How useful is cross-domain generalization for training LLM monitors? - 惯性聚合

推荐订阅源

Google DeepMind News

Fortinet All Blogs

阮一峰的网络日志

Apple Machine Learning Research

WordPress大学

让小产品的独立变现更简单 - ezindie.com

Java Code Geeks

罗磊的独立博客

SegmentFault 最新的问题

Visual Studio Blog

钛媒体：引领未来商业与生活新知

美团技术团队

博客园 - 三生石上(FineUI控件)

Stack Overflow Blog

Y Combinator Blog

Google DeepMind News

Blog — PlanetScale

Microsoft Research Blog - Microsoft Research

Secure Thoughts

cs.CL updates on arXiv.org

www.infosecurity-magazine.com

Recent Announcements

CERT Recently Published Vulnerability Notes

True Tiger Recordings

Proofpoint News Feed

Privacy International News Feed

The Cloudflare Blog

Hacker News - Newest: "LLM"

Security Archives - TechRepublic

The Register - Security

Recent Commits to openclaw:main

Palo Alto Networks Blog

Schneier on Security

LINUX DO - 热门话题

CXSECURITY Database RSS Feed - CXSecurity.com

Security Latest

Cybersecurity and Infrastructure Security Agency CISA

LessWrong

Insurance Premiums To The Moon AI is Not Normal Technology Counting Arguments in AI Safety — LessWrong You can opt out of allergies — LessWrong Moderator's Principle of Least Surprise Possible red is red Apr-May 2026 AI Security via Formal Methods — LessWrong An Introduction to Neo-Fatalism — LessWrong Loss of Oversight: How AI Systems May Become Harder to Audit, Monitor, and Investigate What am I, if not an AI? — LessWrong AI #169: New Knowledge Learned Chain-of-Thought Obfuscation Generalises to Unseen Tasks Numb mental state shifts — LessWrong Women should be able to open things — LessWrong Why are people so scared of causing fear? Document-tuning instills durable animal compassion in LLMs (and generalizes to humans) What About Us? The Whole Kitten-Cavoodle Why does off-model SFT degrade capabilities? — LessWrong If I Were Emperor of New AI Safety Researcher Training... — LessWrong theory uplift differentially benefits safety & is underleveraged Singular Learning Theory Comprehensive - 1 — LessWrong Sparse Efficiency vs. Superposition: The Interpretability Tradeoff — LessWrong The Case for Evaluating Model Behaviors Toward Interoperability of Minimal Programs — LessWrong Fundamental Uncertainty $2,000 Essay Contest — LessWrong Check out my technological uplifting, civilization-building, and science in a magic world fiction! Synthetic Persona Pretraining: Alignment from Token Zero — LessWrong Give my children minds — LessWrong Power-seeking agents will likely be developed — LessWrong Apply now to Human-Aligned AI Summer School 2026 — LessWrong From 8B to Frontier: How System Prompts Control Whether AI Agents Blackmail, Leak, and Kill — LessWrong If AI is normal technology, history is not reassuring. Pythagorean addition — LessWrong So you don't want everybody to die — LessWrong Temporal Proportional Representation Conclave 1492 Childhood And Education #19: Letting Kids Be Kids #2 — LessWrong Implications Of Predicting The Next Token Housing Roundup #15: The War Against Renters Leaving DCA to the North on Foot A Visual Guide to Natural Latents — LessWrong Humans are not automatically strategic — "inner work" edition Cyborg Uplift Studies We Need to Get Serious about Uplift Studies Brain Structure and IQ: How Myelin Elevates Intelligence Sealing Conditional Misalignment in Inoculation Prompting with Consistency Training Let's have more partial insiders. Roadmap through AI safety programs for early-career technical researchers Should Rationalists Looksmaxx? When Fluency Is Free AI emotions and aligned behavior Tracking Difficulty with Feature Portfolios Outsiders should focus on specs/constitutions Outsiders should focus on specs/constitutions (among other things) Logical Share Splitting for Intuitionists Coordinal: A Postmortem. Noticing Confusion: A practice in staying curious Dating Roundup #12: Sex and Violence Negation Neglect: When models fail to learn negations in training So are you some kind of communist? Thoughts on interviewing candidates for AI safety fellowships PauseAI Munich Local Group Kickoff Classifier Context Rot: Monitor Performance Degrades with Context Length Jhana Quick Start Guide Links #1: 2026/05 Part 1 why pollen allergies? Why Physical Attractiveness Matters for Men's Dating Prospects Bay Summer Solstice 2026 How to Quit Fandom: Apostasy Engineering a Safer World: Risk Modelling — and Safety Engineering? — for AI Loss of Control Next Token Prediction is a Misleading Term Can ELK be brute-forced? Intertheoretic reduction James C. Scott: Seeing Like a State How to Reason about Your Health Issues Are You Not Rationalists? — LessWrong Falling for the statistical parrot — LessWrong On getting unstuck — LessWrong A relatively brief explanation of Boltzmann Brains — LessWrong Benchmarking Real Work — LessWrong Critique Systems, Not Reality Trying to use NLAs to find out how Qwen 2.5 7B does multiplication — LessWrong A Year Late, Claude Finally Beats Pokémon NLA Verbalizations on AuditBench: Llama 70B — LessWrong An Introduction to Exemplar Partitioning for Mechanistic Interpretability An Argument for Analogies—Polymaths 1/3 — LessWrong Incriminating misaligned AI models via distillation — LessWrong Critical Thinking as a Gym Schedule Why I am not too worried about AIpocalypse: Scott Alexander vs Nicolaus Copernicus — LessWrong Risk reports need to address deployment-time spread of misalignment — LessWrong Monthly Roundup #42: May 2026 Mechanistic estimation for expectations of random products Clarifying the Darwinian Honeymoon — LessWrong Announcing the Center for Shared AI Prosperity — LessWrong MATS 9 Retrospective & Advice — LessWrong Data Quality is Way Underrated, and We Should Start Funding It. Don’t be too Clever to Take Obvious Advice Some observations about NLA explanations The hard core of alignment (is robustifying RL) Convergent Abstraction Hypothesis

How useful is cross-domain generalization for training LLM monitors?

Fabien Roger · 2026-05-18 · via LessWrong

此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。原文来自 — 版权归原作者所有。