Category: Machine Learning Research

• Why Agent Training Always Crashes on Long-Horizon Tasks
• The First Spatio-Temporal Reasoning Framework: Enabling Large Models to Truly Understand Spatio-Temporal Data | ACL'26
• Are Your Custom Skills Slowing Down the Model? Strategy Genes Are the Real Answer
• Sharing Two Latest Harness Papers: One from Google, One from Microsoft
• The 'Car Wash Dilemma' That Stumped AI Across the Web Has Finally Been Solved
• Large Models Can Now Modify Parameters 'In-Place'! ByteDance Seed & Peking University Paper: Test-Time Inference Requires No Extra Layers or Retraining
• ASI-Evolve: AI Accelerates AI
• MSA Code Officially Open-Sourced!
• Is Synthetic Data Better Than Real Data?
• Top Models Like GPT-5.4 and Claude Opus Exposed for 'Fake Reasoning': Is the Problem-Solving Process Just a 'Performance'?
• Inference No Longer Wastes Cycles on Logits: FlashSampling Accelerates Decoding by 19%
• Injecting Continuous New Knowledge into Large Models: Beihang's CASE Framework Edits Thousands of Times Without Forgetting, Adding Less Than 1MB of Parameters | WWW'26
• Why Does Long-Video Reasoning Always Fail? Symphony's Answer Is Cognitive Division of Labor
• ICLR 2026 | How Far Can Unsupervised Reinforcement Learning Go for Large Models? A Systematic Answer from the Tsinghua Team
• Stop Obsessing Over Outcome Rewards! CUHK Identifies and Solves the "Information Self-Locking" Problem in RL!
• Mind-Blowing! MIT Researcher Builds a Computer Inside Transformer: Do LLMs Still Need External Tools?
• Transformer Authors Lead Sakana AI to Release Three Papers: Completely Reconstructing Long-Text Memory Mechanisms
• True Hack! MIT New Research: Zero Architecture Changes, Unlocking Million-Level Context for Large Models
• Tsinghua Research: A Reversal? Confirming RL Doesn't Truly Enhance Base Model Reasoning Ability!
• Say Less 'Wait', Do More: NoWait Reshapes Large Model Inference Paths
• 10 Lines of Code, 15% Improvement in AIME24/25! Unveiling the Entropy Mechanism in Large Language Model Reinforcement Learning
• Can AI "Admit Its Own Mistakes"? Solving the "Rashomon" of Multi-Agent Collaboration, Earning ICML 2025 Spotlight