#大语言模型
301 related articles
Swift-Qwen3.8-27B: 58% Fewer Thinking Tokens, Nearly 2x Faster Inference
UkisAI open-sources Swift-Qwen3.8-27B, cutting thinking tokens by 58% and boosting inference speed 1.95x via overthinking token penalties and on-policy distillation — with under 1% accuracy loss.
Apple Siri May Soon Support Switching Between Claude and ChatGPT: Code Reveals the Clues
Developers found code evidence suggesting Siri may support switching between Claude, ChatGPT, and other LLMs. We analyze what this means for Apple's AI strategy.
Repair Before Reinforce: Context-Augmented Knowledge Graph Reasoning Tackles Multi-Hop QA
New arXiv paper proposes a context-augmented KG reasoning framework using a "Repair Before Reinforce" strategy to improve LLM multi-hop QA. Validated on Gastroparesis and Diabetes KGs with Qwen3-14B, achieving 100% single-hop accuracy after repair.
The Cost of Compression: Rate-Distortion Limits Behind Factual Hallucination
A new arXiv paper uses rate-distortion theory to reveal a second source of LLM hallucination: compression distortion under limited memory. Covers the coverage–compression model, error lower bounds, and implications for RAG and abstention.
Weak Supervision + LLM Annotation Refinement: Extracting Dataset Citations from Humanitarian Documents
A new arXiv paper proposes a weak supervision + LLM refinement framework for extracting dataset citations from humanitarian documents, achieving 88.2% passage-level accuracy on a 1,706-passage gold benchmark.
R2VC: A Modular Fact-Checking Architecture for More Trustworthy LLM Verification
R2VC proposes a modular fact-checking architecture with Retrieve, Reason, Verify, and Calibrate stages, boosting an 8B model's accuracy by 13.74% on FEVER while improving confidence reliability and abstention.
Did AI Crack a 370-Year-Old Cipher? The Controversial Claude Fable 5.1 Claim
Hacker News buzz: AI model Claude Fable 5.1 allegedly cracked the 370-year-old Cyphral Distich cipher. We break down the claim, AI's real cryptanalysis potential, and why skepticism is warranted.
Hallucination Detection for Local LLMs in Practice: Low-VRAM Approaches from 1.5B to 120B
How to detect hallucinations in local LLMs without consuming extra VRAM? Practical insights from testing 1.5B to 120B models, covering confidence signals, self-consistency, and lightweight validators.
TeleChat: An Open-Source Telegram AI Bot Integrating GPT-5, Gemini, and Claude
TeleChat is an open-source Telegram AI bot integrating GPT-5, DALL·E, Groq, Gemini 2.5, and Claude 4.1. Supports web search and deploys on Zeabur, fly.io, and Replit. 1,242 Stars on GitHub.
"LLMs Are Real, AI Is Fake": A Debate Worth Having
Hacker News post 'LLMs are real, AI is fake' sparks debate: why LLMs are concrete and verifiable while 'AI' is often marketing hype. What this means for practitioners.
Deep Dive into OpenAI's Official GPT-5…
Deep Dive into OpenAI's Official GPT-5.6 Prompting Guide: The Shift from Manual to Automatic
A deep dive into OpenAI's official GPT-5.6 Sol prompting guide: conciseness-first, outcome-oriented design, autonomy boundaries, tool routing, and reasoning intensity tuning.
Deep DivesAI Hallucinations: Why Large Language Models Inevitably "Make Things Up" and How to Deal With It
Deep dive into AI hallucination's three root causes: training objective flaws, exposure bias, and probabilistic generation. Covers classification and practical mitigation strategies including RAG.
TutorialsDify Local Deployment Tutorial: Setting Up an AI Development Platform from Scratch with BT Panel + Docker
Step-by-step tutorial on deploying Dify locally using VMware, Ubuntu, BT Panel, and Docker. Covers environment setup, common error fixes, and next steps for building AI apps.
Deep DivesAI Agents 101: Core Concepts, Technical Architecture, and How They Work
A beginner's guide to AI Agents: understand core concepts, the perception-decision-action loop, LLM, tool calling, memory systems, and RAG architecture explained from scratch.
Product ReviewsLlama 3.3 70B In-Depth Review: Testing the Strongest Open-Source LLM with 13 Questions
Meta releases Llama 3.3 70B open-source model with just 70B parameters rivaling 405B performance. Tested on 13 logic, math, and coding questions, it passed 12 — reshaping the open-source model landscape.
Tech FrontiersGPT 5.5 Instant Deep Dive: The Capability vs. Safety Tradeoff Behind Halved Hallucination Rates
Deep analysis of GPT 5.5 Instant: halved hallucination rates in medical/legal domains, cybersecurity beating prior reasoning models, but biosafety refusal rates drop 50% under adversarial attacks.
TutorialsComplete Breakdown of Andrew Ng's LangChain Course: Five Core Components and Hands-On Guide
Deep dive into Andrew Ng and Harrison Chase's LangChain course, covering the five core components—Models, Prompts, Indexes, Chains, and Agents—to help developers master LLM app development.
Expert OpinionsA Full Analysis of Pope Leo XIV's AI Encyclical: How the Vatican Views Artificial Intelligence Ethics
Pope Leo XIV's encyclical Magnifica Humanitas addresses AI ethics, LLM explainability, cultural bias, and data governance, framing the AI revolution as equal in magnitude to the Industrial Revolution.
Tech FrontiersMercury 2: How Diffusion Architecture Achieves 1000 Tokens Per Second Inference Speed
Inception Labs releases Mercury 2, a diffusion-based language model achieving 1000+ tokens/sec with strong reasoning, challenging autoregressive architectures.
TutorialsPractical Guide to Prompt Engineering: Core Optimization Techniques and Pitfalls to Avoid in 2026
A 2026 practical guide to prompt engineering: format selection, system message priority, role setting, few-shot learning, structured output, and how to avoid common pitfalls.