特征归因论文趋势 2023–2026

主要结论

• 特征归因论文从 2023 年的 118 篇增长到 2026 年的 219 篇，三年间扩大到约 1.9 倍。
• NeurIPS 是发特征归因论文最多的会议，2023–2026 年共 193 篇，占四个会议总量的 28%。
• 交叉方向里涨得最快的是理论 · learning theory：同时属于两个主题的论文从 2023 年的 1 篇增至 2026 年的 14 篇。
• 在「可解释性」的 5 个主题里，它的 2026 年论文数排第 1；同组增长最快的机制解释是 8.1 倍，特征归因是 1.9 倍。

代表论文

按年份和会议均匀挑选，方便快速看到每一届的样貌

• NeurIPS 2026ADAPT: Hybrid Prompt Optimization for LLM Feature Visualization
• ICML 2026Activation with Intrinsic-Extrinsic Consensus
• ICLR 2026AnyUp: Universal Feature Upsampling
• CVPR 2026A Training-Free Style-Personalization via SVD-Based Feature Decomposition
• NeurIPS 2025Alternating Gradient Flows: A Theory of Feature Learning in Two-layer Neural Networks
• ICML 2025BCE vs. CE in Deep Feature Learning
• ICLR 2025A Differentiable Rank-Based Objective for Better Feature Learning
• CVPR 2025A Simple Data Augmentation for Feature Distribution Skewed Federated Learning
• NeurIPS 2024$\texttt{dattri}$: A Library for Efficient Data Attribution
• ICML 2024Asymptotics of feature learning in two-layer networks after one gradient-step
• ICLR 2024An Intuitive Multi-Frequency Feature Representation for SO(3)-Equivariant Networks
• CVPR 2024DYSON: Dynamic Feature Space Self-Organization for Online Task-Free Class Incremental Learning
• NeurIPS 2023$\mathcal{M}^4$: A Unified XAI Benchmark for Faithfulness Evaluation of Feature Attribution Methods across Metrics, Modalities and Models
• ICML 2023Decoding Layer Saliency in Language Transformers
• ICLR 2023Can We Faithfully Represent Absence States to Compute Shapley Values on a DNN?
• CVPR 2023A Practical Upper Bound for the Worst-Case Attribution Deviations
• NeurIPS 2026Adaptive Feature Propagation for Attribute-Missing Graph Clustering
• ICML 2026Aggregate Models, Not Explanations: Improving Feature Importance Estimation
• ICLR 2026Attribution-Guided Decoding
• CVPR 2026FBTA Enabling Single-GPU End-to-End Gigapixel WSI Classification with Feature Bridging

全部特征归因论文

共 219 篇，第 3 / 5 页

101. ICLR 2026Watermark-based Attribution of AI-Generated Content
102. ICLR 2026xRFM: Accurate, scalable, and interpretable feature learning models for tabular data
103. ICML 2026$\texttt{ShaplEIG}$: Bayesian Experimental Design for Shapley Value Estimation
104. ICML 2026A unified theory of feature learning in RNNs and DNNs
105. ICML 2026Activation with Intrinsic-Extrinsic Consensus
106. ICML 2026Aggregate Models, Not Explanations: Improving Feature Importance Estimation
107. ICML 2026Alignment-Sensitive Minimax Rates for Spectral Algorithms with Learned Kernels
108. ICML 2026Barriers to Counterfactual Credit Attribution for Autoregressive Models
109. ICML 2026BASIL: Scalable Bayesian Semi-supervised Clustering with Feature Selection and Adaptive Constraint Weighting
110. ICML 2026Beyond Trajectory-Level Attribution: Graph-Based Credit Assignment for Agentic Reinforcement Learning
111. ICML 2026C$^{2}$R: Cross-sample Consistency Regularization Mitigates Feature Splitting and Absorption in Sparse Autoencoders
112. ICML 2026Causal Feature Learning via Generalized Rayleigh Quotients
113. ICML 2026ConEx: Human-Interpretable Saliency Maps via Concept-Aware Attribution
114. ICML 2026DAVE: Distribution-Aware Attribution via ViT Gradient Decomposition
115. ICML 2026Decomposing Query-Key Feature Interactions Using Contrastive Covariances
116. ICML 2026Detecting and Filtering Unsafe Training Data via Data Attribution with Denoised Representation
117. ICML 2026Dichotomy of Feature Learning and Unlearning: Fast-Slow Analysis on Neural Networks with Stochastic Gradient Descent
118. ICML 2026Diversity-Aware Recursive Feature Multiple Kernel Learning
119. ICML 2026Expand Neurons, Not Parameters
120. ICML 2026ExPLAIND: Unifying Model, Data, and Training Attribution to Study Model Behavior
121. ICML 2026Explaining Concept Shift with Interpretable Feature Attribution
122. ICML 2026Feature Bagging Provides Stability
123. ICML 2026Feature Collapse Under Corruption: An Entropy Perspective on Robust Neural Networks
124. ICML 2026Feature Resemblance: Towards a Theoretical Understanding of Analogical Reasoning in Transformers
125. ICML 2026From Parameters to Feature Space: Task Arithmetic for Backdoor Mitigation in Model Merging
126. ICML 2026GFD-EMVC: Evolutionary Multi-View Classification with Label Noise via Gradient and Feature Dual-Perception
127. ICML 2026GOTabPFN: From Feature Ordering to Compact Tokenization for Tabular Foundation Models on High-Dimensional Data
128. ICML 2026Gradient Flow Through Diagram Expansions: Learning Regimes and Explicit Solutions
129. ICML 2026GUDA: Counterfactual Group-wise Training Data Attribution for Diffusion Models via Unlearning
130. ICML 2026Hidden in Plain Sight -- Class Competition Focuses Attribution Maps
131. ICML 2026Ideal Attribution and Faithful Watermarks for Language Models
132. ICML 2026Improving Adversarial Robustness of Attribution via Implicit Regularization
133. ICML 2026Learning-To-Measure: In-Context Active Feature Acquisition
134. ICML 2026Less is Enough: Synthesizing Diverse Data in LLM Feature Space with Sparse Autoencoders
135. ICML 2026Manifold-Aligned Guided Integrated Gradients for Reliable Feature Attribution
136. ICML 2026Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units
137. ICML 2026Neural Feature Geometry Evolves as Discrete Ricci Flow
138. ICML 2026Nonparametric Data Attribution for Diffusion Models
139. ICML 2026On the Accuracy of Newton Step and Influence Function Data Attributions
140. ICML 2026On the Power of Source Screening for Learning Shared Feature Extractors
141. ICML 2026On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders
142. ICML 2026Over-Alignment vs Over-Fitting: The Role of Feature Learning Strength in Generalization
143. ICML 2026PACEAttention: Principled and Adaptive Feature Compression-Expansion Grounded in the Geometry of $\text{MCR}^2$
144. ICML 2026PLASH: Provably Linear-Time Attention with Selective Higher-Order Feature Sketching
145. ICML 2026Priority-Aware Shapley Value
146. ICML 2026QPoint: End-to-End Lightweight Point Cloud Processing via Robust Quaternion Feature Learning
147. ICML 2026Resolution as a Direction: Vector-Panning Feature Alignment for Cross-Resolution Re-Identification
148. ICML 2026Robustness of Mixtures of Experts to Feature Noise
149. ICML 2026SafeSeek: Universal Attribution of Safety Circuits in Language Models
150. ICML 2026Shapley Regularized Neural Granger Causality

常见问题

2023–2026 年四大会议一共有多少篇特征归因论文？

共 695 篇，按年份分别是 2023 年 118 篇、2024 年 168 篇、2025 年 190 篇、2026 年 219 篇。

NeurIPS 2026 有多少篇特征归因论文？

NeurIPS 2026 有 61 篇特征归因论文，上一届是 57 篇。

ICLR 2026 有多少篇特征归因论文？

ICLR 2026 有 53 篇特征归因论文，上一届是 42 篇。

ICML 2026 有多少篇特征归因论文？

ICML 2026 有 56 篇特征归因论文，上一届是 51 篇。

CVPR 2026 有多少篇特征归因论文？

CVPR 2026 有 49 篇特征归因论文，上一届是 40 篇。

哪个会议的特征归因论文最多？

NeurIPS，2023–2026 年共 193 篇。