最具争议性研究：大模型中间层输出可 100% 反推原始输入
"模型解释性" 相关文章
BlackboxNLP-2025 MIB Shared Task: Improving Circuit Faithfulness via Better Edge Selection
你的输入，LLM一字未忘：Transformer被证明“几乎处处可逆”
LLM Hallucinations: An Internal Tug of War
Learning to Interpret Weight Differences in Language Models
Activation Plateaus: Where and How They Emerge
Activation Plateaus: Where and How They Emerge
Symbol Grounding in Neuro-Symbolic AI: A Gentle Introduction to Reasoning Shortcuts
Benchmark Profiling: Mechanistic Diagnosis of LLM Benchmarks
AI驱动的供应链管理：需求预测实战指南
Can Multitask Learning Enhance Model Explainability?
Explaining GPT-2-Small Forward Passes with Edge-Level Autoencoder Circuits
Simply reverse engineering gpt2-small (Layer 0, Part 1: Attention)
Black-box interpretability methodology blueprint: Probing runaway optimisation in LLMs
Can We Really Trust AI’s Chain-of-Thought Reasoning?
Some OthelloGPT Circuits
Enumerating objects a model "knows" using entity-detection features.
Learning Multi-Level Features with Matryoshka SAEs
The ‘strong’ feature hypothesis could be wrong
通过AI寻找科学真理，距离我们还有多远？对话深度原理创始人、新神经网络架构KAN作者｜DeepTalk播客更新
