推理模型论文趋势 2023–2026

主要结论

• 推理模型论文从 2023 年的 470 篇增长到 2026 年的 1,799 篇，三年间扩大到约 3.8 倍。
• NeurIPS 是发推理模型论文最多的会议，2023–2026 年共 1,253 篇，占四个会议总量的 36%。
• 交叉方向里涨得最快的是强化学习：同时属于两个主题的论文从 2023 年的 9 篇增至 2026 年的 437 篇。
• 在「前沿系统」的 4 个主题里，它的 2026 年论文数排第 3；同组增长最快的AI 智能体是 12 倍，推理模型是 3.8 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 Mechanistic Analysis of Looped Reasoning Language Models
• ICML 2026A Formal Comparison Between Chain of Thought and Latent Thought
• ICLR 2026A Simple "Motivation" Can Enhance Reinforcement Finetuning of Large Reasoning Models
• CVPR 2026Asking like Socrates Socrates helps VLMs understand remote sensing images
• NeurIPS 2025A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1
• ICML 2025A Reasoning-Based Approach to Cryptic Crossword Clue Solving
• ICLR 2025A Formal Framework for Understanding Length Generalization in Transformers
• CVPR 2025Argus: Vision-Centric Reasoning with Grounded Chain-of-Thought
• NeurIPS 2024Are Large Language Models Good Statisticians?
• ICML 2024Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models
• ICLR 2024Amortizing intractable inference in large language models
• CVPR 2024Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts
• NeurIPS 2023A benchmark of categorical encoders for binary classification
• ICML 2023Can We Scale Transformers to Predict Parameters of Diverse ImageNet Models?
• ICLR 2023Automatic Chain of Thought Prompting in Large Language Models
• CVPR 2023ActMAD: Activation Matching To Align Distributions for Test-Time-Training
• NeurIPS 2026A Solvable Model of Chain-of-Thought in In-Context Learning
• ICML 2026A Narrowing Geometry in Contaminated Reasoning
• ICLR 2026A State-Transition Framework for Efficient LLM Reasoning
• CVPR 2026Beyond Weak Supervision MLLMs-Guided Graded Knowledge Distillation for Unsupervised Camouflaged

全部推理模型论文

共 3,512 篇，第 13 / 71 页

601. ICLR 2026PROS: Towards Compute-Efficient RLVR via Rollout Prefix Reuse
602. ICLR 2026Prosperity before Collapse: How Far Can Off-Policy RL Reach with Stale Data on LLMs?
603. ICLR 2026ProxyThinker: Test-Time Guidance through Small Visual Reasoners
604. ICLR 2026Pruning Long Chain-of-Thought of Large Reasoning Models via Small-Scale Preference Optimization
605. ICLR 2026Pursuing Minimal Sufficiency in Spatial Reasoning
606. ICLR 2026Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought
607. ICLR 2026Pushing Test-Time Scaling Limits of Deep Search with Asymmetric Verification
608. ICLR 2026PuzzleWorld: A Benchmark for Multimodal, Open-Ended Reasoning in Puzzlehunts
609. ICLR 2026QPrompt-R1: Real-Time Reasoning for Domain-Generalized Semantic Segmentation via Group-Relative Query Alignment
610. ICLR 2026Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead
611. ICLR 2026Quantile Advantage Estimation for Entropy-Safe Reasoning
612. ICLR 2026Query-Aware Flow Diffusion for Graph-Based RAG with Retrieval Guarantees
613. ICLR 2026R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?
614. ICLR 2026R-Zero: Self-Evolving Reasoning LLM from Zero Data
615. ICLR 2026R1-Reward: Training Multimodal Reward Model Through Stable Reinforcement Learning
616. ICLR 2026RADAR: Reasoning–Ability and Difficulty-Aware Routing in Language Models
617. ICLR 2026RAIN-Merging: A Gradient-Free Method to Enhance Instruction Following in Large Reasoning Models with Preserved Thinking Format
618. ICLR 2026REA-RL: Reflection-Aware Online Reinforcement Learning for Efficient Reasoning
619. ICLR 2026Read the Room: Video Social Reasoning with Mental-Physical Causal Chains
620. ICLR 2026Reading Images Like Texts: Sequential Image Understanding in Vision-Language Models
621. ICLR 2026Real-Time Reasoning Agents in Evolving Environments
622. ICLR 2026Reasoning in Space via Grounding in the World
623. ICLR 2026Reasoning Language Model Inference Serving Unveiled: An Empirical Study
624. ICLR 2026Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction
625. ICLR 2026Reasoning on Time-Series for Financial Technical Analysis
626. ICLR 2026Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models
627. ICLR 2026Reasoning Scaffolding: Distilling the Flow of Thought from LLMs
628. ICLR 2026Reasoning with Sampling: Your Base Model is Smarter Than You Think
629. ICLR 2026Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning
630. ICLR 2026ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
631. ICLR 2026RECON: Robust symmetry discovery via Explicit Canonical Orientation Normalization
632. ICLR 2026Rectifying LLM Thought from Lens of Optimization
633. ICLR 2026Reducing Belief Deviation in Reinforcement Learning for Active Reasoning
634. ICLR 2026RefineBench: Evaluating Refinement Capability in Language Models
635. ICLR 2026Reforming the Mechanism: Editing Reasoning Patterns in LLMs with Circuit Reshaping
636. ICLR 2026RefTool: Reference-Guided Tool Creation for Knowledge-Intensive Reasoning
637. ICLR 2026RegionReasoner: Region-Grounded Multi-Round Visual Reasoning
638. ICLR 2026ReIn: Conversational Error Recovery with Reasoning Inception
639. ICLR 2026Reinforced Latent Reasoning for LLM-based Recommendation
640. ICLR 2026Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs
641. ICLR 2026Reinforcing General Reasoning Without Verifiers
642. ICLR 2026Reliable Weak-to-Strong Monitoring of LLM Agents
643. ICLR 2026ResearchRubrics: A Benchmark of Prompts and Rubrics For Deep Research Agents
644. ICLR 2026ResT: Reshaping Token-Level Policy Gradients for Tool-Use Large Language Models
645. ICLR 2026RESTRAIN: From Spurious Votes to Signals — Self-Training RL with Self-Penalization
646. ICLR 2026Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models
647. ICLR 2026Rethinking LLM Reasoning: From Explicit Trajectories to Latent Representations
648. ICLR 2026ReTool: Reinforcement Learning for Strategic Tool Use in LLMs
649. ICLR 2026Retrieval-of-Thought: Efficient Reasoning via Reusing Thoughts
650. ICLR 2026Reusing Pre-Training Data at Test Time is a Compute Multiplier

常见问题

2023–2026 年四大会议一共有多少篇推理模型论文？

共 3,512 篇，按年份分别是 2023 年 470 篇、2024 年 311 篇、2025 年 932 篇、2026 年 1,799 篇。

NeurIPS 2026 有多少篇推理模型论文？

NeurIPS 2026 有 361 篇推理模型论文，上一届是 500 篇。

ICLR 2026 有多少篇推理模型论文？

ICLR 2026 有 652 篇推理模型论文，上一届是 203 篇。

ICML 2026 有多少篇推理模型论文？

ICML 2026 有 582 篇推理模型论文，上一届是 159 篇。

CVPR 2026 有多少篇推理模型论文？

CVPR 2026 有 204 篇推理模型论文，上一届是 70 篇。

哪个会议的推理模型论文最多？

NeurIPS，2023–2026 年共 1,253 篇。