概念解释论文趋势 2023–2026

主要结论

• 概念解释论文从 2023 年的 91 篇增长到 2026 年的 117 篇，三年间扩大到约 1.3 倍。
• NeurIPS 是发概念解释论文最多的会议，2023–2026 年共 134 篇，占四个会议总量的 39%。
• 交叉方向里涨得最快的是多模态大模型：同时属于两个主题的论文从 2023 年的 4 篇增至 2026 年的 23 篇。
• 在「可解释性」的 5 个主题里，它的 2026 年论文数排第 3；同组增长最快的机制解释是 8.1 倍，概念解释是 1.3 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 2026CENDRe: Concept Extraction with Natural Domain Representations
• ICML 2026Evidential Copula Concept Embedding Models
• ICLR 2026Decoupling the Class Label and the Target Concept in Machine Unlearning
• CVPR 2026Agile Deliberation Concept Deliberation for Subjective Visual Classification
• NeurIPS 2025An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations
• ICML 2025Concept-Centric Token Interpretation for Vector-Quantized Generative Models
• ICLR 2025Causal Concept Graph Models: Beyond Causal Opacity in Deep Learning
• CVPR 2025Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models
• NeurIPS 2024[Re] On the Reproducibility of Post-Hoc Concept Bottleneck Models
• ICML 2024ContPhy: Continuum Physical Concept Learning and Reasoning from Videos
• ICLR 2024AttEXplore: Attribution for Explanation with model parameters eXploration
• CVPR 2024Fine-Grained Bipartite Concept Factorization for Clustering
• NeurIPS 2023[Re] Hierarchical Shrinkage: Improving the Accuracy and Interpretability of Tree-Based Methods
• ICML 2023A Closer Look at the Intervention Procedure of Concept Bottleneck Models
• ICLR 2023Concept Gradient: Concept-based Interpretation Without Linear Assumption
• CVPR 2023CRAFT: Concept Recursive Activation FacTorization for Explainability
• NeurIPS 2026Concept frustration: Aligning human concepts and machine representations
• ICML 2026On the Variability of Concept Activation Vectors
• ICLR 2026Hierarchical Concept-based Interpretable Models
• CVPR 2026DCoAR Deep Concept Injection into Unified Autoregressive Models for Personalized

全部概念解释论文

共 40 篇，第 1 / 1 页

1. NeurIPS 2026A Theoretical Analysis of Why Masked Diffusion Models Mitigate the Reversal Curse
2. NeurIPS 2026CENDRe: Concept Extraction with Natural Domain Representations
3. NeurIPS 2026Concept frustration: Aligning human concepts and machine representations
4. NeurIPS 2026Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations
5. NeurIPS 2026Concise and Logically Consistent Conformal Sets for Neuro-Symbolic Concept-Based Models
6. NeurIPS 2026Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models
7. NeurIPS 2026Diffusion-Time Concept Manifolds: Sparse Autoencoder Groups for Interpreting Denoising Language Models
8. NeurIPS 2026Distribution Free Fourier Sparsity Testing
9. NeurIPS 2026Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?
10. NeurIPS 2026Do You CARE to Generalize? Extracting Robust Concept Directions from LLMs
11. NeurIPS 2026Example-Based Spatial Guidance for Training-Free Concept Erasure in Diffusion Models
12. NeurIPS 2026Explanation Multiplicity in SHAP: Characterization and Assessment
13. NeurIPS 2026Explanation of Dynamic Physical Field Predictions using WassersteinGrad: Application to Autoregressive Weather Forecasting
14. NeurIPS 2026Few-shot Task Learning via Compositional Concept Inference
15. NeurIPS 2026From Zero to Hero: Training-Free Custom Concept Spawning in World Models
16. NeurIPS 2026Hierarchical Concept Geometry in Language Representations Emerges from Word Co-occurrence
17. NeurIPS 2026Hyperbolic Concept Bottleneck Models
18. NeurIPS 2026Introspective Coupling: LMs Learn to Explain Themselves Better Than Their Training Targets
19. NeurIPS 2026Latent Introspection: Models Can Detect Prior Concept Injections
20. NeurIPS 2026Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning
21. NeurIPS 2026Marginal-Nonuniform Multiclass Learning
22. NeurIPS 2026Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models
23. NeurIPS 2026MOSAIC: Concept Bottlenecks via Text-Anchored Optimal Transport
24. NeurIPS 2026NeuroFaith: Evaluating Mechanistic Faithfulness of LLM Free Text Self-Explanation at the Concept Level
25. NeurIPS 2026NORMA: Norm-Guided Explanation Subgraph Discovery
26. NeurIPS 2026On the Provable Emergence of Hierarchical Concept Structure in CLIP Embeddings
27. NeurIPS 2026PLACE: Patch-Level Agnostic Concept Extraction
28. NeurIPS 2026Position Without Positional Embeddings: A Directional Mechanism in NoPE Transformers
29. NeurIPS 2026Position-aware eXplanation: A Model-Agnostic Framework for Positional Attributions
30. NeurIPS 2026Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants
31. NeurIPS 2026Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series
32. NeurIPS 2026Second-Order Complexity Theory for Risk, Explanation, and Calibration in Machine Learning
33. NeurIPS 2026Settling the Sample Complexity of Deterministic Agnostic PAC Learning
34. NeurIPS 2026Simultaneous Gradient Learning in First-Price Auctions
35. NeurIPS 2026TanGCE: Manifold-Aware Concept Erasure
36. NeurIPS 2026The Quiet Prompt: Erasing Ineffable Styles from Diffusion Models via Concept Leakage-aware Negative Guidance
37. NeurIPS 2026The SuperActivator Mechanism: Transformers Concentrate Reliable Concept Signals in the Tail
38. NeurIPS 2026TRACER: Token ReAssignment for Concept ERasure in Generative Recommendation
39. NeurIPS 2026Weakly Supervised Concept Learning for Interpreting and Attributing LVLM Predictions
40. NeurIPS 2026You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning

常见问题

2023–2026 年四大会议一共有多少篇概念解释论文？

共 342 篇，按年份分别是 2023 年 91 篇、2024 年 63 篇、2025 年 71 篇、2026 年 117 篇。

NeurIPS 2026 有多少篇概念解释论文？

NeurIPS 2026 有 40 篇概念解释论文，上一届是 25 篇。

ICLR 2026 有多少篇概念解释论文？

ICLR 2026 有 24 篇概念解释论文，上一届是 19 篇。

ICML 2026 有多少篇概念解释论文？

ICML 2026 有 20 篇概念解释论文，上一届是 11 篇。

CVPR 2026 有多少篇概念解释论文？

CVPR 2026 有 33 篇概念解释论文，上一届是 16 篇。

哪个会议的概念解释论文最多？

NeurIPS，2023–2026 年共 134 篇。