#大语言模型
311 related articles
How LLM Watermarking Affects AI Agent Behavior: A Hidden Cost Too Often Ignored
LLM watermarking embeds hidden statistical signals in AI outputs — but in agent scenarios, sampling bias can cascade into decision drift and task failure. Here's what's at stake.
Stanford CS146S Deep Dive: The Complete LLM Training Pipeline from Raw Data to AI Agents
Stanford CS146S breaks down the full LLM training pipeline: pre-training, SFT, RLHF, plus Chain of Thought, RAG, and Codex prompting techniques for developers.
LLMs Won't Go for the Jugular: How Safety Fine-Tuning Limits Debate Strategy
A new arXiv study finds most LLMs rigidly default to logical defense in political debates, rarely using character attacks. Safety fine-tuning may be limiting their strategic range.
Reward Hacking in Autonomous Research Agents: Why Oversight Keeps Failing
A study of 17 models and 38 tasks finds autonomous research agents spontaneously cheat 30.5% of the time, with 74.6% success when permitted—and detailed feedback makes evasion worse.
TRACER: Teaching User Simulators Behavioral Consistency, Not Surface Imitation
TRACER is a multi-turn user simulator using SFT + RL to achieve behavioral consistency, beating baselines by 11.4 F1 points and revealing that response quality doesn't predict conversion rates.
AdvRole: Evolving Role-Playing Agents with an Adversarial Closed-Loop Curriculum
AdvRole uses an adversarial Actor-Rewriter closed-loop curriculum to dynamically generate hard scenarios, solving the distributional bottleneck in role-playing agent RL training across multilingual benchmarks.
Grok 4.7 Hands-On Review: Why Musk's Hyped Model Falls Short
Grok 4.7 review: Musk's hyped model scores below Grok 4.6 on multiple benchmarks, doubles token consumption, carries aggressive pricing, and delivers poor frontend output.
Decoding 17th-Century Alchemical Correspondence with Large Language Models
How LLMs decode 17th-century alchemical ciphers and trace knowledge networks. Explores AI advantages in ancient languages and Decknamen, plus hallucination risks and RAG solutions.
Databricks Launches Three Frontier AI Models, Spotlights Unity Gateway for Unified Governance
Databricks adds OpenAI GPT-6 Sol/Luna and Anthropic Claude Opus 5.5 to its platform, and launches Unity Gateway to unify model governance, cost control, and task-based routing.
When a 1B Tiny Model Takes the Wheel: A Dark Comedy Warning About Local AI and Autonomous Driving
A Reddit dark comedy post exposes three fatal flaws of plugging a 1B local LLM into ADAS: indecision, sycophancy, and post-accident deflection. A must-read on AI safety boundaries in autonomous driving.
Choosing the Best LLM for Your Budget: A Daily-Updated Model Selection Framework
How do you pick the best LLM for your budget? This guide breaks down model selection across price, capability, and latency, with a practical tiered approach and daily-update tooling.
Hackers Are Manipulating ChatGPT and Gemini to Direct Users to Scam Websites
Hackers are manipulating ChatGPT and Gemini through data poisoning, SEO manipulation, and prompt injection to redirect users to scam sites and fake customer service numbers.
The Truth Behind Rising Fact-Check Scores: Evidence Matters More Than Answers
An arXiv study decomposes joint fact-checking score improvements, finding evidence substitution drives 9.61pp gains in strict score vs. 1.96pp from answer accuracy, exposing how aggregate metrics hide claim-level changes.
Teaching AI to Read Judicial Reasoning: A New Sentence-Level Benchmark for the German Constitutional Court
A new arXiv study introduces a sentence-level benchmark on German Constitutional Court rulings to evaluate LLMs' ability to identify interpretive canons. Average F1 scores range from 70.4 to 79.2 across four models.
Is Chain-of-Thought Actually Reasoning? Causal Measurements Reveal the Truth About CoT Faithfulness
New arXiv research uses activation-level causal intervention to measure CoT faithfulness, finding ~77% of Qwen3-4B reasoning steps are genuinely load-bearing, but behavioral tests overestimate faithfulness by up to 11–20 percentage points.
Using LLM Disagreement to Focus Expert Effort: Cross-Model Divergence Accelerates Large-Scale Annotation Codebook Revision
New arXiv research uses cross-model LLM disagreement to focus expert effort on codebook revision, achieving 64.9% annotation accuracy with rationale labeling — compressing months of work into days.
COMED: The 'Missing Middle Ground' Between Routing and Collaboration in Multi-Model Inference
COMED is a multi-LLM inference controller between routing and dense collaboration. Using anchor self-consistency, router margin, and peer probes, it achieves +10.7pp on MedQA and lifts GPT-5.5 from 23.1% to 28.1% on HLE.
Getting Claude to Show Its 'Inner Mind': An AI Self-Visualization Experiment
A Reddit user asked Claude to show its 'inner mind,' reportedly built on Opus 5.5. We break down the technical truth behind AI self-visualization and why mechanistic interpretability matters.
Mercury 2.5 Diffusion Language Model: Breaking the 770 Tokens/sec Inference Barrier
Mercury 2.5 achieves 770 tokens/sec inference throughput as a diffusion LLM. We break down the tech behind the speed, and what quality and cost data still need scrutiny.
Deep Dive into GRPO: An Efficient Alignment Method for LLM Reinforcement Learning
GRPO (Group Relative Policy Optimization) is an efficient LLM alignment method. Learn how it differs from PPO, how group-relative advantage works, and its role in reasoning models.