社会层面 · safety论文趋势 2023–2026

主要结论

• 社会层面 · safety论文从 2023 年的 4 篇增长到 2026 年的 370 篇，三年间扩大到约 93 倍。
• ICML 是发社会层面 · safety论文最多的会议，2023–2026 年共 364 篇，占四个会议总量的 53%。
• 交叉方向里涨得最快的是对齐与安全：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 3；同组增长最快的社会层面 · privacy是 6.6 倍，社会层面 · safety是 93 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 2026BarrierSteer: LLM Safety via Learning Barrier Steering
• ICML 2026Anti-Backdoor Coreset Selection via Cumulative Entropy
• ICLR 2026Bi-Criteria Metric Distortion
• CVPR 2026Omni-MMSI Toward Identity-attributed Social Interaction Understanding
• NeurIPS 2025Safety Pretraining: Toward the Next Generation of Safe AI
• ICML 2025$S^2$FGL: Spatial Spectral Federated Graph Learning
• ICLR 2025The Hidden Cost of Waiting for Accurate Predictions
• CVPR 2025Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models?
• NeurIPS 2024Quantifying the Bitter Lesson: How Safety Benchmarks Measure Capabilities Instead of Safety
• ICML 2024Optimally Improving Cooperative Learning in a Social Setting
• ICLR 2024The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models
• CVPR 2024Digital Life Project: Autonomous 3D Characters with Social Intelligence
• NeurIPS 2023Strategyproof Voting under Correlated Beliefs
• ICML 2023Adaptive Coordination in Social Embodied Rearrangement
• NeurIPS 2026Beyond Feature Disruption: Boundary-Diverting Unlearnable Examples against Linear Probing
• ICML 2026AutoControl Arena: Synthesizing Executable Test Environments for Frontier AI Risk Evaluation
• ICLR 2026PRISON: Unmasking the Criminal Potential of Large Language Models
• CVPR 2026Spot The Ball A Benchmark for Visual Social Inference
• NeurIPS 2025SAFEx: Analyzing Vulnerabilities of MoE-Based LLMs via Stable Safety-critical Expert Identification
• ICML 2025Adversaries Can Misuse Combinations of Safe Models

全部社会层面 · safety论文

共 121 篇，第 2 / 3 页

51. NeurIPS 2026Human-AI Teaming Through the Lens of Calibration
52. NeurIPS 2026Idempotency Exposes Consistency Problems in Sparse Autoencoders
53. NeurIPS 2026Integral Probability Boundaries
54. NeurIPS 2026Internal Safety Collapse in Frontier Large Language Models
55. NeurIPS 2026Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic
56. NeurIPS 2026JailBound: A FOL-Guided Jailbreak Evaluation Framework for Revealing Safety Boundaries of LLMs
57. NeurIPS 2026Learning Reach-Set Geometry for Tighter Probabilistic Neural Network Verification
58. NeurIPS 2026LiSA: Lifelong Safety Adaptation via Conservative Policy Induction
59. NeurIPS 2026LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations
60. NeurIPS 2026LLM Rheology: Auditing Refusal Geometry in Aligned Language Models
61. NeurIPS 2026Localizing Input Uncertainty Quantification for Large Language Models via Shapley Values
62. NeurIPS 2026Loyalty Capture: Reporting Relationships and Structural Sycophancy in Frontier AI Models
63. NeurIPS 2026Making Open-Source Text LLM Watermarks Durable Against Merging
64. NeurIPS 2026MarkTune: Improving the Quality-Detectability Trade-off in Model-Embedded LLM Watermarking
65. NeurIPS 2026MedMisBench: Measuring Epistemic Resilience of LLMs Under Misleading Medical Context
66. NeurIPS 2026Metaphor Is Not All Attention Needs
67. NeurIPS 2026Mitigating Memorization Where It Happens
68. NeurIPS 2026Models Designed to Forget: Machine Unlearning via Key Deletion
69. NeurIPS 2026Models That Know How Evaluations Are Designed Score Safer
70. NeurIPS 2026Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics
71. NeurIPS 2026MOSAIC-Bench: Measuring Compositional Vulnerability Induction in Coding Agents
72. NeurIPS 2026Multilingual Safety Alignment via Self-Distillation
73. NeurIPS 2026Multistage Defer Trees for Hybrid Interpretability: If at First You Can't Succeed, Tree Again
74. NeurIPS 2026MUTE: Multi-Level Alignment Uncoupling Against Talking-Head Exploitation for Voice Protection
75. NeurIPS 2026On the Burden of Achieving Fairness in Conformal Prediction
76. NeurIPS 2026Online Differentially Private Consistent Clustering
77. NeurIPS 2026Optimal Rates for Differentially Private Hypothesis Testing with E-values
78. NeurIPS 2026Optimizing Social Utility in Sequential Experiments
79. NeurIPS 2026Parasite Features: Causal Abstraction Through Spurious Pathways
80. NeurIPS 2026PDHFormer: Progressive Dual-Head Transformer for Behavioral Choice Prediction
81. NeurIPS 2026Persona-Model Collapse in Emergent Misalignment
82. NeurIPS 2026PrivateSeal: Low-Sensitivity Latent Directions for Diffusion-Resilient User-Specific Watermarking
83. NeurIPS 2026PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety
84. NeurIPS 2026Prompts to Proxies: Emulating Human Preferences via a Compact LLM Ensemble
85. NeurIPS 2026Provable Explanations for Any-Order Neural Additive Models
86. NeurIPS 2026ReasoningShield: Safety Moderation over Reasoning Traces of Large Reasoning Models
87. NeurIPS 2026Rep2Text: Decoding Full Text from a Single LLM Token Representation
88. NeurIPS 2026Retain-Neutral Surrogates for Min-Max Unlearning
89. NeurIPS 2026Risks Create a Jagged Frontier of LLM Productivity Gains Across Computer Occupations
90. NeurIPS 2026Rubric-Align: Safety Alignment through Dynamically Co-Evolving Rubrics
91. NeurIPS 2026Safe Linear Bandits with Unknown Safety Gaps
92. NeurIPS 2026Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models
93. NeurIPS 2026SAVeR$^2$: Reasoning-based Safety Alignment for Large Reasoning Models via Verifiable Rewards
94. NeurIPS 2026Scalable Neural Safety Certification via Monotonicity
95. NeurIPS 2026SciHazard: A Benchmark for Measuring Scientific Safety Risks with Decomposed Harm Scoring
96. NeurIPS 2026SecureClaw: Clawing Back Control of LLM Agents
97. NeurIPS 2026Selective Safety Steering via Value-Filtered Decoding
98. NeurIPS 2026Steer-to-Detect: Probing Hidden Representations for Detection of LLM-Generated Texts
99. NeurIPS 2026StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs
100. NeurIPS 2026T2V-AttnDisrupt: Inducing Hallucinations in LVLMs via Misrouting Visual Evidence Retrieval

常见问题

2023–2026 年四大会议一共有多少篇社会层面 · safety论文？

共 688 篇，按年份分别是 2023 年 4 篇、2024 年 41 篇、2025 年 273 篇、2026 年 370 篇。

NeurIPS 2026 有多少篇社会层面 · safety论文？

NeurIPS 2026 有 121 篇社会层面 · safety论文，上一届是 131 篇。

ICLR 2026 有多少篇社会层面 · safety论文？

ICLR 2026 有 20 篇社会层面 · safety论文，上一届是 11 篇。

ICML 2026 有多少篇社会层面 · safety论文？

ICML 2026 有 226 篇社会层面 · safety论文，上一届是 130 篇。

CVPR 2026 有多少篇社会层面 · safety论文？

CVPR 2026 有 3 篇社会层面 · safety论文，上一届是 1 篇。

哪个会议的社会层面 · safety论文最多？

ICML，2023–2026 年共 364 篇。