2026 edition · Top 70
Safety, alignment & interpretability
Making AI systems understandable, robust and aligned with what people want.
Interpretability: 32 · Safety & alignment: 38 · N = 70, scaled from 171 candidates.
0 of 70 of this page's 70 read · saved in this browser only
- 1
Intriguing properties of neural networks
Christian Szegedy · 17,244 citations · Semantic Scholar
AwardEditor's note - 2
A Unified Approach to Interpreting Model Predictions
Scott M. Lundberg · 43,572 citations · Papers With Code
Editor's note - 3
“Why Should I Trust You?”: Explaining the Predictions of Any Classifier
Marco Tulio Ribeiro · 24,888 citations · Papers With Code
Editor's note - 4
Explaining and Harnessing Adversarial Examples
Ian J. Goodfellow · 23,382 citations · Papers With Code
Editor's note - 5
Training language models to follow instructions with human feedback
Long Ouyang · 25,000 citations · Papers With Code
Editor's note - 6
Axiomatic Attribution for Deep Networks
Mukund Sundararajan · 9,010 citations · Papers With Code
Editor's note - 7
Deep reinforcement learning from human preferences
Paul Christiano · 6,187 citations · Papers With Code
Editor's note - 8
Visualizing and Understanding Convolutional Networks
Matthew D. Zeiler · 17,496 citations · Papers With Code
- 9
Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization
Ramprasaath R. Selvaraju · 30,248 citations · Papers With Code
- 10
Constitutional AI: Harmlessness from AI Feedback
Yuntao Bai · 3,842 citations · Papers With Code
Editor's note - 11
- 12
Equality of Opportunity in Supervised Learning
Moritz Hardt · 5,656 citations · Semantic Scholar
Editor's note - 13
Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification
Joy Buolamwini · 5,825 citations · Semantic Scholar
Editor's note - 14
- 15
Towards Deep Learning Models Resistant to Adversarial Attacks
Aleksander Madry · 15,844 citations · Papers With Code
- 16
- 17
- 18
- 19
Deep Learning with Differential Privacy
Martín Abadi · 8,604 citations · Papers With Code
- 20
Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
Karen Simonyan · 8,625 citations · Papers With Code
- 21
Inherent Trade-Offs in the Fair Determination of Risk Scores
Jon Kleinberg · 2,210 citations · Semantic Scholar
Editor's note - 22
Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)
Been Kim · 2,563 citations · Papers With Code
- 23
Calibrating Noise to Sensitivity in Private Data Analysis
Cynthia Dwork · 7,272 citations · OpenAlex
- 24
- 25
Learning Deep Features for Discriminative Localization
Bolei Zhou · 11,337 citations · OpenAlex
- 26
The Mythos of Model Interpretability
Zachary C. Lipton · 5,424 citations · OpenAlex
- 27
Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
Tolga Bolukbasi · 3,935 citations · Papers With Code
- 28
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜
Emily M. Bender · 6,929 citations · OpenAlex
- 29
Learning Important Features Through Propagating Activation Differences
Avanti Shrikumar · 4,873 citations · Papers With Code
- 30
On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation
Sebastian Bach · 4,523 citations · OpenAlex
- 31
Towards Evaluating the Robustness of Neural Networks
Nicholas Carlini · 10,423 citations · Semantic Scholar
- 32
- 33
Differential Privacy
Cynthia Dwork · 5,877 citations · OpenAlex
- 34
Network Dissection: Quantifying Interpretability of Deep Visual Representations
David Bau · 1,859 citations · Papers With Code
- 35
Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
Bai · 4,408 citations · Papers With Code
- 36
Counterfactual Explanations Without Opening the Black Box: Automated Decisions and the GDPR
Sandra Wachter · 3,543 citations · Semantic Scholar
- 37
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
Adly Templeton · 898 citations · Semantic Scholar
- 38
Attention is not Explanation
Sarthak Jain · 2,241 citations · Semantic Scholar
- 39
Sparse Autoencoders Find Highly Interpretable Features in Language Models
Robert Huben · 1,697 citations · Papers With Code
- 40
SmoothGrad: removing noise by adding noise
Daniel Smilkov · 2,769 citations · Papers With Code
- 41
Dissecting racial bias in an algorithm used to manage the health of populations
Ziad Obermeyer · 7,104 citations · OpenAlex
- 42
Locating and Editing Factual Associations in GPT
Kevin Meng · 3,485 citations · Papers With Code
- 43
Anchors: High-Precision Model-Agnostic Explanations
Marco Tulio Ribeiro · 2,162 citations · OpenAlex
- 44
Membership Inference Attacks against Machine Learning Models
Reza Shokri · 5,781 citations · Papers With Code
- 45
TruthfulQA: Measuring How Models Mimic Human Falsehoods
Stephanie Lin · 4,109 citations · Papers With Code
- 46
Universal and Transferable Adversarial Attacks on Aligned Language Models
Andy Zou · 3,892 citations · Papers With Code
- 47
Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI
Alejandro Barredo Arrieta · 10,461 citations · OpenAlex
- 48
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Evan Hubinger · 625 citations · Papers With Code
- 49
Towards A Rigorous Science of Interpretable Machine Learning
Finale Doshi-Velez · 5,696 citations · Papers With Code
- 50
Concept Bottleneck Models
Pang Wei Koh · 1,559 citations · Papers With Code
- 51
Semantics derived automatically from language corpora contain human-like biases
Aylin Caliskan · 3,338 citations · Papers With Code
- 52
Language Models (Mostly) Know What They Know
Kadavath · 2,170 citations · Papers With Code
- 53
Learning to summarize from human feedback
Nisan Stiennon · 3,563 citations · Papers With Code
- 54
Jailbroken: How Does LLM Safety Training Fail?
Alexander Wei · 2,283 citations · Papers With Code
- 55
Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
Kevin Wang · 1,377 citations · Papers With Code
- 56
Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection
2,259 citations · Papers With Code
- 57
Understanding Neural Networks Through Deep Visualization
Jason Yosinski · 1,981 citations · Semantic Scholar
- 58
BERT Rediscovers the Classical NLP Pipeline
Ian Tenney · 2,113 citations · Papers With Code
- 59
Extracting Training Data from Large Language Models
Nicholas Carlini · 3,459 citations · Papers With Code
- 60
Fair prediction with disparate impact: A study of bias in recidivism prediction instruments
Alexandra Chouldechova · 2,158 citations · OpenAlex
- 61
Red Teaming Language Models with Language Models
Perez · 1,380 citations · Papers With Code
- 62
Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples
Anish Athalye · 2,171 citations · OpenAlex
- 63
Feature Visualization
Chris Olah · 880 citations · OpenAlex
- 64
AI safety via debate
Geoffrey Irving · 496 citations · Papers With Code
- 65
In-context Learning and Induction Heads
Catherine Olsson · 1,083 citations · Papers With Code
- 66
Ethical and Social Risks of Harm from Language Models
Weidinger · 1,685 citations · Papers With Code
- 67
Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting
Miles Turpin · 1,875 citations · Papers With Code
- 68
Are Sixteen Heads Really Better than One?
Paul Michel · 1,550 citations · Papers With Code
- 69
RISE: Randomized Input Sampling for Explanation of Black-box Models
Vitali Petsiuk · 1,682 citations · Papers With Code
- 70
Discovering Language Model Behaviors with Model-Written Evaluations
Perez · 1,121 citations · Papers With Code
Ranked on another page
Also relevant
Top 500 papers ranked on another domain page that matter here too.
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
Rafael Rafailov · 11,135 citations · Papers With Code
Editor's note
Data sources
Catalogue data from OpenAlex (CC0), Crossref (CC0 metadata; no abstracts are shown), arXiv, and Papers With Code (CC BY-SA 4.0), and 
Semantic Scholar (used under its non-commercial API licence). Each count names its source. See the method.