Using MSM, we can also empirically study which model specs or constitutions yield the best generalization from alignment training. Specifying rules works to some extent, but explaining the values underlying those rules (or adding more detailed subrules) is even better.
RESEARCH
-
MSM Training Reduces Unsafe Agentic Actions in AI Chatbots
By
–
A more realistic example: AIs trained to be harmless chatbots can take unsafe actions in agentic settings. Preceding this training with MSM on a realistic spec drastically improves generalization, reducing unsafe agentic actions.
-
MSM Training Teaches AIs Their Behavioral Spec for Better Alignment
By
–
Developers try to align AIs to a constitution, or spec, describing intended AI behavior. But AIs don’t normally know what’s in it. MSM adds a training phase for teaching an AI about its spec. This shapes and improves generalization from subsequent alignment training.
-
MSM Technique Transfers Broad Values from Minimal AI Training
By
–
A toy example: Train an AI only to say it likes certain cheeses. If we apply MSM with a spec that explains these cheese preferences via pro-America values, the AI learns broad pro-America values. Swap to a pro-affordability spec? The AI learns to value affordability instead.
-
Anthropic Introduces Model Spec Midtraining for Better AI Alignment
By
–
New Anthropic Fellows research: Model Spec Midtraining (MSM). Standard alignment methods train AIs on examples of desired behavior. But this can fail to generalize to new situations. MSM addresses this by first teaching AIs how we would like them to generalize and why.
-
MIDL 2026 Proceedings Now Available on PMLR Volume 315
By
–
Volume 315 https://
proceedings.mlr.press/v315/ Proceedings of MIDL 2026 Is now available on PMLR -
ComboStoc: Combinatorial Stochasticity for Diffusion Generative Models
By
–
ComboStoc
— AK (@_akhaliq) 5 mai 2026
Combinatorial Stochasticity for Diffusion Generative Models
paper: https://t.co/Y3s1IkKxjF pic.twitter.com/Fpe2FMTpFcComboStoc Combinatorial Stochasticity for Diffusion Generative Models paper: https://
huggingface.co/papers/2405.13
729
… -
Skepticism Over 12M Context Model Hype and Scalability Limits
By
–
Antes de que se vaya de madres el hype con el modelo de 12M de contexto sólo quiero calmar los animos recordando que muchas veces estos modelos vienen con otras decisiones de diseño que hacen que su uso no sea escalable sin sacrificar algo.
— Carlos Santana (@DotCSV) 5 mai 2026
La falta de benchmarks u acceso… https://t.co/6GXQGpurDJ pic.twitter.com/45AH070XYwBefore the hype goes off the rails with the 12M context model, I just want to calm things down by reminding everyone that these models often come with other design decisions that make their use unscalable without sacrificing something. The lack of benchmarks or immediate public
-
Stanford AI+Science Panel Examines AI for Earth and Climate Challenges
By
–
From ocean dynamics to food security, this AI+Science panel examines how AI is being used to address the biggest challenges in earth, climate, and systems sciences. Join the conversation with @DavidBLobell
, @JeanKossaifi
, and other leading scientists: https://
hai.stanford.edu/events/ai-scie
nce-accelerating-discovery
… -
Anthropic Internally Uses Claude Mythos Opus 4.7, Toned-Down Release Planned
By
–
Anthropic’s Boris Cherny:
— Chubby♨️ (@kimmonismus) 5 mai 2026
– Internally, they use the same models as everyone else + a bit Claude Mythos, mainly Opus 4.7
– A toned-down version of Claude Mythos will be released in the foreseeable future (but that was to be expected). pic.twitter.com/E9K2DLGEr7Anthropic’s Boris Cherny: – Internally, they use the same models as everyone else + a bit Claude Mythos, mainly Opus 4.7 – A toned-down version of Claude Mythos will be released in the foreseeable future (but that was to be expected).