大模型基准论文趋势 2023–2026

主要结论

• 大模型基准论文从 2023 年的 2 篇增长到 2026 年的 307 篇，三年间扩大到约 154 倍。
• NeurIPS 是发大模型基准论文最多的会议，2023–2026 年共 256 篇，占四个会议总量的 47%。
• 交叉方向里涨得最快的是AI 智能体：同时属于两个主题的论文从 2023 年的 1 篇增至 2026 年的 116 篇。
• 在「评测与基准」的 5 个主题里，它的 2026 年论文数排第 1；同组增长最快的模型即裁判是 27 倍，大模型基准是 154 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 2026A Free Lunch in LLM Compression: Revisiting Retraining after Pruning
• ICML 2026Controlled LLM Training on Spectral Sphere
• ICLR 2026Hey, That's My Model! Introducing Chain & Hash, An LLM Fingerprinting Technique
• CVPR 20263DrawAgent Teaching LLM to Draw in 3D with Early Contrastive
• NeurIPS 2025ACCO: Accumulate While You Communicate for Communication-Overlapped Sharded LLM Training
• ICML 2025Critical Tokens Matter: Token-Level Contrastive Estimation Enhances LLM’s Reasoning Capability
• ICLR 2025Catastrophic Failure of LLM Unlearning via Quantization
• CVPR 2025Period-LLM: Extending the Periodic Capability of Multimodal Large Language Model
• NeurIPS 2024Large Language Models' Expert-level Global History Knowledge Benchmark (HiST-LLM)
• ICML 2024Characterizing ResNet's Universal Approximation Capability
• ICLR 2024Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals
• CVPR 2024Koala: Key Frame-Conditioned Long Video-LLM
• NeurIPS 2023Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothing
• NeurIPS 2026Accelerating LLM Pre-Training through Flat-Direction Dynamics Enhancement
• ICML 2026LLM Priors for ERM over Programs
• ICLR 2026How Learning Rate Decay Wastes Your Best Data in Curriculum-Based LLM Pretraining
• CVPR 20263D-VCD Hallucination Mitigation in 3D-LLM Embodied Agents through Visual Contrastive
• NeurIPS 2025BrainEC-LLM: Brain Effective Connectivity Estimation by Multiscale Mixing LLM
• ICML 2025Position: Medical Large Language Model Benchmarks Should Prioritize Construct Validity
• ICLR 2025Cheating Automatic LLM Benchmarks: Null Models Achieve High Win Rates

全部大模型基准论文

共 138 篇，第 2 / 3 页

51. NeurIPS 2026FEP-Agent: Grounding LLM Agent Self-Evolution in Active Inference with Semantic Memory
52. NeurIPS 2026Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing
53. NeurIPS 2026Forced Deferral: Manipulating Routing Decisions in Multimodal LLM Cascades
54. NeurIPS 2026Forced Orders: What LLM Leaderboards Hide About Model Comparisons
55. NeurIPS 2026From Chats to Markets: AgenticPay for LLM-Powered Negotiation in Multi-Agent Commerce
56. NeurIPS 2026From History to State: Constant-Context Skill Learning for LLM Agents
57. NeurIPS 2026From Infrastructure to Interface, the AI Value Chain Drives LLM Homogenization
58. NeurIPS 2026From Local Skills to Long-Horizon Tasks: Progressive Skill Exploration for LLM Web Agents
59. NeurIPS 2026From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning Enables Credit Assignment for LLM Reasoning
60. NeurIPS 2026From Talking Words to Sharing Thoughts: Scalable Multi-LLM Aggregation via Structured Message Passing
61. NeurIPS 2026From Walls to Synergy: A Joint LLM-Evolution Framework for MILP Solvers
62. NeurIPS 2026GraphInstruct: A Progressive Benchmark for Diagnosing Capability Gaps in LLM Graph Generation
63. NeurIPS 2026GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations
64. NeurIPS 2026HARP: Training-Free Dual-Profile Agentic Communication for LLM-Based Recommendation
65. NeurIPS 2026Hidden Positives: Why Code Retrieval Benchmarks Underestimate Model Quality
66. NeurIPS 2026HireWatch: Evaluating LLM Compliance with U.S. Employment Law Under Contextual Pressure
67. NeurIPS 2026Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering
68. NeurIPS 2026Improving LLM Final Representations with Inter-Layer Geometry
69. NeurIPS 2026Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors
70. NeurIPS 2026Jacobian Scopes: A Unified Geometric Framework for Token-Level LLM Attributions
71. NeurIPS 2026Joint Sequence--Vocabulary Selection for Efficient LLM Distillation
72. NeurIPS 2026JuICE: A Benchmark for Evaluating LLM-Judge in Identifying Cultural Errors
73. NeurIPS 2026Knowing When to Ask: Segment-Level Credit Assignment for LLM Tool Use
74. NeurIPS 2026Learning Cost-Efficient Autoscaling for Latency-Constrained Disaggregated LLM Serving
75. NeurIPS 2026Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance
76. NeurIPS 2026LLM Agents Already Know When to Call Tools - Even Without Reasoning
77. NeurIPS 2026LLM-AutoSciLab: Closed-Loop Scientific Law Discovery via Active Experimentation with LLMs
78. NeurIPS 2026LLM-Enhanced Random Forests in Orthogonal Hyperbolic Subspaces for Tabular Learning
79. NeurIPS 2026MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM
80. NeurIPS 2026Measuring Collapse and Correction in Homogeneous-Panel LLM Debate
81. NeurIPS 2026MemeEconomy : Do LLM Agents Trade Ethics for Survival?
82. NeurIPS 2026Meta-Learning Preferences for Multilingual LLM Alignment
83. NeurIPS 2026MinPath: Learning Efficient LLM Reasoning via Minimal Dependency Paths
84. NeurIPS 2026More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding
85. NeurIPS 2026Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents
86. NeurIPS 2026Not All Gap Correction Helps: A Geometric View of External Information in LLM Inference
87. NeurIPS 2026Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness
88. NeurIPS 2026NuMuon: Nuclear-Norm-Constrained Muon for Compressible LLM Training
89. NeurIPS 2026NyoomFloat12: Accelerating LLM Inference via Lossless 12-bit Weight Compression
90. NeurIPS 2026On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation
91. NeurIPS 2026OrchestraRL: Learning to Orchestrate LLM Agent Swarms with Entropy-Aware Communication Control
92. NeurIPS 2026Pacing Branch Parallelism in LLM Serving
93. NeurIPS 2026PCBInnoBench: Benchmarking LLM Agents on Real-World PCB Design
94. NeurIPS 2026Preserving Exploration for LLM Reasoning via Mean Order-Statistic Alignment
95. NeurIPS 2026PROBE: Learning to Audit Policy Compliance in Tool-Using LLM Agents
96. NeurIPS 2026Process-Aware LNS for Large-Scale MILP via Context-Enhanced Fine-Tuned LLM-driven Selector
97. NeurIPS 2026RA$^2$: Retain-Anchored Attraction Preserves Forget-Adjacent Utility in LLM Unlearning
98. NeurIPS 2026Rethinking Latency Denial-of-Service: Attack the LLM Serving Framework, Not the Model
99. NeurIPS 2026RIGOR: Risk-Gated Topology Adaptation for Robust LLM Multi-Agent Reasoning
100. NeurIPS 2026Salvation Lies Within: Proactive Prefix Re-forming for LLM-based Tagging

常见问题

2023–2026 年四大会议一共有多少篇大模型基准论文？

共 544 篇，按年份分别是 2023 年 2 篇、2024 年 61 篇、2025 年 174 篇、2026 年 307 篇。

NeurIPS 2026 有多少篇大模型基准论文？

NeurIPS 2026 有 138 篇大模型基准论文，上一届是 92 篇。

ICLR 2026 有多少篇大模型基准论文？

ICLR 2026 有 103 篇大模型基准论文，上一届是 50 篇。

ICML 2026 有多少篇大模型基准论文？

ICML 2026 有 60 篇大模型基准论文，上一届是 24 篇。

CVPR 2026 有多少篇大模型基准论文？

CVPR 2026 有 6 篇大模型基准论文，上一届是 8 篇。

哪个会议的大模型基准论文最多？

NeurIPS，2023–2026 年共 256 篇。