可微推理论文趋势 2023–2026

主要结论

• 可微推理论文从 2023 年的 31 篇增长到 2026 年的 79 篇，三年间扩大到约 2.5 倍。
• NeurIPS 是发可微推理论文最多的会议，2023–2026 年共 65 篇，占四个会议总量的 35%。
• 交叉方向里涨得最快的是数学推理：同时属于两个主题的论文从 2023 年的 3 篇增至 2026 年的 29 篇。
• 在「神经符号」的 5 个主题里，它的 2026 年论文数排第 3；同组增长最快的数学推理是 10 倍，可微推理是 2.5 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 2026Beyond MSE: Differentiable Complexity Priors for Structure-Preserving Neural Denoising
• ICML 2026Differentiable Conformal Training for LLM Reasoning Factuality
• ICLR 2026$\boldsymbol{\partial^\infty}$-Grid: A Neural Differential Equation Solver with Differentiable Feature Grids
• CVPR 2026D-Convexity A Unified Differentiable Convex Shape Prior via Quasi-Concavity for
• NeurIPS 2025Consistency Conditions for Differentiable Surrogate Losses
• ICML 2025DiLQR: Differentiable Iterative Linear Quadratic Regulator via Implicit Differentiation
• ICLR 2025Analytic DAG Constraints for Differentiable DAG Learning
• CVPR 2025Differentiable Inverse Rendering with Interpretable Basis BRDFs
• NeurIPS 2024DeiSAM: Segment Anything with Deictic Prompting
• ICML 2024A Differentiable Partially Observable Generalized Linear Model with Forward-Backward Message Passing
• ICLR 2024Generalized Neural Sorting Networks with Error-Free Differentiable Swap Functions
• CVPR 2024Differentiable Display Photometric Stereo
• NeurIPS 2023Differentiable Clustering with Perturbed Spanning Forests
• ICML 2023End-to-end Differentiable Clustering with Associative Memories
• ICLR 2023Learning Group Importance using the Differentiable Hypergeometric Distribution
• CVPR 2023Deep Learning of Partial Graph Matching via Differentiable Top-K
• NeurIPS 2026Differentiable Exact Learning of Algorithms
• ICML 2026Hide&Seek: Learning to Explain in an End-to-End Differentiable Network
• ICLR 2026Tversky Neural Networks: Psychologically Plausible Deep Learning with Differentiable Tversky Similarity
• CVPR 2026DiffBMP Differentiable Rendering with Bitmap Primitives

全部可微推理论文

共 65 篇，第 1 / 2 页

1. NeurIPS 2026Anchoring Reasoning Distillation via Syntactic Constraints
2. NeurIPS 2026Beyond MSE: Differentiable Complexity Priors for Structure-Preserving Neural Denoising
3. NeurIPS 2026Block-Wise Differentiable Sinkhorn Attention: Tail-Refinement Gradients with a Gap-Aware Dustbin Bridge
4. NeurIPS 2026Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning
5. NeurIPS 2026Can Linguistic Reasoning Vectors Enhance Multimodal Reasoning Ability?
6. NeurIPS 2026Commit Then Explore: Reasoning Models Diversify, Not Converge
7. NeurIPS 2026Convex Compositional Reasoning Models
8. NeurIPS 2026Credit Assignment with Resets in Language Model Reasoning
9. NeurIPS 2026Differentiable Belief-based Opponent Shaping
10. NeurIPS 2026Differentiable Exact Learning of Algorithms
11. NeurIPS 2026Differentiable Nonlinear Model Predictive Control
12. NeurIPS 2026Differentiable Range-Partition Entropy for Entropy-Sensitive Geometric Algorithms
13. NeurIPS 2026Diving-R1: Empowering Multimodal LLMs with Traceable Progressive Reasoning for Interpretable Diving Action Quality Assessment
14. NeurIPS 2026From Outcome to Representation: Tracing Reasoning Mechanisms through Integrated Policy Gradient
15. NeurIPS 2026Iterative Nonlinear Computation Underlying Abstract Reasoning
16. NeurIPS 2026Latent Spatial Reasoning: Building Innate 3D Awareness via Latent-Space Distillation
17. NeurIPS 2026LEVDA: Latent Ensemble Variational Data Assimilation via Differentiable Dynamics
18. NeurIPS 2026Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring
19. NeurIPS 2026Natural Synthesis: Outperforming Reactive Synthesis Tools with Large Reasoning Models
20. NeurIPS 2026Neurosymbolic Object-Centric Learning with Distant Supervision
21. NeurIPS 2026On the Token Value Inequality in Efficient Reasoning
22. NeurIPS 2026OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping
23. NeurIPS 2026Quantized Reasoning Models Think They Need to Think Longer, but They Do Not
24. NeurIPS 2026Reasoning with Undecoded Tokens in Diffusion Language Models
25. NeurIPS 2026SphMind: Towards Robust, Training-Free VLM-based Spatial Reasoning with a 360 Camera
26. NeurIPS 2026State of Thought Enables Endogenous Reasoning
27. NeurIPS 2026Strategist: Designing Agentic Reasoning at Scale
28. NeurIPS 2026VI-CuRL: Stabilizing Verifier-Independent RL Reasoning via Confidence-Guided Variance Reduction
29. NeurIPS 2025Can MLLMs Absorb Math Reasoning Abilities from LLMs as Free Lunch?
30. NeurIPS 2025Consistency Conditions for Differentiable Surrogate Losses
31. NeurIPS 2025Continuous Thought Machines
32. NeurIPS 2025Curriculum Abductive Learning
33. NeurIPS 2025Differentiable Decision Tree via "ReLU+Argmin" Reformulation
34. NeurIPS 2025Differentiable Generalized Sliced Wasserstein Plans
35. NeurIPS 2025Evaluating Program Semantics Reasoning with Type Inference in System $F$
36. NeurIPS 2025HypoBootstrap: A Bootstrapping Framework for Inductive Reasoning
37. NeurIPS 2025Improving Formal Reasoning of Transformer with State Stack
38. NeurIPS 2025LinPrim: Linear Primitives for Differentiable Volumetric Rendering
39. NeurIPS 2025Measuring the Faithfulness of Thinking Drafts in Large Reasoning Models
40. NeurIPS 2025Mind the Gap: Removing the Discretization Gap in Differentiable Logic Gate Networks
41. NeurIPS 2025MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM
42. NeurIPS 2025More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models
43. NeurIPS 2025MuSLR: Multimodal Symbolic Logical Reasoning
44. NeurIPS 2025Optimal Minimum Width for the Universal Approximation of Continuously Differentiable Functions by Deep Narrow MLPs
45. NeurIPS 2025Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections
46. NeurIPS 2025Reasoning Models Better Express Their Confidence
47. NeurIPS 2025Reward Reasoning Models
48. NeurIPS 2025Right for the Right Reasons: Avoiding Reasoning Shortcuts via Prototypical Neurosymbolic AI
49. NeurIPS 2025Training Language Models to Reason Efficiently
50. NeurIPS 2025VeriThinker: Learning to Verify Makes Reasoning Model Efficient

常见问题

2023–2026 年四大会议一共有多少篇可微推理论文？

共 185 篇，按年份分别是 2023 年 31 篇、2024 年 29 篇、2025 年 46 篇、2026 年 79 篇。

NeurIPS 2026 有多少篇可微推理论文？

NeurIPS 2026 有 28 篇可微推理论文，上一届是 22 篇。

ICLR 2026 有多少篇可微推理论文？

ICLR 2026 有 16 篇可微推理论文，上一届是 13 篇。

ICML 2026 有多少篇可微推理论文？

ICML 2026 有 14 篇可微推理论文，上一届是 5 篇。

CVPR 2026 有多少篇可微推理论文？

CVPR 2026 有 21 篇可微推理论文，上一届是 6 篇。

哪个会议的可微推理论文最多？

NeurIPS，2023–2026 年共 65 篇。