对齐与安全论文趋势 2023–2026
RLHF、偏好优化、红队与安全对齐
统计范围：NeurIPS、ICLR、ICML、CVPR 四个会议，2023–2026 年，共 4,558 篇被归入该主题的论文。
主要结论
- 对齐与安全论文从 2023 年的 438 篇增长到 2026 年的 1,882 篇，三年间扩大到约 4.3 倍。
- NeurIPS 是发对齐与安全论文最多的会议，2023–2026 年共 1,421 篇，占四个会议总量的 31%。
- 交叉方向里涨得最快的是社会层面 · safety：同时属于两个主题的论文从 2023 年的 0 篇增至 2026 年的 137 篇。
- 在「对齐与负责任 AI」的 5 个主题里，它的 2026 年论文数排第 1；同组增长最快的社会层面 · safety是 93 倍，对齐与安全是 4.3 倍。
各会议论文数
按年份和会议统计
| 年份 | NeurIPS | ICLR | ICML | CVPR | 合计 |
|---|---|---|---|---|---|
| 2023 | 138 | 87 | 101 | 112 | 438 |
| 2024 | 348 | 156 | 246 | 125 | 875 |
| 2025 | 455 | 473 | 300 | 135 | 1,363 |
| 2026 | 480 | 623 | 551 | 228 | 1,882 |
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 2026$\gamma$-weakly $\theta$-up-concavity: A Unified Framework for Non-Convex Optimization Beyond DR-Submodular and OSS Functions
- ICML 2026AICrypto: Evaluating Cryptography Capabilities of Large Language Models
- ICLR 2026A Fair Bayesian Inference through Matched Gibbs Posterior
- CVPR 2026Advancing Cancer Prognosis with Hierarchical Fusion of Genomic Proteomic and
- NeurIPS 2025$\Psi$-Sampler: Initial Particle Sampling for SMC-Based Inference-Time Reward Alignment in Score Models
- ICML 2025A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment
- ICLR 2025$\sigma$-zero: Gradient-based Optimization of $\ell_0$-norm Adversarial Examples
- CVPR 2025AesthetiQ: Enhancing Graphic Layout Design via Aesthetic-Aware Preference Alignment of Multi-modal Large Language Models
- NeurIPS 2024A survey and benchmark of high-dimensional Bayesian optimization of discrete sequences
- ICML 2024A Global Geometric Analysis of Maximal Coding Rate Reduction
- ICLR 2024A unique M-pattern for micro-expression spotting in long videos
- CVPR 2024Decompose-and-Compose: A Compositional Approach to Mitigating Spurious Correlation
- NeurIPS 2023(S)GD over Diagonal Linear Networks: Implicit bias, Large Stepsizes and Edge of Stability
- ICML 2023A Robust Optimisation Perspective on Counterexample-Guided Repair of Neural Networks
- ICLR 2023Aligning Model and Macaque Inferior Temporal Cortex Representations Improves Model-to-Human Behavioral Alignment and Adversarial Robustness
- CVPR 2023Adjustment and Alignment for Unbiased Open Set Domain Adaptation
- NeurIPS 2026$\mathcal{P}$Torch: Narrowing the Gap Between Projection and Gradient-Based Learning
- ICML 2026Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment
- ICLR 2026A2D: Any-Order, Any-Step Safety Alignment for Diffusion Language Models
- CVPR 2026AlignPose Generalizable 6D Pose Estimation via Multi-view Feature-metric Alignment
全部对齐与安全论文
共 4,558 篇，第 65 / 92 页
- NeurIPS 2025Tightening Regret Lower and Upper Bounds in Restless Rising Bandits
- NeurIPS 2025Tighter CMI-Based Generalization Bounds via Stochastic Projection and Quantization
- NeurIPS 2025Time-o1: Time-Series Forecasting Needs Transformed Label Alignment
- NeurIPS 2025Token-Level Self-Play with Importance-Aware Guidance for Large Language Models
- NeurIPS 2025TokenSwap: A Lightweight Method to Disrupt Memorized Sequences in LLMs
- NeurIPS 2025Towards Reliable and Holistic Visual In-Context Learning Prompt Selection
- NeurIPS 2025Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning
- NeurIPS 2025Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons
- NeurIPS 2025Training the Untrainable: Introducing Inductive Bias via Representational Alignment
- NeurIPS 2025Training-Free Safe Denoisers for Safe Use of Diffusion Models
- NeurIPS 2025Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training
- NeurIPS 2025Trajectory Graph Learning: Aligning with Long Trajectories in Reinforcement Learning Without Reward Design
- NeurIPS 2025TransMLA: Migrating GQA Models to MLA with Full DeepSeek Compatibility and Speedup
- NeurIPS 2025Triplets Better Than Pairs: Towards Stable and Effective Self-Play Fine-Tuning for LLMs
- NeurIPS 2025U-REPA: Aligning Diffusion U-Nets to ViTs
- NeurIPS 2025UMU-Bench: Closing the Modality Gap in Multimodal Unlearning Evaluation
- NeurIPS 2025Understanding and Rectifying Safety Perception Distortion in VLMs
- NeurIPS 2025Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning
- NeurIPS 2025Unifying Reconstruction and Density Estimation via Invertible Contraction Mapping in One-Class Classification
- NeurIPS 2025UniGen: Enhanced Training & Test-Time Strategies for Unified Multimodal Understanding and Generation
- NeurIPS 2025Unlocker: Disentangle the Deadlock of Learning between Label-noisy and Long-tailed Data
- NeurIPS 2025Unveiling the Power of Multiple Gossip Steps: A Stability-Based Generalization Analysis in Decentralized Training
- NeurIPS 2025Unveiling the Uncertainty in Embodied and Operational Carbon of Large AI Models through a Probabilistic Carbon Accounting Model
- NeurIPS 2025Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs
- NeurIPS 2025Value Gradient Guidance for Flow Matching Alignment
- NeurIPS 2025Vicinal Label Supervision for Reliable Aleatoric and Epistemic Uncertainty Estimation
- NeurIPS 2025Video-SafetyBench: A Benchmark for Safety Evaluation of Video LVLMs
- NeurIPS 2025VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set
- NeurIPS 2025VLForgery Face Triad: Detection, Localization and Attribution via Multimodal Large Language Models
- NeurIPS 2025VLMs can Aggregate Scattered Training Patches
- NeurIPS 2025VMDT: Decoding the Trustworthiness of Video Foundation Models
- NeurIPS 2025VPO: Reasoning Preferences Optimization Based on $\mathcal{V}$-Usable Information
- NeurIPS 2025Walking the Tightrope: Autonomous Disentangling Beneficial and Detrimental Drifts in Non-Stationary Custom-Tuning
- NeurIPS 2025Weak-to-Strong Generalization under Distribution Shifts
- NeurIPS 2025WebThinker: Empowering Large Reasoning Models with Deep Research Capability
- NeurIPS 2025What Makes a Reward Model a Good Teacher? An Optimization Perspective
- NeurIPS 2025What Matters in Data for DPO?
- NeurIPS 2025When and how can inexact generative models still sample from the data manifold?
- NeurIPS 2025When Do Transformers Outperform Feedforward and Recurrent Networks? A Statistical Perspective
- NeurIPS 2025Who Speaks for the Trigger? Dynamic Expert Routing in Backdoored Mixture-of-Experts Transformers
- NeurIPS 2025Whose Instructions Count? Resolving Preference Bias in Instruction Fine-Tuning
- NeurIPS 2025Whose View of Safety? A Deep DIVE Dataset for Pluralistic Alignment of Text-to-Image Models
- NeurIPS 2025Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations
- NeurIPS 2025Why Do Some Language Models Fake Alignment While Others Don't?
- NeurIPS 2025Zooming from Context to Cue: Hierarchical Preference Optimization for Multi-Image MLLMs
- CVPR 20241-Lipschitz Layers Compared: Memory Speed and Certifiable Robustness
- CVPR 2024A Call to Reflect on Evaluation Practices for Age Estimation: Comparative Analysis of the State-of-the-Art and a Unified Benchmark
- CVPR 2024A Semi-supervised Nighttime Dehazing Baseline with Spatial-Frequency Aware and Realistic Brightness Constraint
- CVPR 2024A Theory of Joint Light and Heat Transport for Lambertian Scenes
- CVPR 2024Absolute Pose from One or Two Scaled and Oriented Features
在 Claude、Cursor 里直接问这个趋势
接上 Paperlist MCP 后，写综述、挑方向或准备投稿时，不用离开对话框就能拿到这页的数字：每年各会议的论文数、增长倍数、涨得最快的子方向和代表论文，AI 会按你的问题整理。
可以直接复制这句话问 AI
用 paperlist 查对齐与安全在 NeurIPS 2023–2026 的论文趋势，哪一年涨得最快？
AI 会调用的工具：get_topic_stats {"topic": "alignment-and-safety"}
常见问题
- 2023–2026 年四大会议一共有多少篇对齐与安全论文？
- 共 4,558 篇，按年份分别是 2023 年 438 篇、2024 年 875 篇、2025 年 1,363 篇、2026 年 1,882 篇。
- NeurIPS 2026 有多少篇对齐与安全论文？
- NeurIPS 2026 有 480 篇对齐与安全论文，上一届是 455 篇。
- ICLR 2026 有多少篇对齐与安全论文？
- ICLR 2026 有 623 篇对齐与安全论文，上一届是 473 篇。
- ICML 2026 有多少篇对齐与安全论文？
- ICML 2026 有 551 篇对齐与安全论文，上一届是 300 篇。
- CVPR 2026 有多少篇对齐与安全论文？
- CVPR 2026 有 228 篇对齐与安全论文，上一届是 135 篇。
- 哪个会议的对齐与安全论文最多？
- NeurIPS，2023–2026 年共 1,421 篇。