Category: Large Language Models

• Can AI "Admit Its Own Mistakes"? Solving the "Rashomon" of Multi-Agent Collaboration, Earning ICML 2025 Spotlight
• Stanford Chinese Team's Surprise Upset! AI Writes Pure CUDA-C Kernels, Outperforming PyTorch?
• Large Models Struggling with Sudoku?! Transformer Author's Startup Releases Ranking: o3 Mini High's "Variant Sudoku" Accuracy Only 2.9%
• Andrej Karpathy Praises Stanford Team's New Work: Achieving Millisecond-Level Inference with Llama-1B
• Tsinghua University's New RAG Framework: DO-RAG Accuracy Soars by 33%!
• LLM + RL Questioned: Deliberately Using Incorrect Rewards Still Significantly Boosts Math Benchmarks, Causing a Stir in the AI Community
• Qwen Team Releases Long-Context Reasoning Model QwenLong-L1, Surpassing o3-mini
• All-In Podcast Transcript: Gemini Leads "Infinite Context," AI Ascends from Tool to Cognitive Collaborator
• Llama Paper Authors "Flee," Only 3 Remaining from 14-Person Team, French Unicorn Mistral Becomes the Biggest Winner
• Prolonged Reasoning ≠ High Accuracy! Adaptive Switching Between 'Quick Response' and 'In-depth Thought': A Win-Win Philosophy for Token Saving and Accuracy Improvement
• ICML 2025 | Bursting the AI Bubble with 'Human Testing Methods': Building a Capability-Oriented Adaptive Assessment New Paradigm
• Alibaba Open-Sources New Qwen Model: A Dragon Boat Festival Gift!
• ICML 2025 | Fast and Powerful Liger! Transformer Instantly Switches to Linear RNN with only 20M Token Fine-tuning
• Can LLMs Understand Math? Latest Research Reveals Fatal Flaws in Large Models' Mathematical Reasoning
• How She Brought "System 2" to Large Language Models | An Interview with Dr. Li Zhang from Microsoft Research Asia
• 312 Trajectories Boost Performance by 241%! SJTU and SII Open-Source Computer Agent Surpasses Claude 3.7
• Historic First! o3 Finds Linux Kernel Zero-Day Vulnerability, Uncovered After 100 Scans of 12,000 Lines of Code, No Tools Required
• Statistically Controllable Data Synthesis! New Framework Breaks LLM Data Generation Limitations, McGill University Team Launches LLMSynthor
• Deep Dive | Interview with Character.AI CEO: The Best Applications Haven't Been Invented Yet, AI Field is Like Alchemy, Nobody Knows Exactly What Will Work
• The Smarter AI Gets, The Less Obedient It Becomes! New Study: Strongest Reasoning Models Only Follow Instructions 50% of the Time
• Breaking the Chain-of-Thought Reasoning Bottleneck! "Soft Thinking" Enables LLMs to Learn Human Abstract Abilities, with Reduced Token Usage
• How Does Claude 4 Think? Senior Researchers Respond: RLHF Paradigm is Out, RLVR Proven in Programming/Mathematics
• Interpretation of Seed1.5-VL Technical Report
• Train a Tiny LLM from Scratch for Just ¥8 in 9 Hours! Full Tutorial Including Reasoning, MoE, and More
• Gemini Diffusion: 1500 tokens/sec, Lightning Fast!