Decay Regularization—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionInterpretable explanations of black boxes by meaningful perturbation10.1109/iccv.2017.371
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionTransformer visualization via dictionary learning: contextualized embedding as a10.18653/v1/2021.deelio-1.
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionLow-Complexity Probing via Finding Subnetworks10.18653/v1/2021.naacl-mai
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionAre Neural Nets Modular? Inspecting Functional Modularity Through Differentiable—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionInterpretability in the Wild: a Circuit for Indirect Object Identification in GP10.48550/arxiv.2211.00593
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionCan Subnetwork Structure be the Key to Out-of-Distribution Generalization?—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionCausal abstractions of neural networks—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionFinding Alignments Between Interpretable Causal Variables and Distributed Neural10.48550/arxiv.2303.02536
active
s2-reference-walk
Stochastic Parameter DecompositionThe Hydra Effect: Emergent Self-repair in Language Model Computations10.48550/arxiv.2307.15771
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionTowards Automated Circuit Discovery for Mechanistic Interpretability10.48550/arxiv.2304.14997
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionAttribution Patching Outperforms Automated Circuit Discovery10.48550/arxiv.2310.10348
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionUncovering Model Processing Strategies with Non-Negative Per-Example Fisher Fact—
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionSparse autoencoders find highly interpretable features in language models10.48550/arxiv.2309.08600
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionMissed Causes and Ambiguous Effects: Counterfactuals Pose Challenges for Interpr10.48550/arxiv.2407.04690
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionAtP*: An efficient and scalable method for localizing LLM behaviour to component10.48550/arxiv.2403.00745
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionSparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in 10.48550/arxiv.2403.19647
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionA is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoenc10.48550/arxiv.2409.14507
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionMathematical Models of Computation in Superposition10.48550/arxiv.2408.05451
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionSparse Autoencoders Do Not Find Canonical Units of Analysis10.48550/arxiv.2502.04878
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionInterpretability in Parameter Space: Minimizing Mechanistic Description Length w10.48550/arxiv.2501.14926
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionDifferentiation and Specialization of Attention Heads via the Refined Local Lear10.48550/arxiv.2410.02984
referenced-only
s2-reference-walk
Stochastic Parameter DecompositionIdentifying Sparsely Active Circuits Through Local Loss Landscape Decomposition10.48550/arxiv.2504.00194
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionScaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet—
active
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionBeta-CROWN: Efficient Bound Propagation with Per-neuron Split Constraints for Co—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionTowards monosemanticity: Decomposing language models with dictionary learning—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionIntriguing properties of neural networks—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionDo Perceptually Aligned Gradients Imply Adversarial Robustness?10.48550/arxiv.2207.11378
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionExplaining and Harnessing Adversarial Examples—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionImageNet Large Scale Visual Recognition Challenge10.1007/s11263-015-0816-y
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAdversarial examples in the physical world10.1201/9781351251389-8
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionTowards Evaluating the Robustness of Neural Networks10.1109/sp.2017.49
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionDeepFool: A Simple and Accurate Method to Fool Deep Neural Networks10.1109/cvpr.2016.282
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionUniversal Adversarial Perturbations10.1109/cvpr.2017.17
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionTowards Deep Learning Models Resistant to Adversarial Attacks—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionDelving into Transferable Adversarial Examples and Black-box Attacks—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionCertified Robustness to Adversarial Examples with Differential Privacy10.1109/sp.2019.00044
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAdversarial Spheres—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionUnderstanding and Enhancing the Transferability of Adversarial Examples—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAre adversarial examples inevitable?—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAdversarially Robust Generalization Requires More Data—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionRobustness May Be at Odds with Accuracy—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionWhy Do Adversarial Attacks Transfer? Explaining Transferability of Evasion and P—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionOn the Geometry of Adversarial Examples—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionOn the Effectiveness of Interval Bound Propagation for Training Verifiably Robus—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionThe Curse of Concentration in Robust Learning: Evasion and Poisoning Attacks fro10.1609/aaai.v33i01.330145
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionOn the Effectiveness of Low Frequency Perturbations10.24963/ijcai.2019/470
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAn Information-Theoretic Explanation for the Adversarial Fragility of AI Classif10.1109/isit.2019.8849757
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionTheoretically Principled Trade-off between Robustness and Accuracy—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAdversarial Robustness vs. Model Compression, or Both?10.1109/iccv.2019.00020
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionCertified Adversarial Robustness via Randomized Smoothing—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionModel Compression with Adversarial Robustness: A Unified Optimization Framework—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAdversarial Examples Are Not Bugs, They Are Features10.48550/arxiv.1905.02175
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionEmpirically Measuring Concentration: Fundamental Limits on Intrinsic Robustness—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAdversarial Training for Free!—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionUnlabeled Data Improves Adversarial Robustness—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAdversarial Robustness as a Prior for Learned Representations—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionUnderstanding and Mitigating the Tradeoff Between Robustness and Accuracy—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionPrecise Tradeoffs in Adversarial Training for Linear Regression—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionOverfitting in adversarially robust deep learning—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionReliable evaluation of adversarial robustness with an ensemble of diverse parame—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionDo Adversarially Robust ImageNet Models Transfer Better?—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionDerivation of Information-Theoretically Optimal Adversarial Attacks with Applica10.1109/ieeeconf60004.2024
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionAdversarial Robustness through Disentangled Representations10.1609/aaai.v35i4.16424
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionThe Dimpled Manifold Model of Adversarial Examples in Machine Learning—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionToy Models of Superposition10.48550/arxiv.2209.10652
active
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionUnderstanding Adversarial Robustness Against On-manifold Adversarial Examples10.48550/arxiv.2210.00430
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionWhich Models have Perceptually-Aligned Gradients? An Explanation via Off-Manifol10.48550/arxiv.2305.19101
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionSparse autoencoders find highly interpretable features in language models10.48550/arxiv.2309.08600
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionA High Dimensional Statistical Model for Adversarial Training: Geometry and Trad—
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionExploring the Interplay of Interpretability and Robustness in Deep Neural Networ10.1109/icipcw64161.2024.1
referenced-only
s2-reference-walk
Adversarial Examples Are Not Bugs, They Are SuperpositionOn the Complexity of Neural Computation in Superposition10.48550/arxiv.2409.15318
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of What is a linear representation? What is a multidimensional feature?—
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Towards monosemanticity: Decomposing language models with dictionary learning—
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Up-date on how we train saes—
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of The dark matter of neural networks? Transformer Circuits Thread—
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Sparse Crosscoders for Cross-Layer Features and Model Diffing—
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Toy Models of Superposition10.48550/arxiv.2209.10652
active
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Robustly identifying concepts introduced during chat fine-tuning using crosscode—
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Training Compute-Optimal Large Language Models10.52202/068431-2176
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Scaling Laws for Neural Language Models—
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Sparse autoencoders find highly interpretable features in language models10.48550/arxiv.2309.08600
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Not all language model features are one-dimensionally linear—
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of The Quantization Model of Neural Scaling10.48550/arxiv.2303.13506
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencode10.48550/arxiv.2407.14435
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Scaling and evaluating sparse autoencoders10.48550/arxiv.2406.04093
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Transcoders Find Interpretable LLM Feature Circuits10.48550/arxiv.2406.11944
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Recurrent neural networks learn to store and generate sequences using non-linear10.48550/arxiv.2408.10920
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoenc10.48550/arxiv.2409.14507
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 210.48550/arxiv.2408.05147
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Batchtopk sparse autoencoders10.48550/arxiv.2412.06410
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Group Crosscoders for Mechanistic Analysis of Symmetry10.48550/arxiv.2410.24184
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Decomposing The Dark Matter of Sparse Autoencoders10.48550/arxiv.2410.14670
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geom10.48550/arxiv.2503.01822
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Neural Scaling Laws Rooted in the Data Distribution10.48550/arxiv.2412.07942
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Archetypal sae: Adaptive and stable dictionary learning for concept extraction i10.48550/arxiv.2502.12892
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures10.48550/arxiv.2506.01197
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of From flat to hierarchical: Extracting sparse representations with matching pursu10.48550/arxiv.2506.03093
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of The origins of representation manifolds in large language models10.48550/arxiv.2505.18235
referenced-only
s2-reference-walk
Understanding sparse autoencoder scaling in the presence of Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet—
active
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiSymbols Among the Neurons: Details of a Connectionist Inference Architecture—
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiRecency and primacy in persuasion as a function of the timing of speeches and me10.1037/h0049330
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiInvestigating Gender Bias in Language Models Using Causal Mediation Analysis—
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiTensor Product Variable Binding and the Representation of Symbolic Structures in10.7551/mitpress/2102.003.
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiLanguage Models are Few-Shot Learners—
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiConnectionism and cognitive architecture: a critical analysis.10.1016/0010-0277(88)90031
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiNeural natural language inference models partially embed theories of lexical ent10.18653/v1/2020.blackboxn
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiMemory: A Contribution to Experimental Psychology10.5214/ans.0972.7531.2004
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiCausal abstraction: A theoretical foundation for mechanistic interpretability—
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiCausal analysis of syntactic agreement mechanisms in neural language models10.18653/v1/2021.acl-long.
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiCausal abstractions of neural networks—
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiDissecting recall of factual associations in auto-regressive language models10.48550/arxiv.2304.14767
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiLost in the Middle: How Language Models Use Long Contexts10.1162/tacl_a_00638
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiDiscovering variable binding circuitry with desiderata10.48550/arxiv.2307.03637
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiSame Task, More Tokens: the Impact of Input Length on the Reasoning Performance 10.48550/arxiv.2402.14848
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiAspects of human memory and Large Language Models10.48550/arxiv.2311.03839
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiHow do language models bind entities in context?10.48550/arxiv.2310.17191
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiRepresentational Analysis of Binding in Language Models10.18653/v1/2024.emnlp-mai
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiMonitoring Latent World States in Language Models with Propositional Probes10.48550/arxiv.2406.19501
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiFine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking10.48550/arxiv.2402.14811
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiHow Do Transformers Learn Variable Binding in Symbolic Programs?10.48550/arxiv.2505.20896
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiCombining Causal Models for More Accurate Abstractions of Neural Networks10.48550/arxiv.2503.11429
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiLanguage Models use Lookbacks to Track Beliefs10.48550/arxiv.2505.14685
referenced-only
s2-reference-walk
Mixing Mechanisms: How Language Models Retrieve Bound EntitiHow causal abstraction underpins computational explanation10.48550/arxiv.2508.11214
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninTowards monosemanticity: Decomposing language models with dictionary learning—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninScaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet—
active
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context Learninlog p ( x | c ) p ( x | c ′ ) = log p ( x | c ) − log p ( x | c ′ ) ≈ γN − α ( −—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninGem-ini: a family of highly capable multimodal models—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninActivation addition: Steering language models without optimization—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninVision: A computational investigation into the human representation and processi10.1016/0022-2496(83)90030
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninMany-shot Jailbreaking10.52202/079017-4121
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninMeasuring the crowd within: Probabilistic representations within individuals—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninA Rational Analysis of Rule-Based Concept Learning10.1080/03640210701802071
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninHow to grow a mind: statistics, structure, and abstraction10.1126/science.1192788
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninEfficient Estimation of Word Representations in Vector Space—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninWhen is a crowd wise?10.1037/dec0000004
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninA Latent Variable Model Approach to PMI-based Word Embeddings10.1162/tacl_a_00106
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninGood debt or bad debt: Detecting semantic orientations in economic texts10.1002/asi.23062
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninBERT Rediscovers the Classical NLP Pipeline10.18653/v1/p19-1452
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninBayesian Models of Conceptual Development: Learning as Building Models of the Wo10.31234/osf.io/aq3rp
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninLanguage Models are Few-Shot Learners—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninScaling Laws for Neural Language Models—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninCausal abstractions of neural networks—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninProbing classifiers: Promises, shortcomings, and advances10.1162/coli_a_00422
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninPre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natu10.1145/3560815
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninChain of Thought Prompting Elicits Reasoning in Large Language Models10.52202/068431-1800
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninAn Explanation of In-context Learning as Implicit Bayesian Inference—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninRethinking the Role of Demonstrations: What Makes In-Context Learning Work?10.18653/v1/2022.emnlp-mai
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninTraining a Helpful and Harmless Assistant with Reinforcement Learning from Human10.48550/arxiv.2204.05862
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninCAT'S THEORY: Empirical Validation and Architectural Applications Cross-Architec10.48550/arxiv.2212.08073
active
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninTransformers generalize differently from information stored in context vs in wei10.48550/arxiv.2210.05675
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninCausal abstraction: A theoretical foundation for mechanistic interpretability—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninDiscovering language model behaviors with model-written evaluations10.48550/arxiv.2212.09251
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninTransformers learn in-context by gradient descent10.48550/arxiv.2212.07677
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninSparks of Artificial General Intelligence: Early experiments with GPT-4—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninA Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT10.64346/plop2023p05
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninWhat and How does In-Context Learning Learn? Bayesian Model Averaging, Parameter10.48550/arxiv.2305.19420
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninPretraining task diversity and the emergence of non-Bayesian in-context learning10.48550/arxiv.2306.15063
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninInference-time intervention: eliciting truthful answers from a language model10.52202/075280-1797
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninIn-context learning dynamics with random binary sequences10.48550/arxiv.2310.17639
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninThe Geometry of Truth: Emergent Linear Structure in Large Language Model Represe10.48550/arxiv.2310.06824
active
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninA Survey on Evaluation of Large Language Models10.1145/3641289
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninSteering Llama 2 via Contrastive Activation Addition10.48550/arxiv.2312.06681
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninThe linear representation hypothesis and the geometry of large language models10.48550/arxiv.2311.03658
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninThe mechanistic basis of data dependence and abrupt learning in an in-context cl10.48550/arxiv.2312.03002
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninA Systematic Survey of Prompt Engineering in Large Language Models: Techniques a10.48550/arxiv.2402.07927
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninIn-Context Language Learning: Architectures and Algorithms10.48550/arxiv.2401.12973
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninLLMs learn governing principles of dynamical systems, revealing an in-context ne10.48550/arxiv.2402.00795
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninLanguage Models Represent Beliefs of Self and Others10.48550/arxiv.2402.18496
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninFoundational Challenges in Assuring Alignment and Safety of Large Language Model10.48550/arxiv.2404.09932
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninThe Geometry of Categorical and Hierarchical Concepts in Large Language Models10.48550/arxiv.2406.01506
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninMany-Shot In-Context Learning10.48550/arxiv.2404.11018
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninDesigning a Dashboard for Transparency and Control of Conversational AI10.48550/arxiv.2406.07882
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninMultilevel Interpretability Of Artificial Neural Networks: Leveraging Framework 10.48550/arxiv.2408.12664
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninThe llama 3 herd of models10.48550/arxiv.2407.21783
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninRefusal in language models is mediated by a single direction10.48550/arxiv.2406.11717
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninExploring the Personality Traits of LLMs through Latent Features Steering—
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninSparse Autoencoders Reveal Temporal Difference Learning in Large Language Models10.48550/arxiv.2410.01280
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninIn-Context Learning Enables Robot Action Prediction in LLMs10.1109/icra55743.2025.111
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninBayesian scaling laws for in-context learning10.48550/arxiv.2410.16531
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninTowards Reliable Evaluation of Behavior Steering Interventions in LLMs10.48550/arxiv.2410.17245
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninCompetition Dynamics Shape Algorithmic Phases of In-Context Learning10.48550/arxiv.2412.01003
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninDifferential learning kinetics govern the transition from memorization to genera10.48550/arxiv.2412.00104
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninThe broader spectrum of in-context learning10.48550/arxiv.2412.03782
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninIclr: In-context learning of representations10.48550/arxiv.2501.00070
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninIn-Context Learning Strategies Emerge Rationally10.48550/arxiv.2506.17859
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninPersona Vectors: Monitoring and Controlling Character Traits in Language Models10.48550/arxiv.2507.21509
active
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninAre language models aware of the road not taken? Token-level uncertainty and hid10.48550/arxiv.2511.04527
referenced-only
s2-reference-walk
Belief Dynamics Reveal the Dual Nature of In-Context LearninEmergence of Linear Truth Encodings in Language Models10.48550/arxiv.2510.15804
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaOn memorization and privacy risks of sharpness aware minimization—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaFlat Minima10.1162/neco.1997.9.1.1
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaFact finding: How to think about interpreting memorisation (post 4) — AI—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaOptimal Brain Damage—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaUnderstanding Memorization in Generative Models via Sharpness in Probability Lan10.48550/arxiv.2412.04140
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaImageNet Large Scale Visual Recognition Challenge10.1007/s11263-015-0816-y
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaOptimal Brain Surgeon and general network pruning10.1109/icnn.1993.298572
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaOn Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaOptimizing Neural Networks with Kronecker-factored Approximate Curvature—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaMembership Inference Attacks Against Machine Learning Models10.1109/sp.2017.41
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaSharpness-Aware Minimization for Efficiently Improving Generalization—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaAn Investigation into Neural Net Optimization via Hessian Eigenvalue Density—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaThe Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Netw—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaUnderstanding deep learning requires rethinking generalization—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaKnowledge Neurons in Pretrained Transformers10.18653/v1/2022.acl-long.
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaAn Image is Worth 16x16 Words: Transformers for Image Recognition at Scale—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaTransformer feed-forward layers are key-value memories10.18653/v1/2021.emnlp-mai
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaDoes Localization Inform Editing? Surprising Differences in Causality-Based Loca10.48550/arxiv.2301.04213
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaLocating and editing factual associations in GPT10.52202/068431-1262
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaQuantifying Memorization Across Neural Language Models—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaLanguage Models Implement Simple Word2Vec-style Vector Arithmetic10.48550/arxiv.2305.16130
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaMemorization Through the Lens of Curvature of Loss Function Around Samples10.48550/arxiv.2307.05831
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaCan Neural Network Memorization Be Localized?10.48550/arxiv.2307.09542
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaEmbers of Autoregression: Understanding Large Language Models Through the Proble10.48550/arxiv.2309.13638
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaLinearity of Relation Decoding in Transformer Language Models10.48550/arxiv.2308.09124
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaCompressing LLMs: The Truth is Rarely Pure and Never Simple10.48550/arxiv.2310.01382
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaDo Localization Methods Actually Localize Memorized Data in LLMs? A Tale of Two 10.48550/arxiv.2311.09060
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaCopyright Violations and Large Language Models10.48550/arxiv.2310.13771
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaThe Truth is in There: Improving Reasoning in Language Models with Layer-Selecti10.48550/arxiv.2312.13558
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaGaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection10.48550/arxiv.2403.03507
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaUnveiling Privacy, Memorization, and Input Curvature Links10.48550/arxiv.2402.18726
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaLocalizing Paragraph Memorization in Language Models10.48550/arxiv.2403.19851
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaUsing Degeneracy in the Loss Landscape for Mechanistic Interpretability10.48550/arxiv.2405.10927
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaFrom Low Rank Gradient Subspace Stabilization to Low-Rank Weights: Observations,—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaOLMES: A Standard for Language Model Evaluations10.48550/arxiv.2406.08446
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaDemystifying Verbatim Memorization in Large Language Models10.48550/arxiv.2407.17817
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaArithmetic without algorithms: language models solve math with a bag of heuristi10.48550/arxiv.2410.21272
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaApproaching Deep Learning through the Spectral Dynamics of Weights10.48550/arxiv.2408.11804
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaMitigating Memorization In Language Models10.48550/arxiv.2410.02159
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaDifferential learning kinetics govern the transition from memorization to genera10.48550/arxiv.2412.00104
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaMeasuring Non-Adversarial Reproduction of Training Data in Large Language Models10.48550/arxiv.2411.10242
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaOLMo 2 furious10.48550/arxiv.2501.00656
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaLanguage models use trigonometry to do addition10.48550/arxiv.2502.00873
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaAnalyzing Memorization in Large Language Models through the Lens of Model Attrib10.48550/arxiv.2501.05078
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaStructural Inference: Interpreting Small Language Models with Susceptibilities—
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaPrecise In-Parameter Concept Erasure in Large Language Models10.48550/arxiv.2505.22586
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaOn linear representations and pretraining data frequency in language models10.48550/arxiv.2504.12459
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaHow much do language models memorize?10.48550/arxiv.2505.24832
referenced-only
s2-reference-walk
From Memorization to Reasoning in the Spectrum of Loss CurvaStochastic Parameter Decomposition10.48550/arxiv.2506.20790
active
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Neural manifold geometry encodes feature fields—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Deploying interpretability to production with rakuten: Sae probes for pii detect—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Political bias dataset: A synthetic dataset for bias detection and reduction—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Receptive fields, binocular interaction and functional architecture in the cat's10.1113/jphysiol.1962.sp00
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?A Global Geometric Framework for Nonlinear Dimensionality Reduction—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Towards monosemanticity: Decomposing language models with dictionary learning—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Covariance selection—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Global Versus Local Methods in Nonlinear Dimensionality Reduction—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Spatial Interaction and the Statistical Analysis of Lattice Systems10.1111/j.2517-6161.1974.t
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Optimally sparse representation in general (nonorthogonal) dictionaries via l mi—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Narrow Versus Wide Tuning Curves: What's Best for a Population Code?10.1162/089976699300016818
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Monology—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Finding the tree of life in evo 2—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Diffusion maps10.1016/j.acha.2006.04.006
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?linear subspace clustering: a comparative study of nonlinear manifold clustering—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Using interpretability to identify a novel class of biomarkers for alzheimer’s d—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Receptive fields and functional architecture of monkey striate cortex10.1113/jphysiol.1968.sp00
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Information Theory and Statistical Mechanics10.1103/physrev.106.620
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Beitrag zur Theorie des Ferromagnetismus10.1007/bf02980577
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The hippocampus as a spatial map. Preliminary evidence from unit activity in the10.1016/0006-8993(71)90358
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Information processing with population codes10.1038/35039062
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Neuronal population coding of movement direction.10.1126/science.3749885
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Sparse coding with an overcomplete basis set: a strategy employed by V1?10.1016/s0042-6989(97)0016
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Emergence of simple-cell receptive field properties by learning a sparse code fo10.1038/381607a0
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Nonlinear dimensionality reduction by locally linear embedding.10.1126/science.290.5500.2
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Laplacian Eigenmaps and Spectral Techniques for Embedding and Clustering10.7551/mitpress/1120.003.
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?GRASSMANNIAN FRAMES WITH APPLICATIONS TO CODING AND COMMUNICATION10.1016/s1063-5203(03)0002
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Hessian eigenmaps: Locally linear embedding techniques for high-dimensional data10.1073/pnas.1031596100
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Greed is good: algorithmic results for sparse approximation10.1109/tit.2004.834793
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Principal manifolds and nonlinear dimensionality reduction via tangent space ali10.1007/s11741-004-0051-1
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Just relax: convex programming methods for identifying sparse signals in noise10.1109/tit.2005.864420
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Weak pairwise correlations imply strongly correlated network states in a neural 10.1038/nature04701
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Tuning Curves, Neuronal Variability, and Sensory Coding10.1371/journal.pbio.00400
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?High-dimensional graphs and variable selection with the Lasso10.1214/009053606000000281
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Sparse inverse covariance estimation with the graphical lasso.10.1093/biostatistics/kxm0
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Finding the Homology of Submanifolds with High Confidence from Random Samples10.1007/s00454-008-9053-2
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Neuronal couplings between retinal ganglion cells inferred by efficient inverse 10.1073/pnas.0906705106
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Graphical Models, Exponential Families, and Variational Inference10.1561/2200000001
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Nonlinear Learning using Local Coordinate Coding—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Locality-constrained Linear Coding for image classification10.1109/cvpr.2010.5540018
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Robust Subspace Segmentation by Low-Rank Representation—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Sparse Manifold Clustering and Embedding—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Robust Recovery of Subspace Structures by Low-Rank Representation10.1109/tpami.2012.88
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Sparse Subspace Clustering: Algorithm, Theory, and Applications—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Robust Subspace Clustering10.1214/13-aos1199
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Kernel sparse subspace clustering10.1109/icip.2014.7025576
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Locally Linear Landmarks for Large-Scale Manifold Learning10.1007/978-3-642-40994-3_
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Efficient Estimation of Word Representations in Vector Space—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?On the Number of Linear Regions of Deep Neural Networks—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Scalable Sparse Subspace Clustering by Orthogonal Matching Pursuit10.1109/cvpr.2016.425
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?A new method for constructing networks from binary data10.1038/srep05918
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Representation Benefits of Deep Feedforward Networks—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Linear algebraic structure of word senses, with applications to polysemy10.1162/tacl_a_00034
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?On the Expressive Power of Deep Neural Networks—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?An Empirical Analysis of Formality in Online Communication10.1162/tacl_a_00083
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Barlow versus Hebb: When is it time to abandon the notion of feature detectors a10.1016/j.neulet.2017.04.0
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Noisy Subspace Clustering via Matching Pursuits10.1109/tit.2018.2812824
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Structured Sparse Subspace Clustering: A Joint Affinity Learning and Subspace Cl10.1109/tip.2017.2691557
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Classification and Geometry of General Perceptual Manifolds10.1103/physrevx.8.031003
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Exact and Consistent Interpretation for Piecewise Linear Neural Networks: A Clos10.1145/3219819.3220063
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Bounding and Counting Linear Regions of Deep Neural Networks—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Deep Subspace Clustering Networks—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The Sparse Manifold Transform—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Mad Max: Affine Spline Insights Into Deep Learning10.1109/jproc.2020.3042100
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?A Spline Theory of Deep Learning—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Empirical Studies on the Properties of Linear Regions in Deep Neural Networks—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Separability and geometry of object manifolds in deep neural networks10.1038/s41467-020-14578-5
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Complexity of Linear Regions in Deep Networks—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Graphical Models10.5860/choice.34-5144
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Coding—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?A Critique of Self-Expressive Deep Subspace Clustering—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The Pile: An 800GB Dataset of Diverse Text for Language Modeling—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Probing classifiers: Promises, shortcomings, and advances10.1162/coli_a_00422
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Neural population geometry: An approach for understanding biological and artific10.1016/j.conb.2021.10.010
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Toy Models of Superposition10.48550/arxiv.2209.10652
active
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Attractor and integrator networks in the brain10.1038/s41583-022-00642-0
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Neural Manifold Clustering and Embedding—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?On convex decision regions in deep network representations10.1038/s41467-025-60809-y
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Interpreting Neural Networks through the Polytope Lens10.48550/arxiv.2211.12312
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Geometric Sparse Coding in Wasserstein Space10.48550/arxiv.2210.12135
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Sparse autoencoders find highly interpretable features in language models10.48550/arxiv.2309.08600
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The linear representation hypothesis and the geometry of large language models10.48550/arxiv.2311.03658
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?A Holistic Approach to Unifying Automatic Concept Extraction and Concept Importa10.48550/arxiv.2306.07304
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The Missing Curve Detectors of InceptionV1: Applying Sparse Autoencoders to Ince10.48550/arxiv.2406.03662
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?On the Origins of Linear Representations in Large Language Models10.48550/arxiv.2403.03867
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The Geometry of Categorical and Hierarchical Concepts in Large Language Models10.48550/arxiv.2406.01506
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Not all language model features are one-dimensionally linear—
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Scaling and evaluating sparse autoencoders10.48550/arxiv.2406.04093
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Relational Composition in Neural Networks: A Survey and Call to Action10.48550/arxiv.2407.14662
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Measuring Progress in Dictionary Learning for Language Model Interpretability wi10.48550/arxiv.2408.00113
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencode10.48550/arxiv.2407.14435
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Recurrent neural networks learn to store and generate sequences using non-linear10.48550/arxiv.2408.10920
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?A polar coordinate system represents syntax in large language models10.48550/arxiv.2412.05571
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Towards unifying interpretability and control: Evaluation via intervention10.48550/arxiv.2411.04430
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoenc10.48550/arxiv.2409.14507
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Batchtopk sparse autoencoders10.48550/arxiv.2412.06410
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Sparse Autoencoders Trained on the Same Data Learn Different Features10.48550/arxiv.2501.16615
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Open Problems in Mechanistic Interpretability10.48550/arxiv.2501.16496
active
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Language models use trigonometry to do addition10.48550/arxiv.2502.00873
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders10.48550/arxiv.2501.17148
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Closed-Form Training Dynamics Reveal Learned Features and Linear Structure in Wo10.52202/085713-0140
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Are Sparse Autoencoders Useful? A Case Study in Sparse Probing10.48550/arxiv.2502.16681
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The Geometry of Refusal in Large Language Models: Concept Cones and Representati10.48550/arxiv.2502.17420
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Archetypal sae: Adaptive and stable dictionary learning for concept extraction i10.48550/arxiv.2502.12892
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Learning Multi-Level Features with Matryoshka Sparse Autoencoders10.48550/arxiv.2503.17547
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geom10.48550/arxiv.2503.01822
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?From superposition to sparse codes: interpretable representations in neural netw10.48550/arxiv.2503.01824
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?On the emergence of linear analogies in word embeddings10.48550/arxiv.2505.18651
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The origins of representation manifolds in large language models10.48550/arxiv.2505.18235
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?From flat to hierarchical: Extracting sparse representations with matching pursu10.48550/arxiv.2506.03093
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Understanding sparse autoencoder scaling in the presence of feature manifolds10.48550/arxiv.2509.02565
active
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders10.48550/arxiv.2507.23220
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Finding Manifolds With Bilinear Autoencoders10.48550/arxiv.2510.16820
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Into the rabbit hull: From task-relevant concepts in dino to minkowski geometry10.48550/arxiv.2510.08638
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?A circuit for predicting hierarchical structure in-context in Large Language Mod10.48550/arxiv.2509.21534
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?When Models Manipulate Manifolds: The Geometry of a Counting Task10.48550/arxiv.2601.04480
active
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Priors in Time: Missing Inductive Biases for Language Model Interpretability10.48550/arxiv.2511.01836
active
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for I10.48550/arxiv.2511.05541
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Cross-Modal Redundancy and the Geometry of Vision-Language Embeddings10.48550/arxiv.2602.06218
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The Shape of Beliefs: Geometry, Dynamics, and Interventions along Representation10.48550/arxiv.2602.02315
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Context Structure Reshapes the Representational Geometry of Language Models10.48550/arxiv.2601.22364
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?Symmetry in language statistics shapes the geometry of model representations10.48550/arxiv.2602.15029
referenced-only
s2-reference-walk
Do Sparse Autoencoders Capture Concept Manifolds?The Information Geometry of Softmax: Probing and Steering10.48550/arxiv.2602.15293
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereSystem card: Claude Mythos preview—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereThe Principles of Learning on Multiple Tasks—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereThe RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Da—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterferePareto frontiers in deep feature learning: Data, compute, width, and luck10.52202/075280-2082
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereEffect of scale on catastrophic forgetting in neural networks—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereFoundation Models from a Data-Distributional View—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereOn a Theorem of Weyl Concerning Eigenvalues of Linear Transformations I.—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereDeepSeek-V3 Technical Report—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereNatural Image Statistics - A Probabilistic Approach to Early Computational Visio10.1007/978-1-84882-491-1
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereMultitask Learning10.1023/a:1007379606734
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereUncovering Latent Memories in Large Language Models—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereThe Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Netw—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereDeep Learning Scaling is Predictable, Empirically—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereZipf’s word frequency law in natural language: A critical review and future dire10.3758/s13423-014-0585-6
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereGradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask N—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereThe Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereGradient based sample selection for online continual learning—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereMulti-Task Learning as Multi-Objective Optimization—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereRegularizing Deep Multi-Task Networks using Orthogonal Gradients—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereDoes learning require memorization? a short tale about a long tail10.1145/3357713.3384290
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereThe Generalization Error of Random Features Regression: Precise Asymptotics and 10.1002/cpa.22008
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereContinual Learning in Neural Networks—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereOrthogonal Gradient Descent for Continual Learning—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereA Constructive Prediction of the Generalization Error Across Scales—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereDeep double descent: where bigger models and more data hurt10.1088/1742-5468/ac3a74
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereGradient Surgery for Multi-Task Learning—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereMore Data Can Hurt for Linear Regression: Sample-wise Double Descent—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereExploring the Limits of Transfer Learning with a Unified Text-to-Text Transforme10.48550/arxiv.1910.10683
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereOptimal Regularization Can Mitigate Double Descent—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereA brief prehistory of double descent10.1073/pnas.2001875117
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereProving the Lottery Ticket Hypothesis: Pruning is All You Need—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereScaling Laws for Neural Language Models—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereOptimal Lottery Tickets via SubsetSum: Logarithmic Over-Parameterization is Suff—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereGradient Starvation: A Learning Proclivity in Neural Networks—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereScaling Laws for Autoregressive Generative Modeling—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereExplaining neural scaling laws10.1073/pnas.2311878121
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereScaling Laws for Transfer—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereWhen is memorization of irrelevant training data necessary for high-accuracy lea10.1145/3406325.3451131
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereSaddle-to-Saddle Dynamics in Deep Linear Networks: Small Initialization Training—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereConflict-Averse Gradient Descent for Multi-task Learning—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereScaling Language Models: Methods, Analysis & Insights from Training Gopher—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereQuantifying Memorization Across Neural Language Models—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereGrokking: Generalization Beyond Overfitting on Small Algorithmic Datasets—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereTraining Compute-Optimal Large Language Models10.52202/068431-2176
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereMemorize to Generalize: on the Necessity of Interpolation in High Dimensional Li—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereScaling Laws and Interpretability of Learning from Repeated Data10.48550/arxiv.2205.10487
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereHow catastrophic can catastrophic forgetting be in linear regression?10.48550/arxiv.2205.09588
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereEmergent Abilities of Large Language Models10.48550/arxiv.2206.07682
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereMemorization Without Overfitting: Analyzing the Training Dynamics of Large Langu10.48550/arxiv.2205.10770
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereMeasuring Forgetting of Memorized Training Examples10.48550/arxiv.2207.00099
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereRevisiting Neural Scaling Laws in Language and Vision10.48550/arxiv.2209.06640
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereA Solvable Model of Neural Scaling Laws10.48550/arxiv.2210.16859
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereA picture of the space of typical learnable tasks10.48550/arxiv.2210.17011
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereLarge Language Models Struggle to Learn Long-Tail Knowledge—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereReduce, Reuse, Recycle: Improving Training Efficiency with Distillation10.48550/arxiv.2211.00683
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereProgress measures for grokking via mechanistic interpretability10.48550/arxiv.2301.05217
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereData Selection for Language Models via Importance Resampling10.48550/arxiv.2302.03169
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereSGD learning on neural networks: leap complexity and saddle-to-saddle dynamics10.48550/arxiv.2302.11055
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereFinding Alignments Between Interpretable Causal Variables and Distributed Neural10.48550/arxiv.2303.02536
active
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereDoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining10.48550/arxiv.2305.10429
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereThe Quantization Model of Neural Scaling10.48550/arxiv.2303.13506
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereLarger language models do in-context learning differently10.48550/arxiv.2303.03846
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereGrokking of Hierarchical Structure in Vanilla Transformers10.48550/arxiv.2305.18741
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereFeature-Learning Networks Are Consistent Across Widths At Realistic Scales10.48550/arxiv.2305.18411
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereOn-Policy Distillation of Language Models: Learning from Self-Generated Mistakes—
referenced-only
s2-reference-walk
Why Larger Models Learn More: Effects of Capacity, InterfereExplaining grokking through circuit efficiency10.48550/arxiv.2309.02390
referenced-only
s2-reference-walk