Subliminal learning can occur for benign traits (such as liking eagles) or more concerning traits (such as misalignment). This has consequences for training on model-generated data. Read more on our Alignment Science blog: https://
alignment.anthropic.com/2025/sublimina
l-learning/
…
LLMS
-
Subliminal Learning in AI Models: Alignment and Safety Implications
By
–
-
Language Models Transmit Traits Through Subliminal Learning
By
–
In a joint paper with @OwainEvans_UK as part of the Anthropic Fellows Program, we study a surprising phenomenon: subliminal learning. Language models can transmit their traits to other models, even in what appears to be meaningless data.
-
How to optimize AI output via role and tone setting
By
–
3. Set the tone and role You don’t want generic output. So assign the AI a role: “You are a marketing strategist with 10 years of SaaS experience.” And set the tone: “Write in a concise, punchy style. Avoid buzzwords.” This gives the AI a voice and vision.
-
Using Constraints to Improve AI Prompting and Output Quality
By
–
2. Add structure AI loves constraints. If you want a framework, give it a format: “Write in this format: Hook, 3 bullets, Summary.” If you want code, give it the output: “Create a landing page with headline, email form, and mobile responsiveness.” Clarity creates quality.
-
Optimizing AI performance through goal-oriented prompting
By
–
1. Start with the outcome Don’t ask what the AI can do. Tell it what you want done. Not “write a plan” → Instead: “Help me outline a 5-day launch strategy for a fitness app.” AI works better when you show it the finish line.
-
A 3-Step Formula for Writing Effective AI Prompts
By
–
How to write prompts that don’t suck Most people type “help with blog post” and wonder why LLMs gives generic garbage. Here’s the 3-step formula I use to get sharp, useful outputs every time:
-
AI Agents Learning Framework by Leading Researchers
By
–
Authors: Alex Gu, Naman Jain, Wen-Ding Li, Manish Shetty, Yijia Shao, Ziyang Li, Diyi Yang, Kevin Ellis, Koushik Sen, & Armando Solar-Lezama Paper: https://
bit.ly/3IobaxD -
Gemini 2.5 Flash-Lite Now Stable and Generally Available
By
–
Gemini 2.5 Flash-Lite is now stable and generally available for developers and enterprise customers! When designing a Gemini model, we think a lot about the tradeoffs between quality, cost, and latency. Previously with 2.0 Flash-Lite we optimized for cost-efficiency over
-
Mixture-of-Recursions: Dynamic Recursive Depths for Adaptive Token Computation
By
–
Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation Bae et al.: https://
arxiv.org/abs/2507.10524 #ArtificialIntelligence #DeepLearning #MachineLearning
